<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
<channel>
  <title>Expedient AI</title>
  <link>https://expedientai.co.uk/</link>
  <atom:link href="https://expedientai.co.uk/feed.xml" rel="self" type="application/rss+xml"/>
  <description>Practical AI for UK small businesses: workflow automation, enquiry handling, CRM and reporting, and honest advice on whether AI is the right answer. AI for the rest of us.</description>
  <language>en-GB</language>
  <copyright>Expedient AI Ltd</copyright>
  <lastBuildDate>Tue, 13 Oct 2026 00:00:00 +0000</lastBuildDate>
  <item>
    <title>10 practical changes in Hermes Agent v0.22.0 (and what they're for) &amp;middot; Expedient AI</title>
    <link>https://expedientai.co.uk/feed/hermes-0-22-practical-changes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/hermes-0-22-practical-changes/</guid>
    <pubDate>Tue, 13 Oct 2026 00:00:00 +0000</pubDate>
    <description>Hermes Agent v0.22.0, the Threshold Release, landed on 2026-10-13. It is the largest window Hermes has shipped, about 2.5x the v0.21.0 cycle, and it adds connectors so the agent works inside Gmail, Linear and Notion, one-click local models on your own hardware, one gateway for every profile, live subagent steering, and a 34 percent smaller codebase. These are the ten changes that actually shift day-to-day work.</description>
    <content:encoded><![CDATA[Tools 10 practical changes in Hermes Agent v0.22.0 (and what they're for) Hermes Agent v0.22.0, the Threshold Release, landed on 2026-10-13. Upstream calls it the largest window it has shipped, about two and a half times the v0.21.0 cycle, with roughly 24,200 commits and 6,080 merged pull requests. Fifteen highlights are listed. Ten of them actually change how a one-person operator runs the thing day to day, and those are the ones below. 2026-10-13 · 11 min read Before the list, the shape of it. Hermes was, as the release puts it, the god of thresholds: the doorway, the crossing, the boundary stone. v0.21.0 assembled the pantheon. v0.22.0 is about getting through the door. A fresh install now opens on a setup chat that runs before you own an API key, the agent reaches into the apps you already use, local models come home to your own hardware in one click, and the packaging underneath became a tested stable channel rather than a moving target. 1. A first launch that sets itself up Open the desktop app on a machine where you have never used Hermes and it starts a short setup chat instead of a settings screen. One card at a time it asks your name, an accent colour and a layout, which apps to connect, which plugins to install, then gives a quick tour and offers up to four first tasks picked for you. It runs on a free trial model, so it needs no key. Skip setup ends it, and Settings can bring it back. The agent also gained a set_theme tool, so asking for a colour and a dark mode now works in any desktop chat. 2. Connectors: the agent works inside your apps This is the one that changes what the thing is for. The agent can now act in hosted apps such as Gmail, Linear and Notion through the Nous tool gateway. You ask for an app, sign in once in your browser, and the tools are usable in the same turn. You do not need a Nous account to start. The first time you use an app, Hermes creates a free connectors-only identity. The desktop has a Connectors page that puts hosted apps and your MCP servers in one list, replacing the MCP tab. You can connect several named accounts of one app, a work and a home Gmail for instance, and manage them with a connectors command. 3. Local models in one click Hermes can run open models entirely on your own machine with nothing leaving it. No account, no API key. It installs and maintains llama.cpp with the right engine for your hardware, whether CPU, CUDA, Metal, Vulkan or ROCm, rates every catalog model against your GPU and RAM, and picks the highest-quality build that fits. Downloads pause and resume, the server starts and stops with Hermes, and engine updates are one click. The catalog includes Qwen3.8 27B, Qwen3.6 35B-A3B and DeepSeek V4. This is the single most relevant change if you are selling self-hosted AI to a business that does not want data leaving the building. The gap between "we can run it on your own hardware" and "you can run it on your own hardware" used to be a weekend of setup. It is now a menu item. 4. Stable release channels Hermes now ships tested SemVer releases. Every release passes source CI, Docker, Nix, signed desktop builds and install/upgrade runs on every operating system before it is published. Source installs move from release to release by default instead of tracking every commit on main. If you prefer the bleeding edge it is still there. But for anyone running this in production, the default is no longer development. 5. Signed desktop packages and one package manager The desktop app is now a sealed, signed package that carries its own Python and Node, so first launch builds nothing. Windows gets a signed MSIX for both x64 and ARM64 through App Installer or the Microsoft Store. macOS gets a signed, notarized build for Apple Silicon. Underneath, a single package manager handles install and update. First launch no longer spends minutes building an environment. 6. The plugin catalog, now 600 reviewed plugins The catalog launched this cycle and grew to 600 entries, each pinned to a reviewed commit and installable from the desktop or the CLI. You can browse by category and star ranking, with a page per plugin and per author. Catalog rules keep listed plugins from patching Hermes internals or shipping unexpected code. 7. One backend and one gateway for every profile Hermes now runs a single messaging gateway and a single desktop backend per machine that serve all of your profiles, instead of one process per profile. Creating a profile makes its bots live without a restart. Each profile keeps its own credentials, allowlists, cron jobs and sessions, and the updater migrates existing per-profile setups automatically. Fewer processes, no token clashes, one thing to start and stop. If you run more than one profile, this is the change you will notice first. 8. See and steer your subagents live Running subagents now stay visible above the composer in the CLI, the TUI and the Desktop, each with its task, elapsed time and live transcript, plus Steer and Stop controls. The stop command reaches background children and hands back their partial work rather than discarding it, and delegated tasks can now pass screenshots and files through. 9. A model picker that knows your plans Star your favourite models, see a subscription running low with the reset time, and see a rate limit lift right in the picker. Reasoning effort is now part of choosing a model in the terminal, the slash command and the desktop. 10. Leaner in every sense The codebase was simplified with no behaviour change. Source went from about 1.06 million lines to 698 thousand, a 34 percent cut, and the gateway runner shrank from 34,847 lines to 5,512. Prompts got lighter too. About 19 rarely used tools now load only when the agent searches for them, which roughly halves the desktop tool schema sent on every call, and another 2,100 tokens came off the fixed prompt. Local models prefill less on every turn. Also in this window, briefly Bot Screen lets you watch a bot running on a Linux gateway and take the screen over when it hits a login, a two-factor prompt or a CAPTCHA, then hand control back while the bot carries on with your session. The browser agent can sign into sites using 1Password, Bitwarden or Hermes's own encrypted vault, with the model never seeing the password. Claude 5.5 and GPT-6 are pickable where offered, and prior reasoning is replayed by default on providers that can read it. Every user-facing string now goes through one translation catalog with full coverage in all 16 bundled languages. And there are about 19 rarely used tools on demand rather than always loaded, which is the prompt saving above. The breaking changes, before you upgrade This is the part worth reading twice. Python 3.14 is the runtime for managed installs, and the package manager installs it for you. KittenTTS is unavailable on 3.14. The Windows desktop is MSIX-only and needs Windows 11 22H2 or newer. The old NSIS and MSI installers are gone. One gateway per machine is the default. The updater folds per-profile gateways into the multiplexed default when nothing conflicts. Chats no longer reset on an idle or daily timer. Those modes are retired; a config that still sets them logs a warning. Use a fresh chat instead. Old sessions are pruned automatically. Ended sessions inactive for 90 days are deleted at startup. Pinned and open sessions are never touched, and you can set the auto-prune flag to false to keep everything. Memory providers moved to catalog plugins , along with Home Assistant, Spotify and Langfuse. Existing setups are migrated with data intact. Two providers remain bundled. Some tools were renamed for plugin authors. The old names still work, so nothing breaks on your side. Who this matters for If you are running Hermes yourself, connectors and local models are the two changes that alter what you can hand a client. If you are selling AI systems to a small business, local models in one click is the one to lead with, because it turns "self-hosted" from a claim you have to defend into a setting a non-technical owner can turn on. And the honest caveat: 600 plugins and a plugin catalog is a lot of surface area. The rules that pin plugins to reviewed commits and block them from patching internals are the right call, but it is a new place for supply-chain risk to live. Read the catalog rules before you install anything into a client machine. Running this yourself? The set up your own AI agent guide walks through getting a self-hosted agent running, and your data, your rules covers what can and cannot safely go into a hosted model. Frequently asked questions What is new in Hermes Agent v0.22.0? v0.22.0 is the Threshold Release, dated 2026-10-13, and the largest window Hermes has shipped at roughly two and a half times the v0.21.0 cycle. The headline additions are connectors that let the agent act inside hosted apps such as Gmail, Linear and Notion, one-click local models running entirely on your own machine, a single gateway and backend for all profiles, live subagent steering, and a 34 percent smaller codebase. Does v0.22.0 let me run AI models locally? Yes. Hermes installs and maintains llama.cpp with the right engine for your hardware, whether CPU, CUDA, Metal, Vulkan or ROCm. It rates every catalog model against your GPU and RAM and picks the highest quality build that fits. No account and no API key are needed, and downloads pause and resume. What are connectors in Hermes v0.22.0? Connectors let the agent act inside hosted apps such as Gmail, Linear and Notion through the Nous tool gateway. You ask for an app, sign in once in your browser, and the tools are usable in the same turn. You do not need a Nous account to start, and you can connect several named accounts of the same app. Are there breaking changes in v0.22.0? A few. Python 3.14 is now the runtime for managed installs and KittenTTS is unavailable on it. The Windows desktop is MSIX-only and needs Windows 11 22H2 or newer. One gateway per machine is now the default. Chats no longer reset on an idle or daily timer. Sessions inactive for 90 days are pruned automatically unless you set sessions.auto_prune to false. What does the smaller codebase mean for speed? Source went from about 1.06 million to 698 thousand lines, a 34 percent cut, with no behaviour change. Roughly 19 rarely used tools now load only when the agent searches for them, which roughly halves the tool schema sent on every call, and another 2,100 tokens came off the fixed prompt. Can I watch and steer my subagents in v0.22.0? Yes. Running subagents stay visible above the composer in the CLI, TUI and Desktop, each with its task, elapsed time and live transcript, plus Steer and Stop controls. The stop command reaches background children and hands back their partial work, and delegated tasks can pass screenshots and files. Twenty minutes. A short diagnostic. If you are a UK small business with a workflow that has been costing you time for months, the cheapest answer is a twenty-minute call. You will leave with a yes or no on whether AI is the right fix, and a rough cost and time range if it is. Get in touch via the contact form]]></content:encoded>
  </item>
  <item>
    <title>AI Employees for Small Business | AI Agents That Do Real Work</title>
    <link>https://expedientai.co.uk/ai-employees/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/ai-employees/</guid>
    <pubDate>Tue, 13 Oct 2026 00:00:00 +0000</pubDate>
    <description>What an AI employee actually is, what an autonomous AI agent can genuinely do for a UK small business, and how to implement one. ChatGPT Dots, Grok Bot, Meta Muse, Claude and self-hosted Hermes compared honestly.</description>
    <content:encoded><![CDATA[AI employees and AI agents You cannot hire an AI employee. You can build one. An AI employee is not a person and it is not magic. It is an autonomous agent: a piece of software that takes one recurring job, works through it without asking you at every step, and uses your own accounts to do it. For a UK small business that is genuinely useful. It is also narrower than the advertising suggests, and you deserve to know where the line sits before you spend money finding it. Book a free 20-minute call See the honest answers Free 20-minute diagnostic No lock-in You own what gets built What you are actually buying Three different things get called an AI employee. Most disappointing results come from buying one of these and expecting another. A chatbot answers you. An automation follows a script. An agent decides what to do next. 01 · Chatbot A conversation You ask, it answers. Useful for support deflection and internal questions. It will not chase anything or do anything on its own. 02 · Automation A script that fires When this happens, do that. Reliable and cheap, but it only handles the case you wrote. Nothing outside the script runs. 03 · Agent Decides for itself Given a goal and access to your tools, it plans the steps, uses them, and reports back. This is the one that saves you time. The honest bit What agents are genuinely good at today. The pattern that holds across every credible test is consistent. Give an agent work that is high volume, digital, measurable, and cheap to correct when one goes wrong, and it does well. Give it anything else and it does not. They work Enquiries arriving by email, web form or WhatsApp, answered the same way every time Quoting from your own price list and past jobs Customer support where the answer already exists in your documentation CRM updates, meeting notes and weekly reporting Chasing invoices and booking appointments They do not, yet Anything needing physical work, or a relationship you have built over years Judgement calls where the right answer depends on context you have not written down The job your best person does by knowing something unwritten Anything where being wrong is expensive or public Independent testing found the best agents complete only a fraction of open-ended office work, and Gartner expects a large share of agentic projects to be cancelled by 2027. Both point the same way: this replaces defined tasks, not people. The honest pitch is hours back, not staff cut. The five worth knowing What each one is actually for. Every one of these arrived in the last ten weeks, which is why so much of the coverage reads as hype. Here is the useful difference between them. Released 8 September 2026 · US only Meta Muse Meta's personal AI agent. It runs on its own virtual machine in Meta's cloud, with a browser you can watch as it works and a permissions system called Sentinel that separates read access from write access. It has its own email address, so you can forward something to it or copy it into a thread. Meta named Shopify, Stripe, Walmart, GitHub and Notion as commerce partners, so ordering and checkout inside a conversation is the intended path. Know this: Amazon blocked Muse from its retail site on 21 September, saying the agent did not identify itself and appeared to capture credentials. Meta says it cannot see passwords or payment details. Two weeks after launch, the clearest example of why agent access needs reviewing rather than assuming. Released 11 August 2026 Grok Bot xAI's agent, and the most literal reading of the term. It has its own cloud computer, signs into your apps using your accounts, and works through multi-step tasks rather than answering a single question. Worth watching if your work lives in the apps you already log into daily. It pairs well with a current-information model, which is what Grok is good at. Released 29 September 2026 · not available in the UK ChatGPT Dots An always-on agent with its own cloud computer, more than four thousand app integrations, and the ability to keep working when you close the tab. It handles scheduled jobs, messages your team in Slack when it needs a decision, and can hand coding work to Codex. Know this: Dots are not available on ChatGPT Pro in the UK, the European Economic Area or Switzerland, and OpenAI has not published a date for when that changes. It also needs Pro at around $100 a month rather than Plus. If you are in the UK, this is a read-and-watch page, not a buy button. I covered the UK options when it launched. Read the ChatGPT Dots breakdown for UK businesses Available now Claude The strongest option when the work is harder than the easy 10%. Claude runs multi-step tasks without waiting for your next prompt, and its Skills feature packages your firm's own repeatable expertise into named procedures your team can run. The Agent SDK lets Claude be built into production systems rather than used as a chat window, which is how it earns a place in the back office rather than the browser. What I run Hermes, self-hosted An open-source agent you host yourself. No per-seat licence, because the software is not rented to you, and your data stays on your own infrastructure instead of leaving the building. It has memory, tools, approvals and logs, so it can run the same workflow repeatedly rather than improvising each time. This is the default I recommend, and the reason is the licence and the data rather than the model quality, which is close across the board now. Side by side The five, compared. Tool Released Where you can get it Good for Meta Muse 8 September 2026 US only, free tier or $20 and $100 a month Commerce and checkout inside a conversation, with a permissions model Grok Bot 11 August 2026 Beta Signing into the apps you already use, with current information ChatGPT Dots 29 September 2026 Not in the UK or EEA. Pro at around $100 a month Scheduled background jobs across thousands of app integrations Claude Available now Claude plans and API Harder multi-step work, and packaging your own process as Skills Hermes Open source, now Your own hardware, your own choice of model Running your own workflows with no licence and no data leaving the building The only sensible way to choose Choose on access, not on capability. Capability has converged. The models are close enough that capability is a bad tie-breaker. What decides it is which one can reach your accounts, what it costs, and where your data ends up. For the tightest control Self-hosted and open source Hermes with open-weight models such as Llama, Mistral, Qwen or DeepSeek, running on your own hardware. No licence per seat, nothing leaving the building, and nobody able to change what it does without your permission. For the newest consumer agents What you already pay for If you already have a ChatGPT or Claude subscription, an agent inside it is capability you have bought and not used. That is usually the cheapest first move, because the cost is already committed. For the rest of your stack Wire them into the real work The bigger win is usually not one agent. It is your quoting, inbox and reporting talking to each other. That is closer to an integration project than a purchase, which is where most of the value is. Free, whether you buy or not Two guides that answer this properly. If these answer your question, you do not need me. That is the point of them. 12 pages · PDF Your First AI Employee What to automate first, in what order, and how to tell whether a task is actually a good candidate before you spend a day on it. Includes the prompts and the checks. Download the PDF 11 pages · PDF Set Up Your Own AI Agent Step by step through building an agent you run yourself, choosing between the hosted options and an open-source setup, and the mistakes that cost people the weekend. Download the PDF There are ten more guides on the site, including where AI actually helps and whether it is worth doing at all . Two of them answer the sceptical version of this question better than I can. Straight answers The questions worth asking first. What is an AI employee? An AI employee is a piece of software that takes a recurring job, works through it without being prompted at every step, and uses your own accounts and systems to do it. The name is a sales term rather than a technical one. Underneath, an AI employee is an AI agent: a model given tools, permissions and a memory of your process, then left to get on with a task. Can AI agents actually replace staff? Not in the way the marketing suggests. Independent testing found the best agents complete only a fraction of open-ended office work, and Gartner expects a large share of agentic projects to be cancelled by 2027. They replace defined, digital, measurable tasks rather than people. If the job involves judgement, relationships or physical work, it is not an AI employee yet. What can an AI employee do for a small business? The jobs that are high volume, digital, measurable and cheap to correct when one goes wrong. Enquiry handling, quoting from your own price list, customer support, CRM updates, meeting notes and reporting, and chasing invoices. A single mistake is a fix, not a crisis, which is exactly why these tasks suit an agent and physical or relationship work does not. Are AI employees worth it for a small business? If you own one of the platforms already and it can reach your accounts, they are worth testing, because the capability is already paid for. Expect to put in real effort up front: connecting the right accounts, writing the instructions and checking the output before you trust it. The gain is hours returned each week, not staff removed. What does an AI agent cost to set up? It depends entirely on what you point it at. A single agent reading email and updating your CRM is a much smaller job than one that quotes, books, chases payment and talks to customers. I scope after a free twenty-minute call, because the number is set by your process and your systems, not by a package price. Which AI agent should I choose? Pick on access rather than capability, because capability is now similar across them. If you can reach a self-hosted open-source agent on your own infrastructure, there is no per-seat licence and your data stays on your machine. If you want the newest consumer agents first, note that ChatGPT Dots is not available in the UK and Meta Muse launched in the US only. What happens on the call Bring the workflow, not the tool. 01 You bring one process The one eating the most hours or causing the most errors. Not a tool you read about, a job you actually want gone. 02 I tell you if it is a fit Including when it is not. Some workflows are cheaper to fix with a checklist than with software, and you will hear that on the first call. 03 You get a plain-English scope What gets built, on what, and what it will not do. If it is not worth doing, that is the answer and there is nothing more to pay. If you would rather learn first, start with AI training and enablement , or read how Bristol service businesses are already using agents . Bring me the workflow you wish would go away. Twenty minutes, no pitch, and a straight answer on whether AI is the right fix. Book a free 20-minute call]]></content:encoded>
  </item>
  <item>
    <title>Video SEO for Local Business · YouTube Channel &amp; GBP Video</title>
    <link>https://expedientai.co.uk/video-seo/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/video-seo/</guid>
    <pubDate>Mon, 12 Oct 2026 00:00:00 +0000</pubDate>
    <description>Video SEO for UK trades and local businesses. Three 60-90 second videos on your Google Business Profile and a YouTube channel, tuned for the searches your customers actually type. Filmed £897 or AI-produced £397.</description>
    <content:encoded><![CDATA[Video SEO for local business Video is the second step of the journey. Customers ask ChatGPT or Google for a shortlist, then watch the recommended business before they pick up the phone. AI Overviews cite YouTube in nearly a third of local-intent answers, and a Google Business Profile with a video gets five times the calls of one without. If you have no video, you are invisible at the moment the decision is made. Book a call See the packages Filmed or AI-produced Uploaded to GBP and YouTube Channel setup included Two routes Three videos either way. How we make them is your call. Both routes deliver the same thing: three finished short-form videos, titled and tagged for the searches your customers actually make, on your Google Business Profile and your YouTube channel. The filmed route captures your actual work. The AI route builds from what you already have. AI video · £397 No filming. Real videos. From what you already have. We script three short-form videos from your existing photos, your website copy, and the kind of jobs you do day to day. AI tooling produces a 60-90 second finished video for each: motion graphics, captions, voiceover (real voice or AI voice, your call), and titles and tags tuned for local search. Five days from intake. Best for businesses that want to test the video surface cheaply before committing to a filmed day. If the AI videos pull, most upgrade to the filmed package later. Book the AI route Filmed video · £897 One half-day. Three videos. Yours forever. A local camera crew spends half a day with you on a real job or in your workshop. Three finished videos: typically one "what we do" intro, one "how we do it" process, and one "before/after" or customer question. Edited, captioned, with titles and tags tuned for local search. Crew sourced through our partner network (Bristol, Bath, Weston-super-Mare, Stroud, Cheltenham, Swindon, plus UK-wide on request). One week from booking to delivery. You keep the raw footage. Book the filmed route Both routes include Local-search titles, descriptions and tags written for the terms your customers actually type Uploaded to your Google Business Profile and your YouTube channel A 30-minute YouTube channel setup call, so the channel exists with the right categories and a real upload Captions on every video, because most local searches happen on mute Why this is also AI search Video helps you twice. It helps the older search, where your customers find you on Google Maps. It also helps the newer one, where they ask an AI who to call and the answer cites video rather than a listing. For Google Maps A profile with video gets five times the calls. A Google Business Profile with a video is a different listing to one without. It ranks better in the Map Pack, it shows in the image carousel, and it converts better because a prospect can see the work before they call. This is the most direct return on the money. For AI answers AI Overviews cite YouTube in local searches. When someone asks an assistant for a recommendation in your area, YouTube is frequently one of the sources it reads. A channel with captioned, titled, local-search-optimised videos is a citable source that a business listing is not. This is the part most people have not thought about, and it is why the captions and titles are tuned rather than generic. Making your site readable to AI agents is a separate discipline, and it is where most of the search is going. If that is your problem rather than your local ranking, see Agentic Engine Optimization . If it is your Map Pack, see Local SEO . How it works Five days from brief to uploaded. 01 Channel setup call. 30 minutes. We set up or tidy your YouTube channel, set the right categories, write the channel description around what you actually do and where you work, and get a real first upload on it. So the channel exists before anyone searches for you. 02 Scripts and titles, written for search. Three short scripts, each built around a term your customers type rather than a phrase you like. Titles, descriptions and tags written for those terms. You see them before we shoot or produce, because getting a title wrong is expensive to fix after filming. 03 Shoot or produce. Filmed route: half a day on location, we direct and capture. AI route: produced from your existing assets with motion graphics and captions. Either way you keep the raw material. 04 Edit, caption, upload. Finished to 60-90 seconds, captioned, uploaded to both your Google Business Profile and your YouTube channel, with the metadata written in step two. The honest bit What this will not do. It will not make you number one. Video is one ranking factor among many for the Map Pack, alongside proximity, reviews and profile completeness. If your reviews are weak or your category is wrong, video will not fix it, and I would rather fix those first. The £149 audit tells you which is which before you spend on video. It will not go viral. Three local videos do not make you a creator. They are built to be found by people already searching for your trade in your area, which is a smaller number of people than a viral video and a far better one. If you want reach rather than calls, this is the wrong spend. Start with the audit, not the video. If you are not sure whether video is your problem, the £149 Local SEO audit will tell you where you actually stand on Maps, what your competitors have, and whether video is worth doing at all. If it is, the video packages are above. Book a call]]></content:encoded>
  </item>
  <item>
    <title>Agentic Engine Optimization (AEO) · Make Your Site Readable to AI Agents</title>
    <link>https://expedientai.co.uk/agentic-seo/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/agentic-seo/</guid>
    <pubDate>Mon, 12 Oct 2026 00:00:00 +0000</pubDate>
    <description>AEO audit for UK businesses. Can AI agents and answer engines actually find, read and use your site? Fixed-price audit and fixes for robots.txt, llms.txt, structured data and agent readability.</description>
    <content:encoded><![CDATA[Agentic Engine Optimization Can an AI agent actually read your site? Agents fetch, parse and reason over your pages. Most of the time they succeed quietly. When they do not, nothing breaks, no error appears, and no analytics event fires. They simply skip you, or they answer from bad assumptions. This audit tells you which of the two is happening. Book the audit Try the free check first Fixed price Plain-English report No jargon What this is SEO taught us to be found. AEO is being used. The term comes from Addy Osmani, Director of Engineering at Google Cloud AI, who published an Agentic Engine Optimization framework in April 2026. He is explicit that this is a practitioner's view from inside the developer-tools world, not a Google-wide recommendation. That caveat matters, and we will not pretend otherwise when we sell this to you. The difference A search crawler indexes you. An agent reads you. A crawler fetches pages and stores signals. An agent fetches pages and tries to do something with them: answer a question, complete a task, or write code against your API. It reads the whole page in one or two requests, counts tokens, and decides whether the content fits its context window. If it does not fit, it may silently truncate it, skip it, or fall back on assumptions and invent something. Why you would not notice The agent leaves no fingerprint in your analytics. There is no scroll depth, no time on page, no button clicks, no form fills. A human takes four to eight minutes navigating your site; an agent compresses that into one GET request and moves on. Your funnel shows nothing, which is exactly why this goes unnoticed for years. The six checks What the audit actually tests. Each check maps to something an agent needs in order to use your site. If any of them fails, you are invisible to the part of the market that is already asking an assistant rather than a search box. 01 Access control Does your robots.txt let AI crawlers through, or quietly block them? A block means total invisibility with no error message. We check GPTBot, ClaudeBot, PerplexityBot, Google-Extended and the rest against your actual file. 02 Discovery Do you have an llms.txt? A flat Markdown directory that tells agents what you do and where to find it, so they do not have to crawl your whole site guessing. Cheap to add, most sites do not have one. 03 Token efficiency How big is each page in tokens? Agents working in the 100K to 200K context range will skip or truncate anything oversized. We measure your real pages and tell you which ones are too heavy to be read in full. 04 Machine-readability Is your content parsable without a visual layout? Does it need JavaScript to render? Agents do not see your design. They see structure, headings and text. We check whether your key content survives that translation. 05 Capability signalling Does anything tell an agent what you can actually do, in a format it can read? An agent that has to infer your services from marketing prose is guessing. A structured capability file is not guessing. 06 Structured data Is your schema.org markup complete and valid? It is the cheapest thing an agent can read about your business, because it is machine-written. Gaps there are the most expensive omissions, since it is where you would have been found. Where to start Do the free check first. The audit below is fixed price and covers all six checks, including the two the free tool cannot do: reading your robots.txt and fetching your llms.txt. Start with the free check, and only pay if it tells you something you did not already know. Free · no login Website AI Visibility Checker Scores one page for the four checks it can judge from HTML alone: structured data, headings, metadata and content depth. Nothing to sign up for. Most people start here and never need the paid audit. Run the free check Fixed price The full AEO audit Every check run against your live site, with a written report explaining what is failing, why it matters to AI agents specifically, and what to change. Plain English, no jargon, and yours to keep whether or not you do the work with us. Book the audit Who this is for The honest version. It is for you if... You get asked for recommendations rather than found through search. You have noticed referrals arriving from ChatGPT or Perplexity with no obvious source. You have a website that looks fine to you and you have no idea whether anything can read it. You are the kind of business an assistant would recommend if it had a clean enough picture of you. It is probably not for you if... You get all your customers from referrals and word of mouth, and you have no intention of changing that. You need more local Map Pack visibility, which is a different problem with a different answer. You want to be told that AI will replace your website. It will not, and this is not that pitch. Looking for the other half of the picture? Video SEO covers the surface AI answers most often cite for local searches, and Local SEO covers ranking in the Map Pack itself. Start with the free check. Two minutes, no login, and it tells you whether your page is readable. It cannot see your robots.txt or llms.txt, so a clean score is a good sign, not a guarantee. We will tell you where the line is. Run the free check]]></content:encoded>
  </item>
  <item>
    <title>Running ads without generative AI, for brands whose customers would notice</title>
    <link>https://expedientai.co.uk/feed/marketing-without-generative-ai/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/marketing-without-generative-ai/</guid>
    <pubDate>Sat, 03 Oct 2026 00:00:00 +0000</pubDate>
    <description>How to run paid ads with no AI-generated copy and no AI-generated imagery, including the platform settings that switch themselves back on. For artists, makers and handmade brands.</description>
    <content:encoded><![CDATA[Marketing Running ads without generative AI. If you make the work by hand, you should not have to hand it over to a machine to sell it. Here is exactly how we run paid ads for artists and makers with no AI-written copy and no AI-generated imagery, what it costs, and where it genuinely does not work. 2026-10-03 8 min read Marketing We learned this properly working with Jim'll Paint It. When the whole value of a brand is that a real person sat down and made the thing, using a model trained on other people's work to advertise it is not a bad look. It is a contradiction the audience will spot immediately, and artists' audiences are, in our experience, the most switched-on audience on the internet when it comes to spotting AI slop. So we made it a standing option, and for some clients a standing rule. If you want that written into the engagement, we put it in writing. It costs us nothing to commit to, because it is how we prefer to work anyway. Why artists are right to be wary The objection is not technophobia. It is specific, and it holds up. Training data. Image models were trained on scraped work, largely without consent, credit or payment. Using one to advertise an artist's work launders that problem straight into their brand. It undercuts the offer. If your pitch is "a human painted this", an AI-generated ad image tells the buyer the human part is negotiable. Audiences retaliate. Art communities are fast and vocal. One AI-looking ad in the feed can turn a comment section into a pile-on that costs more than the campaign earns. Voice matters more than volume. Most artist brands sell on a distinctive voice. Generative copy flattens it into the same cheerful mush every other brand is running. What "no generative AI" actually means in practice It is worth being precise, because "no AI" is a broad claim and the platforms are full of AI whether you like it or not. Here is where the line gets drawn. What we do not do. No AI-generated or AI-extended images, backgrounds or product mockups. No AI-written ad copy, headlines, captions, emails or blog content for the brand. No AI-generated avatars or synthetic spokespeople. No AI voiceover pretending to be the artist. No feeding the artist's catalogue into an image model for "variations". What we do use. The platforms' own delivery algorithms, because Meta and Google are machine learning end to end and you cannot advertise there otherwise. That is bidding and targeting, not generation. Advantage+ and Performance Max with the generative features switched off. Analytics, forecasting, spreadsheets and dashboards, which is maths rather than generation. Ordinary photo editing: cropping, colour correction, retouching a dust speck, nothing that invents content that was not photographed. The setting that quietly switches itself back on This is the practical bit that catches people out. Meta and Google will auto-generate text variants, expand images and swap backgrounds unless you tell them not to. We switch it off per campaign, document it, and re-check it after every platform update, because it has a habit of switching itself back on. If you do not audit this monthly, you do not actually have a no-generative-AI account. You have one that was true for about six weeks. Where the creative actually comes from This is the part people assume is the problem. It is not. Artists have an advantage most e-commerce brands would pay a fortune for, and it is an endless supply of genuinely interesting raw material. Process footage. The single best-performing format for artist brands is watching the thing get made. Phone on a tripod, timelapse, thirty seconds. It outperforms polished studio work consistently, it is free, and it is the one thing a model cannot fake. The artist talking. Unscripted, badly lit, one take. Why they made it, who asked, what went wrong halfway through. The rougher it is the more it reads as real, which is exactly the signal this audience scans for. The archive. Most artists are sitting on years of work, commission briefs, sketchbooks and rejected drafts. That is a creative library. We audit it, catalogue it and build a rotation from it before discussing anything new. Real customer content. The work on real walls, in real houses, with a real caption from the person who bought it. Ask at delivery, offer nothing more than a repost, and most people say yes. Copy from the artist's own words. Their captions, their emails, their replies to commission enquiries. The voice already exists. The job is editing, not inventing. Does it perform worse? No, and that is the practical argument rather than the ethical one. On the accounts we run this way, human creative built from process footage and the artist's own voice consistently beats slick generic assets. It stops the scroll differently, because it does not look like an ad. It carries proof: watching a piece get painted is evidence, where a rendered mockup is only a claim. And it fits the buying decision, because nobody buys original art on price. They buy on trust in the maker, so the creative has to build trust in the maker. Where the rule does cost you It costs you volume, and I would rather say so than sell it as costless. You cannot spin out forty creative variants a week from a prompt. So the testing plan has to be tighter: fewer concepts, held longer, judged on contribution margin rather than on how many hooks you burned through. We plan around that from day one. Fewer, better, longer. It also happens to be cheaper. How we set the account up Creative audit first: everything already shot, painted, filmed or written gets catalogued into a usable bank. Platform settings lockdown, generative enhancements disabled across Meta and Google, documented, re-audited monthly. Then a capture routine rather than a shoot, which is one timelapse, one talking clip and one finished-piece photo a week. Ten minutes, not a production day. Profit-first measurement, because originals, prints and commissions have wildly different margins and scaling prints should not quietly cannibalise the commission pipeline. And slow, deliberate testing: fewer concepts, longer read windows, decisions on profit rather than on ROAS theatre. Who this is for Artists, illustrators, printmakers, ceramicists, custom commission studios, small handmade brands, and anyone whose audience would be actively offended by a generated image. It works particularly well for people who start on Etsy or the marketplaces and build their own shop from there. It is also for brands who simply do not want their name attached to it. You do not need an ethical manifesto. "I don't want it, don't use it" is a complete instruction. If you want help with this It is not complicated to do badly and it is not complicated to do well. What is costly is finding out six weeks later that the platform quietly re-enabled the thing you switched off, or that your creative library was never built, and you have been running generic assets the whole time. That work runs through Magic Dolphin, where I run the ads and the measurement side. This is a fixed-scope engagement rather than something I will try to sell you AI for. If you would like someone to audit an account and tell you exactly what is live on it, book a call and say you want the no-generative-AI audit. The full write-up of the method is on the Magic Dolphin blog , and the client results are on this site under Magic Dolphin . Frequently asked questions Can you really run Google and Meta ads without any AI? Without generative AI, yes. You cannot opt out of machine learning entirely, because both platforms' bidding and targeting are machine learning and you cannot advertise there otherwise. That is delivery, not generation. What you can switch off is the generative layer: auto-written text variants, image expansion, background swapping on Advantage+ and Performance Max, synthetic avatars and voiceovers. Does avoiding generative AI hurt performance? On the accounts we run this way it does not, and for handmade work it usually helps. Process footage and the artist's own voice consistently outperform generic assets, because they carry proof rather than a claim. What you do lose is volume: you cannot produce forty creative variants a week, so testing is planned around fewer concepts held longer. What is the most common mistake people make? Believing they have a no-AI account when they do not. Meta and Google re-enable generative features after platform updates. If it is not documented and re-audited monthly, you have a no-generative-AI account that was true for about six weeks. Do you charge more for this? No. Committing to no generative AI in your marketing costs us nothing, because it is how we prefer to work. If you want it written into the engagement so nobody has to ask, we will put it in writing.]]></content:encoded>
  </item>
  <item>
    <title>ChatGPT Dots: what OpenAI's always-on agents mean for a small business</title>
    <link>https://expedientai.co.uk/feed/chatgpt-dots-small-business/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/chatgpt-dots-small-business/</guid>
    <pubDate>Thu, 01 Oct 2026 00:00:00 +0000</pubDate>
    <description>OpenAI launched Dots at DevDay: AI agents with their own cloud computer. What they can do for a small business, why they are not available in the UK, and where the real cost and risk sit.</description>
    <content:encoded><![CDATA[Tools ChatGPT Dots, and what they mean for a small business. OpenAI launched Dots at DevDay on 29 September 2026. They are always-on AI agents with their own cloud computer, connected to 4,000+ apps, that keep working when you close the tab. Useful, expensive, and not available in the UK at launch. Here is the honest read. 2026-10-01 7 min read Tools OpenAI announced Dots on 29 September 2026. The idea is simple: instead of asking ChatGPT a question and getting an answer you then have to act on, you hand it the job. Each Dot gets its own virtual machine in the cloud, browses the web, opens applications, and works through the connected services you have given it. You are not giving it your laptop. It runs somewhere else, which is the whole point. That is a genuinely different shape of tool from the chatbot in your browser tab. And for a small business, it is worth spending twenty minutes understanding what it does, what it costs, and why you may not be able to switch it on at all. 1. What a Dot actually is Each Dot is a persistent agent. It runs continuously rather than waiting for a prompt, it works inside an isolated cloud machine, and it keeps going when you are not looking at ChatGPT. You can reach it through the web app, and OpenAI has said voice calls, Slack and Microsoft Teams are coming, with SMS planned. Underneath sits GPT-6 Astra. The Dot connects to more than 4,000 apps through plugins, which means it can work across the tools you already pay for rather than being sealed inside ChatGPT. The practical difference is that a Dot can finish the job. Ask ChatGPT where to order dinner and it suggests somewhere. Ask a Dot and it builds itemised cart comparisons across delivery services, accounts for local tax and tip, and sends them back for you to choose. 2. What that looks like for a small business The examples OpenAI gave are deliberately unglamorous, and they are the right ones. A Dot that turns an interview recording into social captions and show notes. A Dot that spots an invoice you forgot to send, prepares it, and puts it in front of you for approval. A Dot that keeps tabs on supplier pricing while you do something else entirely. For a small business the honest version is narrower than the launch material. The work that suits this shape is the work that is repetitive, spans two or more systems, and has an obvious finish line. Chasing an unpaid invoice. Getting a quote out the same day. Turning an enquiry into a booked job. Nightly reporting that nobody had to ask for. What it is not good at, yet, is anything where a wrong answer costs you money or a customer. A Dot that quietly books the wrong slot, or sends a price to the wrong client, is a bigger problem than a person making the same mistake, because the person would probably have checked. 3. The part that stops most UK businesses: it is not available here OpenAI's own help documentation says Dots are rolling out to Pro users excluding the European Economic Area, Switzerland, and the UK . Business Premium users get them across supported ChatGPT regions. Enterprise, Edu and Healthcare are in beta and off by default until a workspace admin turns them on. So if you are reading this from the UK on a Pro subscription, there is nothing to switch on yet. That is not a technicality, it is the single most important fact about Dots for this audience. The rest of this article is still worth reading, because the shape of the tool tells you where the industry is going, and because OpenAI has said it plans to widen access. 4. What it costs There is no separate price for Dots. Your first one is included with ChatGPT Pro, which starts at $100 a month, or with Business Premium seats. Free and Plus do not get Dots at launch. Additional dots are promised later, along with options to trade speed against monthly workload. Put that next to the alternative and the picture is clearer. A small business can get most of the same operational benefit from a properly configured setup on open-source and open-weight models running on hardware it already owns, for a monthly bill measured in tens rather than hundreds. The Dot is a different product, not a better one. It buys convenience and a large app ecosystem, and it charges you in cash and in dependency. 5. Why this can genuinely overwhelm a small business I have watched owners read a launch like this and conclude they are now behind. Usually they are not behind, they are looking at a tool built for a different size of company. The overwhelm has three sources, and they are worth naming because each has a different answer. Volume. Four thousand connected apps is a decade of accumulated enterprise tooling. A twelve-person firm uses maybe forty. Installing integrations you will never open is how you create an admin burden instead of removing one. Permission. An agent that browses, buys, books and sends is only as safe as the permissions you gave it and the oversight you built around it. That oversight is real work, and it is the part no launch demo shows you. The Your Data, Your Rules guide covers how to write the rulebook your team can actually follow. Cost ambiguity. Pro at $100 a month is not a rounding error for a business doing forty thousand a year, and it is easy to adopt several tools at once and arrive at a bill nobody can attribute to anything. The free ROI calculator puts a number on the hours first, so the tool has to beat a figure rather than a feeling. 6. What I would actually do this week Nothing, until you know whether you can even get it. Then, in order: Take the AI readiness scorecard . Ten questions, one honest score, free. Most businesses that score badly do not need a better tool, they need one workflow picked properly. Run the ROI calculator on a single workflow. If the answer is a few thousand pounds a year, that is the number the tool has to beat. Read the Where AI Actually Helps guide. It maps which tasks are worth building against and which are not yet, which is the question a launch announcement is worst at helping you answer. Then bring that one workflow to a twenty-minute call. You get a clear yes or no, and if the answer is that Dots, or an agent of any kind, is not the right tool yet, you will hear that instead of a proposal. 7. The honest take Dots are a good idea executed by people who know where the industry is going. The always-on agent with its own machine is the right architecture for work that spans systems and does not stop when you close the laptop. That is a real capability, and it will get cheaper. But for a UK small business today it is a preview of a direction, not a purchase decision. The capability is not available to you, the price point is aimed at a different size of company, and the thing that actually makes agents useful, a named person who owns the workflow and checks the output, is the part you cannot buy off a shelf. Get that right first, and the tool choice becomes obvious. Get it wrong, and no agent will save you. Frequently asked questions What are ChatGPT Dots? Dots are always-on AI agents announced by OpenAI on 29 September 2026. Each one runs on its own virtual machine in the cloud, browses the web, connects to more than 4,000 apps through plugins, and keeps working when you are not in the ChatGPT tab. They run on GPT-6 Astra. Can I use ChatGPT Dots in the UK? Not at launch. OpenAI's help documentation states Dots are rolling out to Pro users excluding the European Economic Area, Switzerland and the UK. Business Premium users get access across supported ChatGPT regions. OpenAI has said it plans to widen access, but there is no date for the UK. How much does ChatGPT Dots cost? There is no separate price. Your first Dot is included with ChatGPT Pro, which starts at $100 a month, or with Business Premium seats. Free and Plus plans do not include Dots at launch. Additional Dots are promised later with options to trade speed against workload. What can ChatGPT Dots do for a small business? The realistic near-term use is repetitive work that spans more than one system and has a clear finish line: chasing unpaid invoices, turning an enquiry into a booked job, same-day quoting, overnight reporting. It is not yet suited to anything where a wrong action costs money or a customer without a human approving it first.]]></content:encoded>
  </item>
  <item>
    <title>AI Training for Your Team · Get More From the AI You Already Pay For</title>
    <link>https://expedientai.co.uk/ai-training/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/ai-training/</guid>
    <pubDate>Wed, 30 Sep 2026 00:00:00 +0000</pubDate>
    <description>AI training for UK small businesses. Get your team using ChatGPT, Claude and Gemini properly, and run Hermes with open-source models like DeepSeek, MiniMax and Qwen locally.</description>
    <content:encoded><![CDATA[AI training & enablement You already pay for AI. Now make it work. Most teams I meet use AI like a faster search box. A little prompting, a bit of rewriting, back to the spreadsheet. The tools are capable. The gap is training, not software. I fix that, and where the job needs more than the tools give you, I build the system. Book a call See the questions For non-technical teams Your tools, or mine, or both On your site or remote The offer Training, tools, and the system when you need one. 01 Get more from what you have. ChatGPT, Claude, Gemini, Copilot. If your team already has them, this is where the return is. Context about your business, reusable instructions, connecting it to your inbox and docs, and the habits that turn a lucky afternoon into an actual process. 02 Or set up your own, locally. Hermes as the agent layer, running open-source and open-weight models such as DeepSeek, MiniMax, Qwen and Llama, hosted on your hardware. No per-seat cost, your data stays in the building, and nothing gets switched off by someone else's pricing decision. 03 Or I build the system. When the tools are not enough, the workflow gets built properly. That is where twenty years of running companies comes in: I know where the hours actually go, what your people will genuinely use, and what a process costs to run at volume. What I know how to do Across both sides of the fence. Most people who train you use one or two of these. I use all of them, and I can tell you which is right for the job in front of you rather than selling you the one I happen to prefer. Open source, self-hosted Your models, your machine. Hermes and OpenCode running DeepSeek, MiniMax, Qwen, Llama, Mistral and the rest on local hardware or a box I set up for you. Setup, configuration, upgrades, and the training so your team is not frightened of it. For volume work, private data, and anyone who has been burned by a price change. The tools everyone knows Used properly, not just used. ChatGPT, Claude, Claude Code, Gemini and Google AI Studio, Grok, and Perplexity Computer. ChatGPT Business and Claude's team plans, structured properly with shared instructions, projects, and the admin side set up so it does not become a shadow IT problem. Not sure which is right for you? That is the first conversation, and it is free. Twenty minutes on a workflow you are trying to improve, and you leave with a clear answer either way. The questions I actually get Ask me these, or ask me yours. These are the real ones from the last few years of small businesses figuring out AI. If yours is not here, it is probably better than the ones that are. Usage How do I get more out of the ChatGPT we already pay for? This is the most common question I get. The answer is almost never a new tool. It is context, reusable instructions, and a decision about where the work lives so two people doing the same job stop getting two different results. Safety What can we safely put into these tools? More than most people assume, less than vendors would like. Customer names, contracts and pricing need handling; a rewritten product description usually does not. I give you a plain rulebook your team can follow, which is the Your Data, Your Rules guide in short form. Capability Can it do our job, or only the easy 10%? Honest answer per task. Some of it is fully handled today. Some is a good first draft. Some is not worth automating yet, and I will say so rather than build something that quietly underperforms. The Where AI Actually Helps guide covers which is which. Staff What do we tell staff who worry it replaces them? That question gets a straight answer, because getting it wrong causes real damage. In practice AI removes the boring parts of a role and shifts what the role is. That conversation is part of the work, not a footnote to it. Cost Is it worth it, or are we being sold a platform? Sometimes it is not, and I will tell you on the first call. Before spending anything, the AI automation ROI calculator sizes the hours, and the Is This Worth Doing guide walks the decision. Both are free. Lock-in What happens if the tool we pick gets expensive or disappears? That is the argument for self-hosting in one line, and it is why open source is my default. The AI vendor interrogation kit asks the questions that tell you how exposed you are. What you get to keep Guides, tools and references. Yours whether you buy or not. Everything below is free and public. Reading these first is genuinely the right thing to do, whether or not you ever work with me. If they answer your question, you do not need me. Free tool AI Council Six AI advisers argue your next decision from different angles before you spend on it. Useful on its own, and the fastest way to see what these tools are actually capable of. Free tool AI Readiness Scorecard Ten questions, one honest score. Where you actually are before you start, so the plan is built on the business you have rather than the one in the pitch deck. Free tool AI Automation ROI Calculator Size the hours a workflow costs you in pounds. The number makes the internal conversation much shorter. Eight of the twelve guides, download them here 01 Where AI Actually Helps Which tasks are worth automating and which are not. Start here. Download the PDF 02 The 30-Day AI Pilot The shape of a first project that finishes in a month instead of a year. Download the PDF 03 Your Data, Your Rules What can go into these tools and what should not. Download the PDF 04 Writing AI Instructions So everyone on the team gets the same result from the same prompt. Download the PDF 05 AI Vendor Interrogation Kit The questions that show you how exposed you are to a price change. Download the PDF 06 Your First AI Employee Setting up the named owner who makes the system work day to day. Download the PDF 07 Set Up Your Own AI Agent Running an agent on your own machine, for when the data cannot leave. Download the PDF 08 Is This Worth Doing The decision, before you commit money to it. Download the PDF All twelve guides, including build-the-thing, get-found, get-the-conversation and close-the-deal → How it works Three ways to work together. 01 Consultation. Twenty minutes, free. Bring one workflow that eats your team's time. You leave with a clear yes or no on AI for it, what it would take, and one thing you can run yourself today. No pitch, no obligation. 02 Training. Your team, your tools. A working session built around what your people actually do, on what you already use or the stack I set up for you. Practical, on your own examples, with the written material to keep afterwards. 03 Build. When the tools are not enough. The full system: workflow mapped, built, handed over, with your team trained to run it. Fixed price, and you own the system, the data and the roadmap when it is done. See the full service list . Start with one question. Pick the question on this page that matters most to you, and ask it on the call. If the answer is that you do not need me yet, that is a genuinely useful twenty minutes and I will tell you. What helps me help you: the workflow, who does it, roughly how many hours a week it takes, and what you have already tried. Book a call]]></content:encoded>
  </item>
  <item>
    <title>Local SEO for UK trades: four ways to win the Map Pack</title>
    <link>https://expedientai.co.uk/feed/local-seo-managed-and-video/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/local-seo-managed-and-video/</guid>
    <pubDate>Wed, 23 Sep 2026 00:00:00 +0000</pubDate>
    <description>Four ways to win the Map Pack if you run a UK trades business. £149 audit, £497 setup, £150/mo monthly, or from £397 with video.</description>
    <content:encoded><![CDATA[News Local SEO for UK trades: four ways to win the Map Pack Four ways to win the Map Pack if you run a UK trades business. £149 audit, £497 setup, £150/mo monthly, or from £397 with video. 2026-09-23 · 5 min read Your phone hasn't rung this week and you don't know why. You have a Google Business Profile. It says the right things. But your competitors are in the Map Pack at the top of the page and you're not, so the customers searching for an emergency plumber in your town right now are calling them, not you. The page at expedientai.co.uk/local-seo/ now offers four ways to fix that, depending on how much of the work you want to hand over. If you just want to know where you stand £149 audit. One page, plain English, 5 working days. We check your Google Business Profile score, your Map Pack position for the searches your customers actually type, your review count and velocity, your citations, and the local signals on your website. Top 3 quick wins ranked by impact. Yours to keep, act on, or hand to a web agency. Best if: you want to know where you stand before spending anything else. If you know what to fix but don't want to do it yourself £497 setup, one-off. We claim and verify your Google Business Profile if it isn't already, fill it out properly (services, hours, photos, Q&A), build your citations across the 12 directories that actually move the needle, fix any NAP drift on your existing listings, add LocalBusiness schema to your site, and embed the map on your contact page. Two weeks from intake. Best if: you've had a look at your local SEO and you know what needs doing, but you don't have a Saturday between jobs. If you've already won the Pack and want to stay there £150/month. We run your Google Business Profile every month: 2-4 posts, a photo refresh, Q&A seeding, review responses (you send us the new ones, we reply). Citation drift monitoring across the 12 directories. Monthly rank report for your main keywords with your Map Pack position and a 1-line action note. No contract. Cancel any month after the first. Best if: you've got into the Map Pack and you're worried about losing the position because you don't have time to keep the profile active. If you want to be found on YouTube as well as Maps From £397. Two routes. The AI route: we script and produce three short-form videos from your existing photos and website copy. No filming. 60-90 seconds each, with motion graphics, captions, voiceover (yours or AI, your call), and titles and tags tuned for the local searches your customers type. Five days from intake. £397. The filmed route: a local camera crew spends half a day with you on a real job or in your workshop. Three finished videos, edited, captioned, with the same local-search tuning. You keep the raw footage. One week from booking to delivery. £897. Both routes include a 30-minute YouTube channel setup call, so your channel actually exists, has the right categories, and starts with a real upload. Uploaded to your GBP and YouTube. Best if: you've noticed that more of your customers are asking ChatGPT for shortlists, and they're watching the recommended tradesperson on YouTube before they pick up the phone. What hasn't changed The £149 audit is still £149, still one page, still no upsell. The honest bit at the bottom of the page still applies: if you don't have a website yet, start with Website Build from £300 . If you serve the whole UK from one office, Map Pack ranking isn't the right channel for you. Adding the three new options doesn't change who the page is for. You can stack any combination. Most people start with the £149 audit, then add either the £497 setup or the £150/month once they've seen the report. The video option works either before or after the audit. Where to start If you've never done a Map Pack audit before, the £149 read tells you exactly what's wrong and what to fix first. If you've already had a look and you know what needs doing, the £497 setup is the shortest path from where you are to in-the-Pack. If you're already in the Pack and you're losing ground, the £150/month keeps the work going without you having to do it. expedientai.co.uk/local-seo/ has the full breakdown.]]></content:encoded>
  </item>
  <item>
    <title>How to use AI in your UK small business without buying the wrong thing &amp;middot; Expedient AI</title>
    <link>https://expedientai.co.uk/feed/how-to-use-ai-in-your-uk-small-business/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/how-to-use-ai-in-your-uk-small-business/</guid>
    <pubDate>Tue, 22 Sep 2026 00:00:00 +0000</pubDate>
    <description>A practical first-week playbook for UK small businesses adopting AI. If you have searched 'how to use AI for my business', 'help me implement AI in my company', or 'AI for my workflow', this is the order of operations. What to try first, what to defer, and what to skip.</description>
    <content:encoded><![CDATA[Tools How to use AI in your UK small business without buying the wrong thing If you have searched "how to use AI in my business", "help me implement AI in my company", or "AI for my workflow", this is the order of operations. The first week is for mapping the bottleneck and the second is for picking one workflow to ship. Everything else (the tooling, the platform, the model) follows from those two decisions, not the other way around. 2026-09-22 · 8 min read 1. Stop looking at the tools. Map the bottleneck. The first search most operators do is for the tool. The right first search is for the bottleneck. Which workflow in your business costs the most hours, produces the most errors, or is the reason another hire keeps getting delayed? That answer drives everything else. The tool follows from the workflow, not the other way around. Concrete first-week exercise: write down the five workflows that take the most hours in your operation. For each, note the cost per hour and the number of judgement calls per task. The workflow with the highest cost-per-hour and the fewest judgement calls is your first AI candidate. The others are not, even if they look more interesting. The free AI automation ROI calculator sizes the cost-per-hour in pounds, and the is this worth doing PDF guide walks through the same exercise with worked examples. 2. Pick one workflow. Ship it in weeks, not quarters. The single biggest predictor of whether AI lands in a UK small business is the size of the first project. Big pilots fail. Small, scoped workflows ship, get used, get measured, and become the case for the next one. A workflow that takes six to eight weeks to build and starts paying back inside three months is the right shape. The 30-day AI pilot PDF guide is the matching cadence for the first build, and the build the thing guide is what to do once the pilot is signed off. For most operators I work with, the first shipped workflow is one of four: enquiry handling across web chat and WhatsApp, a quote generator for a trades business, a customer-service reply workflow that escalates the hard ones to a human, or a Monday-morning report that summarises the week's pipeline. Each one is small enough to ship, large enough to measure, and useful enough to keep. 3. Buy the smallest tool that does the job. The tooling landscape is loud. Most operators come to me having already paid for three or four AI tools that didn't deliver, and the pattern is the same: the tools were right, but they were bolted onto the wrong workflow. Buying tools first is how this happens. Picking the workflow first usually surfaces a smaller, cheaper, less glamorous stack that actually does the job. The AI vendor interrogation kit PDF guide is the checklist for spotting the difference. A working stack for a UK small business in 2026 looks like this in practice: an open-weight or hosted model for the routine work (the bulk of volume), a frontier model for the small percentage of tasks where stakes justify the cost, a workflow layer that ties them to your existing tools (CRM, calendar, email, spreadsheets), and monitoring with fallback paths so the system holds up on a Monday morning. The total monthly cost is usually under one hundred pounds for a single workflow at small-business volumes. 4. Make sure a real person owns it. AI projects without a named owner on the customer side tend to drift. The pattern is familiar: the consultant builds, the consultant hands over, the system sits unused, and six months later nobody remembers why it exists. The fix is a single named operator on your side who actually uses the system, gives the consultant feedback, and is the person the team goes to when something is unclear. It does not need to be a technical role. It needs to be the person who runs the workflow day to day, with the time and the mandate to make the system work. Without that, even a well-built system will sit unused. The your first AI employee PDF guide is the working draft of how to set that operator up for success. 5. Measure before you scale. The discipline that separates AI projects that survive from AI projects that die is measurement. Hours saved, errors reduced, calls answered, leads captured, quotes sent on time. Pick two or three numbers, write them down before the system goes live, and check them at one month and three months. If they do not move, the workflow was the wrong choice. If they do, the case for the next workflow writes itself. Most of the operators I work with skip this step and pay for it later. The numbers are what make the second and third AI workflow easier to justify internally. They are also what stops the first one from being quietly turned off in month four because nobody could tell whether it was working. The AI readiness scorecard is a free 10-question self-assessment that scores where your business sits before you start, and the ROI calculator is the post-launch check. What this looks like for a typical UK small business A trades business with three vans, an enquiry inbox, and a part-time office manager. The first shipped workflow is an enquiry handler that answers the WhatsApp and web-chat messages out of hours, books the visit into the calendar, and writes a clean handover to the office manager at 8am. The build takes about six weeks. The monthly cost is around forty pounds. The measurable result is fewer missed calls and a faster quote turnaround, both of which can be checked at the end of month one against the same week last year. A professional services firm with five people and a partner who still does the proposals. The first shipped workflow is a draft-and-review engine that turns a one-page brief into a structured proposal, with the partner in the loop on every output. The build takes about eight weeks. The monthly cost is in the same range. The measurable result is proposal turnaround time and win rate on the proposals the system touches. A small retail brand with a Shopify store and an inbox full of customer emails. The first shipped workflow is a customer-service reply handler across email and web chat that drafts a reply for every message and only escalates the small percentage that needs a human. The build takes about six weeks. The measurable result is first-response time and ticket resolution rate, both of which can be tracked inside the existing help-desk tool. What to skip on the first project Skip the platform that wants to be your single source of truth. Skip the model with the most impressive benchmark. Skip the consultant who wants to redesign your whole operation before shipping anything. Skip the multi-month roadmap that delivers a deck instead of a system. The first AI project in a UK small business should ship a working tool in weeks, on a stack you understand, against a number you can measure. The where AI actually helps PDF guide is the working map of which workflows are worth the first build and which ones are not. If a vendor or a consultant cannot tell you, in plain English, what the first workflow will be, how long it will take, what it will cost to run, and how you will know it worked, the answer is no. There will be a better proposal behind that one. The honest version Most UK small businesses that adopt AI well start small, ship fast, and measure honestly. The ones that fail start with the tool, the platform, or the multi-workflow roadmap, and they pay for the bloat in months when the system is unused or the bill is unrecognisable. The first project is the one that pays for the second and the third. Make it small enough to ship, scoped enough to measure, and boring enough to actually run on a Monday morning. Tools and guides referenced in this post AI automation ROI calculator , free, no login. Size the cost-per-hour of your bottleneck in pounds. AI readiness scorecard , free 10-question assessment. Where does the business actually sit before you start. AI Council , six AI advisers argue your next decision through different perspectives, before you spend on it. Is this worth doing , first-week decision guide. Download the PDF . 30-day AI pilot , the matching cadence for the first build. Download the PDF . Build the thing , what to do once the pilot is signed off. Download the PDF . AI vendor interrogation kit , the checklist for spotting the difference between a tool that works and one that demos well. Download the PDF . Your first AI employee , the working draft of how to set up the named operator for success. Download the PDF . Where AI actually helps , the working map of which workflows are worth the first build. Download the PDF . Frequently asked questions How do I start using AI in my small business? Map the bottleneck first, not the tool. Pick one workflow that costs the most hours or produces the most errors, write down what good looks like, and only then ask whether AI is the right answer. Most first AI projects fail because the workflow was the wrong starting point, not because the tool was wrong. What is the first AI task a UK small business should automate? The one with the highest cost-per-hour and the lowest judgement-per-step. Enquiry handling, quote generation for trades, customer-service reply drafting, lead list building and Monday-morning reports are the four I see work first time. Anything that needs a human in the loop at every step is the wrong place to start. How much does it cost to implement AI in a small business? Less than most operators expect if you start narrow. The first workflow can be built and shipped for a few thousand pounds in consulting fees and run on an open-weight model stack for under forty pounds a month. The cost climbs when scope grows, so the discipline is to ship one workflow, measure it, then decide whether the next one is worth the build. Do I need technical staff to use AI in my business? No. The systems that work for UK SMEs are built around the team that already runs the workflow. Training and runbooks are part of the handover. The aim is a system your team trusts and can use on a Monday morning, not one that needs a specialist on call.]]></content:encoded>
  </item>
  <item>
    <title>8 practical changes in Hermes Agent v0.21.4 (and what they're for) &amp;middot; Expedient AI</title>
    <link>https://expedientai.co.uk/feed/hermes-0-21-4-practical-changes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/hermes-0-21-4-practical-changes/</guid>
    <pubDate>Tue, 22 Sep 2026 00:00:00 +0000</pubDate>
    <description>Hermes Agent v0.21.4 is a tagged rollup of the PRs since v0.21.3, dated 2026-09-21. The headline change is a host-wide gateway singleton lock with Desktop attaching to the running host backend instead of spawning a second one. Eight practical wins worth flagging for a one-person UK studio: skills.auto_load pinning, --format stream-json structured JSONL, session_search after/before bounds plus OR-relaxed recall, decline unauthorized_dm for the gateway, MCP discovery_concurrency cap, hermes sessions set-journal-mode, Desktop chat/UI font picker, and LTX 2.5 plus Kling O3 in the video catalogs.</description>
    <content:encoded><![CDATA[Tools 8 practical changes in Hermes Agent v0.21.4 (and what they're for) Hermes Agent v0.21.4, the Singleton rollup, landed on 2026-09-21. The upstream notes call out a host-wide gateway singleton lock with Desktop attaching to the running host backend instead of spawning a second one as the headline change, then list a dozen friction-reducers without curating them. These are the eight that actually shift how a one-person operator runs Hermes day to day: skills.auto_load pinning, --format stream-json structured JSONL output, session_search after/before bounds plus OR-relaxed recall, decline unauthorized_dm for the gateway, MCP discovery_concurrency cap, hermes sessions set-journal-mode, the Desktop chat/UI font picker plus one-click local engine updates, and LTX 2.5 plus Kling O3 in the video catalogs. 2026-09-22 · 6 min read Context: I run Hermes as my main agent on multiple named profiles (default, pclm_commander, vereby_ops_manager) for separate coding projects. Telegram is the primary channel, with daily cron briefings and delegated multi-step builds. v0.21.4 (the Singleton rollup, dated 2026-09-21) is a tagged patch on the 0.21 line. The window since v0.21.3 contains 5,071 non-merge commits across 5,169 changed files, 1,812 merged PRs and 2,116 closed issues. The upstream maintainers flagged exactly one fix as the reason this tag exists: the host-wide gateway singleton lock with Desktop attaching to the running host backend instead of spawning a second one. Everything else in the window is deferred to v0.22.0, which will roll every curated highlight from v0.21.0 onward into a single changelog. Even so, eight of those items genuinely move how I work. They are the ones below, with the singleton fix listed first because it is the reason this tag exists. 1. Host-wide gateway singleton lock, and Desktop no longer spawns a second backend What it is: The gateway now holds a host-wide singleton lock backed by a rendezvous record, and the Desktop app attaches to the running host backend instead of spawning its own. The two processes (CLI/Daemon gateway and Desktop) are now one process per host, regardless of which surface starts it. The rendezvous record is what makes the second-attached process find the first one cleanly across reboots. Why it matters: Before, a desktop install could spawn a second backend, doubling the port count and the WAL writer load on state.db . The second backend often silently won the race for new sessions, so a desktop user could be talking to a stale copy. That class of bug is gone. Practical example: A daemon-style gateway is already running on the host from a previous session. The Desktop app launches. Instead of forking its own backend, it attaches to the existing gateway. The session list in the Desktop is the same list as the CLI sees, and writing to state.db from one surface shows up in the other without a refresh. 2. skills.auto_load pins skills into every new session What it is: A new skills.auto_load config key accepts a list of skill names. Each new session is opened with those skills already injected into the prompt as if they had been loaded on first request. The skill is not duplicated, it is pinned. Why it matters: Operators who run a fixed set of house-style skills (brand voice, imagegen recipe, voice-and-tone) used to add them in the first turn of every session, every time. For a cron job that fires dozens of times a day, that turned into real prompt-token spend. Pinning means the skill is in scope without anyone asking. Practical example: A cron fires every hour to summarise yesterday's briefs. The skill hermes-release-hero is pinned via skills.auto_load . The first turn does not have to spend 200 tokens loading it, because the agent already has it in scope. The cron job runs and writes its output without the loading tax. 3. --format stream-json structured JSONL on the CLI What it is: The CLI now emits structured JSONL output for every turn when called with --format stream-json . One JSON object per line, line-prefixed by event kind ( turn_start , tool_call , tool_result , turn_end ), with the same fields as the in-process Python events. Why it matters: For any operator who pipes Hermes output into another tool (a Notion logger, a Telegram digest, a custom dashboard), the previous output was markdown or plain text. The only stable parser was grep . JSONL means every downstream tool can ingest the stream and re-emit it without losing structure. Practical example: A cron job runs hermes --format stream-json --query "summarise the last 24 hours" and pipes the output to a custom log shipper. The log shipper reads tool_call objects, builds a per-tool latency table, and writes a row to a Supabase table. A week later, the operator queries the table to find which tool is slowest on the morning brief. 4. session_search gets after/before bounds and OR-relaxed recall What it is: session_search accepts --after and --before date bounds (ISO YYYY-MM-DD ) and runs an OR-relaxed recall retry when the first pass returns nothing. The OR-relaxed retry drops one constraint at a time until at least one hit surfaces, so a query that was just outside the strict window still finds a session. Why it matters: Until now, session_search was a single-pass recall. A query that almost matched returned nothing. The new bounds plus the retry are the difference between “I remember doing this last Tuesday” and “I can't find it.” Practical example: An operator searches for a session about a Supabase ticket from last week. The first pass misses because the date is one day off. The OR-relaxed retry drops the date constraint, finds the session, and the operator has the context they need. 5. decline unauthorized DM behavior for the gateway What it is: The gateway can now refuse direct messages from senders who are not on the allowlist, instead of forwarding them into the chat session. decline is the default behaviour for any sender whose user_id is not explicitly allowed; the gateway returns a one-line refusal to the platform and the message is dropped. Why it matters: For a Telegram-bound gateway that handles production work, the surface area for unsolicited inbound DMs used to be wide-open. A typo in a command (or a misrouted group reply) could land a non-owner message in the active session. The decline behaviour is the default-deny posture. Practical example: A Telegram user not in the allowlist DMs the gateway. Before, the message was processed as if it were from the operator. Now the gateway declines the message, logs the attempt with the sender's user_id, and the active session is untouched. 6. Configurable MCP discovery connect cap ( mcp.discovery_concurrency ) What it is: The new mcp.discovery_concurrency setting caps the number of concurrent MCP discovery connections the gateway will open during boot. The default is small (a handful); operators who run many MCP servers can raise it. Why it matters: Discovery is what builds the tool catalogue the gateway advertises to a session. With many MCP servers, an unbounded concurrent connect turns into a thundering-herd at boot and a slow start. The cap is a knob for the operator, not a magic number. Practical example: A host runs 14 MCP servers. Without the cap, boot spends 10 seconds in discovery. Setting mcp.discovery_concurrency = 4 brings boot discovery to under 3 seconds. The tool catalogue is identical at the end. 7. hermes sessions set-journal-mode , Desktop font picker and one-click local engine updates What it is: Three things in one slot because they share a theme: (a) hermes sessions set-journal-mode lets operators flip the SQLite journal mode for state.db between WAL and DELETE without a manual sqlite3 session; (b) the Desktop chat/UI font picker lets the operator pick the chat font and the UI font independently; (c) one-click local engine updates and a Plugins hub plugin-uninstall action round out the Desktop polish for this window. Why it matters: (a) is the answer to “my WAL keeps growing under load” without going to a shell. (b) is the answer to “the chat font is too narrow on this screen” without editing CSS. (c) is the answer to “I want a clean uninstall” without grepping the install directory. Practical example: hermes sessions set-journal-mode WAL after a busy week of cron output. The state.db file size stabilises. The chat font is set to Cabinet Grotesk. The plugins hub shows every installed plugin with a one-click uninstall. None of this needed a code change. 8. LTX 2.5 and Kling O3 join the video catalog What it is: Two new video models are added to the FAL-style video catalog the agent uses for generative video. LTX 2.5 is the new motion-realism tier from the LTX line; Kling O3 is the latest generation of Kling's motion model. Why it matters: Picking a generative video model used to mean reinstalling the integration or editing a config. Now it is a name change in the call. The two models cover the two failure modes the operator usually hits: motion realism (LTX 2.5) and long-form coherence (Kling O3). Practical example: A short promo video needs to feel grounded in motion. The call swaps to LTX 2.5. A 30-second reel needs to hold a single subject without drift. The call swaps to Kling O3. Neither swap is a reinstall or a config edit. What did not make the list (but is in the notes) v0.21.4 is also a stated rollup. The window since v0.21.3 contains the catalog-website work (a page for every catalog plugin and author, with pinned-commit READMEs and added/updated sorting), the dozen new community plugins (tailscale, ssh, shodan, terminal, rss, resetwatch, done-bell, kiwi, cognee, Octen), the backend-owned connector operation with a setup card on Desktop/TUI/CLI, and a large run of profile/multiplex isolation, cron, kanban, Desktop and state.db fixes. Real changes, but either narrower than the eight above or only material for specific workflows. Full curated release notes for the whole 0.21 window ship with v0.22.0, which will roll everything from v0.21.0 onward into a single changelog. Nothing in this window is skipped. The honest version of how to read this release v0.21.4 is a patch tag, not a feature release. The upstream notes are explicit: it exists because the singleton fix needed to reach Cloud agents, which auto-update to the newest release tag, and everything else in the window was clean enough to ride the same train. The headline list runs to one named fix plus about a dozen friction reducers; the daily-impact list runs to eight; the rest are mentioned in the section above. That is the honest shape. Everything in the notes is worth a skim when you hit the specific problem it solves, not before. Frequently asked questions Is Hermes Agent v0.21.4 a big release? No. Despite the version number, 0.21.4 is a stable patch tag for the 0.21 line. The upstream notes say it exists primarily so the host-wide gateway singleton fix reaches Cloud agents, which auto-update to the newest release tag. The window since v0.21.3 contains 5,071 non-merge commits across 5,169 changed files, 1,812 merged PRs and 2,116 closed issues, with curated release notes deferred to v0.22.0 which will roll everything from v0.21.0 onward into a single changelog. What does the host-wide gateway singleton lock actually fix? Before v0.21.4, a Desktop install could spawn a second backend on the same host as the CLI/Daemon gateway. The two processes would each hold their own writer handle on state.db, which is the same bug class that produced the \"N live SessionDB handles\" precursor warning in v0.21.3. The singleton lock, backed by a rendezvous record, means one host backend per host regardless of which surface started it. Desktop now attaches to the running host backend instead of forking its own. Does skills.auto_load mean the skill gets loaded every session? Yes, by design. The skill is pinned into the prompt as if it had been loaded on first request, but it is not reloaded on each turn. For a cron job that fires dozens of times a day, that turns into real prompt-token savings: the skill's body is in scope without anyone asking, and the loading tax is paid once at session start rather than once per turn. When do the full curated notes ship? With v0.22.0. The upstream release notes for v0.21.4 are explicit: full curated release notes for this window ship with v0.22.0, which will roll everything from v0.21.0 onward into a single changelog with highlights, feature areas and complete contributor credits. Nothing in the v0.21.0 to v0.21.4 window is skipped.]]></content:encoded>
  </item>
  <item>
    <title>Top 10 practical changes in Hermes Agent v0.21.3 (and what they're for) &amp;middot; Expedient AI</title>
    <link>https://expedientai.co.uk/feed/hermes-0-21-3-practical-changes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/hermes-0-21-3-practical-changes/</guid>
    <pubDate>Mon, 14 Sep 2026 00:00:00 +0000</pubDate>
    <description>Hermes Agent v0.21.3 is a tagged rollup of the PRs since v0.21.2, dated 2026-09-14. Two named fixes for remote Desktop / Cloud users and ten practical wins worth flagging for the UK studio workflow: refresh-coalesce on the gateway, shared state.db writer handles, JSON-RPC reverse calls, per-model reasoning-effort control, OpenRouter OAuth PKCE, HEIF/HEIC/AVIF decoding, FAL catalog expansion (Wan 3.0, Kling 3.0, MiniMax H3 Max Turbo, Gemini Omni Flash 1.1, Meta Muse), Slack pasted tables, multiplexed-profile isolation, daily MCP re-auth nudge.</description>
    <content:encoded><![CDATA[Tools Top 10 practical changes in Hermes Agent v0.21.3 (and what they're for) Hermes Agent v0.21.3, the Remote Fix rollup, landed on 2026-09-14. The upstream highlight list runs to two named fixes for remote Desktop / Cloud users plus a dozen friction-reducers. These are the ten that actually shift how a one-person operator runs Hermes day to day: refresh-coalesce on the gateway, shared state.db writer handles, JSON-RPC reverse calls, per-model reasoning-effort control, OpenRouter OAuth PKCE, HEIF/HEIC/AVIF decoding, FAL catalog expansion, Slack pasted tables, multiplexed-profile isolation, and a daily MCP re-auth nudge. 2026-09-14 · 7 min read Context: I run Hermes as my main agent on multiple named profiles (default, pclm_commander, vereby_ops_manager) for separate coding projects. Telegram is the primary channel, with daily cron briefings and delegated multi-step builds. v0.21.3 (the Remote Fix rollup, dated 2026-09-14) is a small tagged patch on the 0.21 line. The window since 0.21.2 contains 1,036 non-merge commits across 2,642 changed files and 338 merged PRs, and yet the maintainers flagged exactly two as urgent enough to ship under a new tag: the remote-Desktop refresh burst fix and the state.db writer-handle leak. Everything else is either a small friction reducer or plumbing. The "size" of 0.21.3 is the two named fixes plus everything else that turned out to be ready in the same window, not the commit count for one tag. Even so, ten of those items genuinely move how I work. They are the ones below, with the two named fixes listed first because they are the reason this tag exists. 1. Remote desktop sessions no longer expire on refresh bursts What it is: Both refresh paths on the gateway (the cookie gate and the desktop's native bearer route) now coalesce concurrent requests carrying the same rotating refresh token, so a Desktop wake burst can no longer replay an already-rotated token into the Portal's reuse detection and revoke the whole session. Refresh also runs off the event loop, so a slow identity provider no longer freezes /api/status . Pairs with Portal-side hermes-portal #1209 (sliding 30-day idle horizon, 5-minute rotated-token grace). Why it matters: The bug was loud enough to warrant the tag on its own. A remote operator who woke up to a revoked session every other morning has a working backup pattern now. Practical example: First thing in the morning, three devices with the desktop app wake at the same instant and try to refresh. Before, the rotation window was open and the second rotated token got flagged as replay. Now those three requests coalesce into one, and the session lives until you choose to leave it. 2. Long-lived processes stop leaking duplicate state.db writer handles What it is: Gateway, dashboard/Desktop backend, ACP and CLI readers attach read-only; in-process writers share the registry handle. The " N live SessionDB handles " precursor warning stops firing on a healthy topology. Why it matters: Long-running backends used to log this warning roughly every thirty minutes even when nothing was wrong. The warning now waits to mean something. Practical example: A gateway and an ACP client run side-by-side for a week. Previously, the warning ended up trained to be ignored. Now a warning actually happening means there is a problem, so hermes logs --level WARNING becomes useful again. 3. JSON-RPC reverse calls land in the desktop What it is: The TUI/Desktop now have a generated TS/OpenRPC contract for server-to-client requests, with a Pydantic wire-contract registry on the server. New request types can land in the desktop without a desktop rebuild; the server side is the source of truth. Why it matters: For a desktop that runs alongside a session that's been going for hours, having the protocol declared in a typed registry is the difference between a desktop that upgrades and a desktop that freezes for a tool rename. Once the registry is in, new request types from the server can show up in the desktop without rebuilding it. Backwards-compat is handled on the wire, not in the app bundle. Practical example: The desktop is at version N, the server is at N+2. Three new request types have been added server-side since the desktop shipped. Before, those would not render. Now they do, because the desktop treats them as opaque contracts and the server is the source of truth. 4. Reasoning-effort per model, with a composer pill What it is: Every model picker now has reasoning-effort selection. The Desktop surfaces it both as a composer pill and as a per-auxiliary control. Why it matters: Reasoning-effort used to be a picker-wide setting that you set at session start. Picking a "deep" model for one branch of a worktree and a "shallow" model for another used to require a second agent. Now it is per-picker, mid-session, without dropping model identity. Practical example: The morning routine: draft replies shallow, audit the day's brief deep, switch pickers, run the audit, switch back. Each switch is a single click and the rest of the session state holds. 5. OpenRouter OAuth PKCE login What it is: OpenRouter joins Anthropic and the rest as a first-class PKCE provider, alongside the existing OAuth flows. Why it matters: For an operator who jumps between provider accounts daily, one fewer bespoke env var. Practical example: The OpenRouter auth used to require an API key in an env file. PKCE means the desktop login flow works for OpenRouter the way it works for every other OAuth provider. 6. HEIF, HEIC and AVIF are decoded locally What it is: Image decoding for those formats is in the upstream now, locally, no third-party decoder round-trip. Why it matters: Until now, an iPhone photo pasted into a Telegram chat arrived as a pink square (or went through a third-party decoder). Now it arrives as itself. Practical example: A client sends an iPhone screenshot of their CRM. The agent reads the screenshot rather than guessing from a pink square. 7. FAL catalog expands What it is: Wan 3.0, Kling 3.0 and Kling Image v3, MiniMax H3 Max Turbo, Gemini Omni Flash 1.1 and Meta Muse join the catalog. Why it matters: Picking a generative video model used to mean reinstalling the integration. Picking a model that was a different version of an existing provider used to mean editing config. Now it is a name change in the call. Practical example: Moving a video generation call from Kling 2 to Kling 3 is a string change in the prompt, not a reinstall. 8. Slack pasted tables, and the Agent Sessions API What it is: Slack messages now keep tabular context when they arrive as messages rather than text dumps, and an Agent Sessions API exposes logs and state for a running agent session. Why it matters: Pasting an agent's Markdown table into a spreadsheet is a real copy of workflow, not a metaphor. The Slack change plus the Sessions API make it one paste, not five. Practical example: The morning briefing posts the day's KPI table into Slack. Pasting it into a Numbers tab now preserves cells, headers, and number formatting. The Sessions API lets you point a separate job at a specific agent run from a tab in the desktop. 9. Multiplexed-profile isolation What it is: The gateway now isolates session state cleanly across profiles when you run multiple sessions under one backend, and refuses to write to a state.db hosted on a cross-VM filesystem where the WAL would lose its guarantees. Why it matters: Two profiles on the same host used to share state.db quietly because both home directories resolved to ~/.hermes . They no longer do. The WAL refusal also keeps a flaky filesystem mount from corrupting session history. Practical example: A workspace profile and a client profile both point at the same Hermes backend. Each gets its own state.db. Switching profiles does not accidentally leak a conversation from one into the other. 10. Daily MCP re-auth nudge in Desktop What it is: Refresh tokens from MCP OAuth now bind to their issuer, and the Desktop surfaces a daily nudge for any server whose token is about to expire. Why it matters: Silent MCP token expiry is one of the failure modes that costs a morning. The nudge flips the failure from silent to skippable. Practical example: Every morning at 09:00, the desktop shows a list of MCP servers whose tokens expire in the next 24 hours. You rotate the ones that matter before any tool call actually breaks. What did not make the list (but is in the notes) v0.21.3 is also a stated rollup. The window since 0.21.2 contains the Honcho peer-model rework, Zapier-style tooling internals, TUI splash rewrite, and quiet infrastructure work (Pydantic wire-contract stabilisation, gateway-liveness fixes, MCP refresh-token issuer-binding). Real changes, but either narrower than the ten above or only material for specific workflows. Full curated release notes for the whole 0.21 window ship with v0.22.0, which will roll everything from 0.21.0 onward into a single changelog. Nothing in this window is skipped. The honest version of how to read this release v0.21.3 is a patch tag, not a feature release. The upstream notes are explicit: it exists because the two named fixes needed to reach Cloud agents, which auto-update to the newest release tag, and everything else in the window was clean enough to ride the same train. The headline list runs to two named fixes plus about a dozen friction reducers; the daily-impact list runs to ten; the rest are mentioned in the section above. That is the honest shape. Everything in the notes is worth a skim when you hit the specific problem it solves, not before. Frequently asked questions Is Hermes Agent v0.21.3 a big release? No. Despite the version number, 0.21.3 is a stable patch tag for the 0.21 line. The upstream notes say it exists primarily so the remote-Desktop refresh fixes below reach Cloud agents, which auto-update to the newest release tag. The window since 0.21.2 contains 1,036 non-merge commits across 2,642 changed files and 338 merged PRs, with curated release notes deferred to 0.22.0 which will roll everything from 0.21.0 onward into a single changelog. What was actually broken with remote Desktop sessions before 0.21.3? A Desktop wake burst could replay an already-rotated refresh token into the Portal's reuse detection and revoke the whole session. Both refresh paths on the gateway now coalesce concurrent requests carrying the same rotating refresh token, so the burst is no longer mistaken for replay. Refresh also runs off the event loop, so a slow identity provider no longer freezes /api/status. Pairs with Portal-side hermes-portal #1209 (sliding 30-day idle horizon, 5-minute rotated-token grace). Do long-lived processes still leak state.db writer handles? No, not on a healthy topology. Gateway, dashboard/Desktop backend, ACP and CLI readers attach read-only, and in-process writers share the registry handle, so the 'N live SessionDB handles' precursor warning stops firing. The fix is included in 0.21.3 as part of the same stable rollup. Is OpenRouter OAuth supported now? Yes. PKCE login is in the window and works from the login surface alongside the existing provider list. The desktop now exposes a small composer pill plus a per-auxiliary control for picking reasoning effort per model.]]></content:encoded>
  </item>
  <item>
    <title>6 things that actually change in Hermes Agent v0.21.0 (and what they're for)</title>
    <link>https://expedientai.co.uk/feed/hermes-0-21-practical-changes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/hermes-0-21-practical-changes/</guid>
    <pubDate>Sat, 05 Sep 2026 00:00:00 +0000</pubDate>
    <description>Hermes Agent v0.21.0, the Pantheon release, landed on 2026-08-31. Six changes from the headline list that actually shift how a one-person operator runs Hermes day to day: bot-to-bot DMs, cron memory, live delegation, the MCP command center, the in-app browser, and protected instruction files.</description>
    <content:encoded><![CDATA[Tools 6 things that actually change in Hermes Agent v0.21.0 (and what they're for) Hermes Agent v0.21.0, the Pantheon release, landed on 2026-08-31. The upstream highlight list runs to a dozen items. These are the six that actually shift how a one-person operator runs Hermes day to day: bot-to-bot DMs, cron memory, live delegation, the MCP command center, the in-app browser, and protected instruction files. 2026-09-05 · 9 min read Context: I run Hermes as my main agent on multiple named profiles (default, pclm_commander, vereby_ops_manager) for separate coding projects. Telegram is the primary channel, with daily cron briefings and delegated multi-step builds. v0.21.0 (the Pantheon release, dated 2026-08-31) is the most visible new-shape release since v0.20.0. It also has a problem the headline count hides: between the v0.20.6 and v0.21.0 release tags there are 265 commits and only six carry a feat prefix. The rest are fixes. The "size" of v0.21.0 is the rolling-up of v0.20.1 through v0.20.6 plus a small but sharply different surface change, not the commit count for one tag. Even so, six of those headline items genuinely move how I work. They are the ones below. 1. hermes peer : bot-to-bot DMs between your own agents What it is: Any Hermes agent can now message any other agent by handle, from the CLI or from inside a conversation. Replies land in each agent's canonical Bot Chat, so a conversation between two agents is durable and inspectable rather than fire-and-forget. Why it matters: Until now, multi-agent on Hermes meant plumbing: spin up a subagent, wait for it, hand the result back yourself. With peer, a research agent can hand findings to a coding agent and you can read the exchange after the fact. The interesting part is not the chat UI, it is that two of your agents can now coordinate without you being the message bus. Practical example: A research agent on the default profile finishes a market scan and DMs the result to a coding agent on the vereby_ops_manager profile. The coding agent replies with a draft plan, you wake up to both transcripts in your inbox. You only had to ask once. 2. Cron jobs that remember : the morning briefing job knows what it told you yesterday What it is: Cron agents now load and update persistent memory like every other agent. A continuity=true flag carries each run's output into the next one. Every job gets a durable notepad scratchpad. Monitor-mode jobs skip the LLM entirely when nothing has changed. Why it matters: Scheduled jobs were goldfish. A briefing job at 9am had no idea what the 9am job said yesterday. continuity=true changes that: the briefing can dedupe against its last report, scratchpad gives the job somewhere to write between runs, and monitor mode stops paying for "no change" notifications. For anyone running daily briefings, this is the single biggest behavioural change in the release. Practical example: A daily jobs-and-pipeline briefing at 08:30 reads its own notepad, sees it told you yesterday that Site B's queue was down, and skips that line unless something changed. The output is two lines instead of fifteen, and the two lines are the ones that matter today. 3. Live subagent steering : list, steer, stop the running children What it is: delegate_task grew a real orchestration layer. List the running children, steer one mid-flight with a course correction, stop one early and keep the partial result. Optional JSON-schema validation on child outputs, per-delegation cost surfaced in the result, and the defaults moved to 250 iterations and 10 concurrent children. Why it matters: Until now, a delegated subagent either ran to the end or had to be abandoned. Steering mid-flight is the missing primitive for long research or build jobs that drift off the brief. Combined with JSON-schema validation, the parent's result is a typed object you can rely on, and the cost figure means you can see which branch of a fan-out actually cost what. Practical example: Five parallel subagents are auditing a site. One of them is going deep on a section you do not care about. Steer it back to the part of the brief that matters, it adjusts without losing the work already done. When they all finish, the cost row tells you one branch cost four times the others. Next time you assign that branch to the cheaper model. 4. The MCP command center : paste anything, see cost per server, get re-auth nudges What it is: MCP servers and the catalog merged into one coherent desktop page. Drag-in "paste anything" import. Background health checks that nudge you to re-auth before a tool call fails. A fleet cost and usage overlay showing schema token estimates and 30-day usage per server. hermes:// deep links that install an MCP server with explicit confirmation. Why it matters: For a long time, managing more than a handful of MCP servers meant editing config files and forgetting which token had expired. The command center turns that into a dashboard: paste a server description in, see it land, get a heads-up when one is about to break a tool call. The cost overlay is the second half: it lets you see which servers are carrying the schema-cost budget, which is the line item most easily missed. Practical example: Drag a new MCP server description from a GitHub README into the import box. Background health check runs. Five minutes later, a nudge tells you one existing server's token is about to expire. You rotate it before any tool call actually fails. The 30-day overlay shows one server has 12x the schema cost of the others, you schedule a review. 5. The in-app browser the agent can drive : navigate, click, read What it is: The desktop app's in-app browser stopped being a window the agent could only look at. Hermes now navigates, clicks, and reads it directly. Pages can still be popped out to the system browser, with full link context menus. Why it matters: For desktop automation and web research inside the app, the difference between "I can see the page" and "I can drive the page" is the difference between a viewer and a tool. For a one-person operator, this is most visible on the long, multi-step web tasks where you used to copy-paste between the agent and a browser tab. Now the agent operates the tab and you watch. Practical example: You ask the agent to walk a vendor's docs site, find the rate-limit page, and screenshot the relevant lines. It opens the in-app browser, navigates, clicks through, and reads the table. You get the answer and the screenshot in one turn. 6. Protected instruction files : AGENTS.md, skills and memory always require approval What it is: Writes to protected agent-instruction files (AGENTS.md, skill files, memory stores) always require explicit write approval. A redaction sweep closed secret-leak gaps across terminal output, env-name variants, control-splits, process listings and checkpoints. Why it matters: This is a self-protection fix more than a feature. An agent that can quietly rewrite its own standing orders, or quietly leak a secret into a log line, is an agent that can be quietly steered by a prompt-injected email. The approval gate on protected files is unconditional; the redaction sweep covers the residual gap. Together they make the kind of footgun that bit earlier agent platforms much less likely. Practical example: A session opens an inbound email that turns out to be a prompt-injection attempt asking the agent to "update" its AGENTS.md with new rules. The write is held pending approval, you see the attempted change, you reject it, the session continues normally with the original standing orders intact. What did not make the list (but is in the notes) v0.21.0 also ships a CLI power wave (Ctrl+P command palette, /model picker, /status, status bar with cache-hit and tokens-per-second toggles, plus rotating composer placeholders and terminal pets), the desktop "glass" surface (frost picker, full-range tint and peek), a Slack native live cards path, a Telegram inline-picker that bypasses Telegram's command-menu cap, six new built-in model providers (Meta Model API Muse Spark, CommandCode, Tencent TokenPlan, Nebius Token Factory, Ramp Router, Actual Computer), a stalled-provider recovery that stops burning the retry budget, structured-output translation for Anthropic wires, session pin/unpin from the CLI, and a publish-site command for versioned GitHub/Cloudflare/Netlify Pages deploys. Real changes, but either narrower than the six above or only material for specific workflows. What this means for a UK service business running Mercury OS For the clients we run Hermes-style agent workflows for, the meaningful upgrades are the cron memory and the protected-instruction gate. Together they move "agent platform" from something you supervise constantly to something you can schedule, hand standing orders to, and walk away from. hermes peer is the one to watch next: bot-to-bot DMs between named agents are the prelude to a small business running one research agent, one ops agent and one customer-comms agent that all know about each other. Today it is a CLI. Six months from now, it is the shape of a tiny operator team. The honest version of how to read this release v0.21.0 is labelled a minor but it rolls up the v0.20.1 through v0.20.6 patch tags plus a small amount of new shape. The 265 commits between the v0.20.6 and v0.21.0 tags look like a hotfix because almost all of them are fixes. The headline list runs to twelve items; the daily-impact list runs to six; the rest are mentioned in the section above. That is the honest shape. Everything in the notes is worth a skim when you hit the specific problem it solves, not before. Frequently asked questions Is Hermes Agent v0.21.0 a big release? Yes. It is labelled 0.21.0 rather than a patch tag because the surface changed: Bot Mode is built into the desktop, subagents can be steered live, cron has memory and continuity, and the MCP surface became a dashboard. The underlying commit count between v0.20.6 and v0.21.0 is 265, almost all of it fixes. The size of the release is in what shipped, not in the window between tags. What does Bot Mode in 0.21.0 actually do for one person running Hermes? It makes named agents visible. Every profile gets a deterministic avatar face and a place in a shared roster, and you can put two or more of them in the same group chat. The new piece is hermes peer, which lets any agent DM any other by handle from the CLI or from inside a conversation. The use case is a research bot handing findings to a coding bot and the reply landing back in a chat you can read. Do cron jobs remember anything in v0.21.0? Yes. Cron agents now load and update persistent memory like every other agent, continuity=true carries each run's output into the next, every job gets a durable notepad scratchpad, and monitor-mode jobs skip the LLM entirely when nothing changed. The morning briefing job now knows what it told you yesterday. Why is there no item 7, 8, 9 or 10 in this post? The upstream v0.21.0 highlights list is twelve items. This post deliberately stops at six, the ones a one-person operator running named profiles plus Telegram plus cron plus delegated builds actually feels day to day. The other six are listed honestly in the 'what did not make the list' section so nothing is hidden.]]></content:encoded>
  </item>
  <item>
    <title>Anthropic, Google and Meta all shipped in three days: a working reading of the first week of September 2026</title>
    <link>https://expedientai.co.uk/feed/sept-2026-anthropic-google-meta-week/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/sept-2026-anthropic-google-meta-week/</guid>
    <pubDate>Thu, 03 Sep 2026 00:00:00 +0000</pubDate>
    <description>Claude Fable 5.1 on 1 September, Gemini 3.8 Flash on 2 September, Muse Spark 1.3 on 2 September. What each one actually changed for a UK service business.</description>
    <content:encoded><![CDATA[News Anthropic, Google and Meta all shipped in three days: a working reading of the first week of September 2026 Claude Fable 5.1 on 1 September, Gemini 3.8 Flash on 2 September, Muse Spark 1.3 on 2 September. What each one actually changed for a UK service business. 2026-09-03 · 10 min read Between 1 and 3 September 2026 the three big American frontier labs all shipped a new model within roughly thirty-six hours of each other. Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September. Google released Gemini 3.8 Flash on 2 September, with a gated Gemini 3.8 Flash Cyber variant for vetted defenders. Meta released Muse Spark 1.3 on 2 September, rolling out in Muse Code and the Meta Model API. This is a working reading of what each one actually shipped, and what it means for a UK service business running AI in production. What Anthropic actually shipped Claude Fable 5.1 and Claude Mythos 5.1 are the same model with two different safeguard levels. Fable 5.1 is generally available. Mythos 5.1 is gated through Anthropic's trusted access programmes and is positioned for cybersecurity and life-sciences work where the additional safeguards make sense. If you are a normal UK service business, the one you will touch is Fable 5.1. The headline change for operators is price, not capability. Anthropic cut cache read pricing, which is where the bill lives on any agentic workflow that re-reads the same tool outputs, file contents or conversation history multiple times. Anthropic's own framing is roughly 25% cheaper for typical workloads and up to 45% cheaper for highly agentic workflows. For a small business running long context windows with code editors, document Q&A or multi-step research, that is the line item that matters. Data retention gets a proper enterprise answer. Anthropic's new Enterprise Frontier Safeguards (EFS) keep customer data inside cloud infrastructure the customer controls, while still giving Anthropic the signal it needs for safety monitoring. EFS rolls out in phases from late autumn 2026. Until then, eligible customers can use Fable 5.1 under zero data retention. For any UK business in financial services, legal or healthcare, the data retention story is the thing that was previously the reason not to use Claude in production. That barrier has now been answered on paper. Capability lifts are concentrated on long-horizon work and code. Anthropic's benchmarks for Fable 5.1 emphasise Terminal-Bench, agentic scientific research, multidisciplinary reasoning and agentic coding, where the gap to Fable 5 widens as the task gets longer. The model's default effort level is High inside Claude Code and Medium inside Claude Cowork and Claude.ai, which means you do not have to do anything to get the lift. If you were already paying for Fable 5 in your coding workflow, you should see the difference without changing anything. The Mythos 5.1 angle is mostly for the security crowd. Mythos 5.1 is the gated build with safeguards tuned for cyber and biology work. Anthropic also fixed a specific false-positive problem: cyber safeguards now block 60% fewer benign queries than before, and Fable 5.1 can be used to discover software vulnerabilities (though not to develop exploits for them). If you are running a security consultancy or red team, Mythos is the one to ask for access to. If you are not, ignore it. What Google actually shipped Gemini 3.8 Flash is the third Flash release in six weeks. Google has been running a tight release cadence on the Flash line, and 3.8 Flash is positioned as their best reasoning and coding model yet, at the same introductory price as 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens. That introductory price expires on 31 December 2026, at which point list pricing steps up to $1.50 and $7.50. For a small business planning spend into 2027, that step-up matters and is the most easily missed line in the announcement. It is a long-horizon coding and agent model. Google explicitly says 3.8 Flash "works harder" on complex tasks, executing extra reasoning steps and calling tools iteratively rather than going for raw tokens-per-answer efficiency. On the long-horizon software engineering benchmark DeepSWE v1.1, 3.8 Flash reportedly outperforms larger frontier models end-to-end. On HLE-Verified it hits 54.9%. The honest read is that Flash has moved from "good enough for the cheap layer" to "strong enough for the middle of the stack, including serious agent work." Gemini 3.8 Flash Cyber is the gated security variant. It uses the same underlying intelligence as the regular Flash, but is exposed through Google's new Fairwind programme to a vetted set of defenders. If you are a UK MSSP, blue team or critical infrastructure operator, this is the one to apply for. Same caveats as Mythos 5.1: do not bother applying if you are not in that pool. The practical shift for a UK business already on Google Workspace is that the model switched on you. If you are paying for Workspace and using Gemini for email triage, document summarisation or sheets work, 3.8 Flash is what arrives automatically behind the scenes. You do not need to deploy anything. The lift you will see over the next two to four weeks is in long-document reasoning and tool use, which is exactly where the Workspace integrations spend most of their time. What Meta actually shipped Muse Spark 1.3 is a Meta Superintelligence Labs update focused on agentic workflows and coding. It rolls out on 2 September 2026 in Muse Code and the Meta Model API. Meta positions it as drawing on what they learned from months of broad adoption of Muse Code and the Meta Model API. Previously available reasoning modes are in from launch; the higher reasoning effort setting is "coming shortly after additional safety testing." The agentic framing matters more than the benchmark. Muse Spark 1.3 is tuned for longer-horizon work in a single thread, asking clarifying questions when prompts are ambiguous, and confirming before taking consequential actions. Meta highlights three things: better preservation of detailed requirements in long tasks, more accurate routing of incoming prompts to the right task in messy multi-threaded contexts, and a better sense of the model's own capability boundaries (less hallucination when it hits a wall). If you have used Muse Spark 1.1 in production you will recognise the failure modes those three things are aimed at. Source caveat. Meta's main AI blog at ai.meta.com/blog still lists Muse Spark 1.1 (from July) as the most recent Muse Spark post. The Muse Spark 1.3 announcement is on Meta's research subdomain at research.meta.ai/blog/introducing-muse-spark-1-3 , not on the main blog. This is worth knowing because it is where to look for future Muse Spark drops, and because the absence of a parallel press release on the main blog is, for now, a small but real signal that this is a developer-focused release rather than a headline-grabbing one. For a UK service business the read is straightforward. Muse Spark 1.3 is the new default if you were already on Muse Spark 1.1, and it does not change the broader answer to whether to run Meta models. If you were not on Meta models before, there is nothing in 1.3 that should pull you in. Meta is still going after the developer and personal-agent layer, not the enterprise compliance layer. What actually changed for a UK service business Three shifts. None of them are about a single model. They are about the operating economics of running AI in September 2026. 1. The cheap layer got both cheaper and more capable at the same time. Anthropic cut cache pricing, which is the line that matters most for agentic workflows. Google's 3.8 Flash is positioned as the best Flash yet at the same introductory price. Meta's Muse Spark 1.3 is an iteration on the agentic workflow the previous model already had. Net effect: the tier of work your business does most of (drafting, summarising, classification, triage, long-document reasoning) is now either cheaper, more capable, or both. 2. The frontier layer got two answers to the same question. Anthropic and Meta both shipped a "long-horizon agentic model that knows its limits" in the same week. Google shipped a coding-and-reasoning workhorse that is positioned as good enough for the middle of the stack. The frontier is no longer "one big model that does everything slightly better." It is "two or three specialised models that each do one thing very well, with a cheap layer underneath that can do most of what the frontier used to do." The Mercury OS install shape already routes by task; the news this week is that there are more routing choices and the cheap layer is now genuinely good enough for more of the work. 3. The compliance and data story got a proper answer on Anthropic. Enterprise Frontier Safeguards, paired with zero data retention until EFS rolls out, is the answer to the question UK regulated businesses have been asking for eighteen months. If you are in financial services, legal or healthcare, and you have been waiting for Claude to be deployable under your data-handling rules, this is the week that answer arrived. It is not the same as a UK-specific certification, but it is a serious step. What to do on Monday If you are running AI in your business today, the boring thing is the right thing. Re-price your Anthropic bill. If you are on Fable 5 with significant agentic workflow volume, the 25%-to-45% cache read cut translates directly into a smaller bill. Recalculate against last month's actual usage. If the saving is meaningful, raise the work you route to Claude. Test Gemini 3.8 Flash against your top ten prompts. If you have a Google Workspace spend, you will get 3.8 Flash automatically behind the scenes. The honest test is the same one we have been recommending for months: take your ten most common prompts, run them on the new model against the old, and see whether the lift is real on your work, not on the leaderboard. Wait on Muse Spark 1.3 unless you are already in Meta. There is no UK enterprise compliance story for Meta yet. If you are using Muse Spark 1.1 in production, the 1.3 update is a free lift. If you are not, this is not the week to switch providers. Do not chase the headlines. Three frontier releases in three days is the kind of week that makes operators want to redo their stack. Resist the urge. The wins available this week are in the line items (cache pricing, default Flash model behind Workspace, agentic reliability on Muse Spark), not in reshuffling which provider you pay. The honest version If you are not in the AI business and you are a UK service business owner, the short version is this. The AI you are paying for just got either cheaper or more capable on Tuesday, Wednesday and Thursday this week, depending on which provider sits behind the tool you use. If your vendor is on top of their game you will see the lift in the next two to four weeks. If you do not, that is the signal to ask them. If you are running AI yourself, do the boring thing. Run your top ten prompts against the new models that touch your work. Change one thing. Leave the rest. This was not a breakthrough week. It was a week where the cheap layer got cheaper, the frontier layer got specialised, and the compliance story got a proper answer for the customers who needed one. That is good news for the economics of running AI in a small business, and not much else. Frequently asked questions Should a UK small business actually switch models after a week like this? No. The bottleneck for most UK small businesses is not which model is newest. It is the prompts, the data, the integrations, and the operator habit. Switch one piece per quarter, run the same ten prompts against the new option, and move on. The risk of reshuffling providers every time a launch hits is that you spend the week learning a new tool surface instead of running your business. What does the Anthropic cache pricing cut actually mean in pounds for a small business? Anthropic's own framing is 25% cheaper for typical workloads and up to 45% cheaper for highly agentic work. The 45% number is the one to model on if your Claude usage involves code editors, long document Q&A or any multi-step research with significant context reuse. On a bill that was previously £400 a month, the realistic range is £100 to £180 off, depending on how much of your volume is agentic and how much cache reuse you actually do. Is Gemini 3.8 Flash good enough to replace Claude or Anthropic for our workflow? Probably not wholesale, and you should not try to swap providers wholesale anyway. 3.8 Flash is genuinely strong on long-horizon coding and reasoning and is positioned as the best Flash yet. The honest answer is to test it against your top ten prompts, not to assume a leaderboard number translates to your work. For a small UK business already on Google Workspace, the more relevant question is whether 3.8 Flash is good enough to do more of the work you currently route to a separate provider. Does Enterprise Frontier Safeguards make Claude usable in regulated UK industries? It is a serious step, not the final one. EFS keeps customer data inside customer-controlled infrastructure while still giving Anthropic the safety signal it needs, and zero data retention is available until EFS rolls out. For UK financial services, legal and healthcare, this answers the data-handling question that has been the main blocker to Claude in production for eighteen months. It is not a UK-specific certification, and your compliance team will still want to review it. But it is the first Anthropic announcement on this front that a UK compliance lead can actually work with. What is Meta Muse Spark 1.3 good for, and is it relevant to a UK service business? Muse Spark 1.3 is Meta's update on its Muse Code developer experience and the Meta Model API, focused on agentic workflows, longer-horizon tasks and a better sense of the model's own capability limits. For a UK service business already using Muse Spark 1.1 it is a free lift. For everyone else it does not change the broader case for or against Meta, which is still a developer-and-personal-agent story rather than an enterprise compliance story. The honest read is: upgrade if you are already in, do not switch providers to get this one.]]></content:encoded>
  </item>
  <item>
    <title>Hermes Desktop startup failure: diagnosing and fixing stale model fallbacks</title>
    <link>https://expedientai.co.uk/feed/hermes-desktop-stale-fallback-boot-fix/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/hermes-desktop-stale-fallback-boot-fix/</guid>
    <pubDate>Thu, 27 Aug 2026 00:00:00 +0000</pubDate>
    <description>The backend was healthy. The failure was stale model routing in a scheduled job and an inactive provider in the fallback chain. Here is the safe recovery.</description>
    <content:encoded><![CDATA[News Hermes Desktop startup failure: diagnosing and fixing stale model fallbacks The backend was healthy. The failure was stale model routing in a scheduled job and an inactive provider in the fallback chain. Here is the safe recovery. 2026-08-27 · 4 min read A recent Hermes Agent Desktop update caused an issue where the desktop application appeared unable to start or log in, even though the local Hermes backend and configured AI provider were working normally. The visible error was: Desktop boot failed: Timed out waiting for Hermes backend port announcement (90000ms) At the same time, the logs showed that Hermes had actually started its backend successfully: HERMES_BACKEND_READY port=50465 Hermes backend listening on 127.0.0.1:50465 This created a confusing situation: the Desktop app timed out waiting for its local backend, while the backend was already alive and accessible. What caused it The fault had two overlapping parts: 1. A retired OpenRouter model remained in old task routing. A scheduled task was still attempting to use stealth/ox-alpha , an OpenRouter preview model that had been retired. OpenRouter returned HTTP 404 and directed users to its replacement, z-ai/glm-5.3-flash . 2. Hermes then fell back to an inactive Alibaba Coding Plan model. The fallback chain still contained qwen3.8-max through alibaba-coding-plan . As the relevant Alibaba entitlement was no longer active, the request failed with: HTTP 403: Access to model denied AccessDenied.Unpurchased The updated primary model was already MiniMax M3, but saved scheduled jobs and fallbacks can retain historical model choices. That meant some jobs continued trying obsolete providers even after the global model was changed. The safe recovery process The first priority is to avoid deleting Hermes data. Do not remove or reset: ~/.hermes This folder contains the Hermes configuration, API credentials, sessions, memories, skills, profiles, cron jobs, logs, and local databases. Before changing anything, create a backup: mkdir -p ~/Desktop/hermes-backups tar -czf ~/Desktop/hermes-backups/hermes-before-fix-$(date +%Y%m%d-%H%M%S).tar.gz ~/.hermes A successful archive command may display: tar: Removing leading '/' from member names That is normal and does not indicate a problem. Check model routing Use the Hermes CLI to inspect the active model and fallback chain: ~/.hermes/hermes-agent/venv/bin/hermes fallback list In this case, the output showed: `Primary: MiniMax-M3 (via minimax) Fallback chain: qwen3.8-max (via alibaba-coding-plan)` The invalid Alibaba fallback should be removed through the interactive fallback manager: ~/.hermes/hermes-agent/venv/bin/hermes fallback remove Then check it again: ~/.hermes/hermes-agent/venv/bin/hermes fallback list The goal is for MiniMax to remain the primary model, with no inaccessible provider left in the fallback path. Fix scheduled jobs Scheduled Hermes jobs can retain prior model settings or stop deliberately when Hermes detects configuration drift. List jobs with: ~/.hermes/hermes-agent/venv/bin/hermes cron list Look for failed jobs using old models or errors such as: HTTP 403: Access to model denied or: global inference config drifted Pin affected jobs explicitly to the current working provider and model: ~/.hermes/hermes-agent/venv/bin/hermes cron edit JOB_ID --provider minimax --model MiniMax-M3 For example: ~/.hermes/hermes-agent/venv/bin/hermes cron edit 3c006b59e927 --provider minimax --model MiniMax-M3 This updates only the selected job's model routing. It does not delete the job, its schedule, history, skills, delivery settings, or other Hermes data. Confirm the provider works A simple one-shot request confirms whether Hermes can reach the configured model outside Desktop: ~/.hermes/hermes-agent/venv/bin/hermes -z "Reply with exactly: Hermes MiniMax test passed" A successful result: Hermes MiniMax test passed confirms that the MiniMax API key, provider configuration, local Hermes runtime, and AI request flow are working correctly. When Desktop still fails If the terminal test succeeds but Hermes Desktop still reports a backend port timeout, the problem is likely limited to the Electron Desktop interface rather than the agent, credentials, or stored data. Run a health check: ~/.hermes/hermes-agent/venv/bin/hermes doctor Useful indicators include: Configuration version is current Required Python packages are installed API credentials and config files exist SQLite databases are healthy Sessions, memory, skills, and cron directories are present No deprecated configuration keys are detected If those checks pass, use Hermes through its terminal interface while the desktop issue is investigated: ~/.hermes/hermes-agent/venv/bin/hermes --tui Or start the local browser dashboard: ~/.hermes/hermes-agent/venv/bin/hermes dashboard This retains access to the same sessions, memory, skills, scheduled jobs, and model configuration while bypassing the Desktop application. Key takeaway A Desktop boot timeout does not necessarily mean Hermes itself has failed. In this case, the backend was healthy, MiniMax was working, and all local data remained intact. The effective fix was to remove an inaccessible Alibaba fallback, pin legacy scheduled jobs to the active MiniMax model, verify the provider through the CLI, and use the dashboard or terminal UI until the Desktop connection issue is resolved. Frequently asked questions Frequently asked questions Why does Hermes Desktop show 'Timed out waiting for Hermes backend port announcement'? The visible timeout comes from the Desktop shell waiting for the local Hermes backend to announce a port. In the case we fixed, the backend had started successfully, but a scheduled job kept hitting obsolete providers, so Desktop never finished its startup routine. Is ~/.hermes safe to keep during a Hermes recovery? Yes. ~/.hermes holds the Hermes configuration, API credentials, sessions, memories, skills, profiles, cron jobs, logs and local databases. The recovery steps in this post only edit model routing on individual jobs and remove one entry from the fallback chain. Nothing is wiped. How do I remove an inactive provider from the Hermes fallback chain? Run the interactive fallback manager from the Hermes CLI: ~/.hermes/hermes-agent/venv/bin/hermes fallback remove. Confirm the chain afterwards with hermes fallback list. The goal is the active MiniMax primary with no inaccessible provider left behind. How do I pin a Hermes cron job to the current model? Use the Hermes CLI to edit the affected job directly: ~/.hermes/hermes-agent/venv/bin/hermes cron edit JOB_ID --provider minimax --model MiniMax-M3. This updates only that job's model routing and leaves its schedule, history, skills and delivery settings intact. What should I do if Hermes Desktop still fails after the fallback and cron fixes? Run ~/.hermes/hermes-agent/venv/bin/hermes doctor for a full health check. If the doctor report is clean, use Hermes through the terminal UI with hermes --tui or through the local browser dashboard with hermes dashboard while the Desktop connection issue is investigated separately.]]></content:encoded>
  </item>
  <item>
    <title>The Ox Alpha test: one prompt, a game and a poster</title>
    <link>https://expedientai.co.uk/feed/ox-alpha-one-prompt-test/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/ox-alpha-one-prompt-test/</guid>
    <pubDate>Sun, 23 Aug 2026 00:00:00 +0000</pubDate>
    <description>Two one-shot prompts to an AI agent. A polished Flappy Bird clone and an interactive kinetic poster. Both playable live. The prompts and the results, in full.</description>
    <content:encoded><![CDATA[News The Ox Alpha test: one prompt, a game and a poster Two one-shot prompts to an AI agent. A polished Flappy Bird clone and an interactive kinetic poster. Both playable live. The prompts and the results, in full. 2026-08-23 · 7 min read Play the results first: Flappy Deluxe, the game and Heavy Type, the interactive poster . Both run in your browser. Then read how they got here. Test one: the game The prompt, in full, was one sentence: “Can you make me a flappy birds clone but with nice upgraded graphics and sfx” No spec. No feature list. No design direction. That is the point of the test: how much quality can the agent infer from almost nothing? What came back was Flappy Deluxe , a single self-contained HTML file, about 19KB, no external assets: Five-layer parallax background. Gradient sky on a full day-night cycle, stars that twinkle at night, a sun and moon that cross-fade, drifting clouds, a city silhouette with lit windows, and two depth layers of hills. A hand-drawn animated bird. Gradient body, belly, beak, eye with a highlight, and a wing that flaps on its own phase and snaps harder when you flap. Rotation follows vertical velocity, exactly like the original. Procedural sound effects. Flap, score chime, hit, death and swoosh, all synthesised live with the Web Audio API. No sound files exist; every noise is generated from oscillators and filtered noise at the moment you trigger it. M to mute. Game feel. Flap puffs, a sparkle burst on every point, an explosion on death, screen shake, a difficulty curve that narrows the pipe gap and raises the speed as you score, medals at 10, 20 and 30, and a best score saved locally. Flappy Deluxe, captured mid-flight. Every pixel drawn in code, every sound synthesised live. Total elapsed time from prompt to playable: about two minutes of build. Then the interesting part happened. The bug the agent caught before shipping The first delivery had a flaw the agent only found by testing its own work in a real browser: the game opened on the night phase of its sky cycle, and the night sky colour was nearly identical to the page background. On some first paints the canvas looked completely blank with just a faint box in the middle. The fix took three small patches: start mid-cycle in daylight, wrap the localStorage save in a safety check so the game cannot crash when storage is blocked, and surface any runtime error as visible red text instead of failing silently. That loop, build, self-test in a browser, find the real-world failure, patch, re-verify, is the part that matters. A demo that only works when the person who wrote it is watching is not a product. Test two: the poster The second prompt, again one sentence: “Generate a 1 page website that showcases all your design and animation skills. It can be about anything but the goal is to impress and not look like anything anyone has done before. Make it interactive.” The agent chose the concept itself and built Heavy Type , a living poster built around one idea: typography with physical weight. The hero. Letter weight and width bend toward your cursor in real time. A kinetic hero. The headline is set in a variable font, and every letter's weight and width bends toward your cursor in real time. Move the mouse across the type and the glyphs visibly strain and relax. A gravity pit. A physics sandbox where the alphabet drops into a bordered arena as rigid bodies. You can grab letters, fling them, stack them, and watch them settle with real momentum, rotation and friction. Drop more, or reset. A generative ink field. Hundreds of particles flow along an invisible mathematical field, painting ribbons of ink that never repeat. Your cursor drags and stirs the flow like stirring paint. Editorial art direction. Warm paper, ink black, one signal orange. Film-grain texture over the whole page, a kinetic type marquee, scroll-triggered reveals, and reduced-motion support throughout. The gravity pit mid-session: letters dropped, dragged and piled. Grab one and throw it. Again, the first version had a bug only visible in testing: the physics letters were drawn with a CSS font rule that is invalid inside a canvas, so the bodies fell and bounced invisibly. The agent caught it by probing the canvas pixels, patched the font string, and re-verified that letters now land visibly. Are they really one-shot? Honestly: the first working version of each was one-shot. The shipping version is one prompt plus the agent's own verification loop, which caught two real bugs a human tester would have hit in the first ten seconds. We think that distinction is the honest one to draw, and it is the same distinction we draw with client work: generation is fast, verification is what makes it shippable. What this proves for a small business Nothing about a Flappy clone is directly useful to a Bristol SME. The loop that built it is. The same sequence, prompt, build, self-test in a real browser, fix, deploy, is how we build the things that do matter: enquiry handlers that qualify leads at 2am, quote systems that draft replies in your voice, internal dashboards that pull the numbers off your spreadsheets. The toy demos are the proof that the loop works end to end, from a one-sentence idea to a hosted, tested, public URL, in minutes. The gap between describing software and using it has collapsed. What is left is knowing which software is worth describing. That is the part we do. Try both Play Flappy Deluxe (tap, click or space to flap; medals at 10, 20, 30) and wreck the Heavy Type poster (drag the letters, paint with your cursor). Both are hosted on this site, free, no login. If you want the same loop pointed at a process in your business, book a chat . Bring the bottleneck. Frequently asked questions Frequently asked questions What is the Ox Alpha test? A capability test we ran on our own AI agent: two single-shot prompts, no follow-up corrections allowed in the brief. One asked for a Flappy Bird clone with upgraded graphics and sound effects. The other asked for a one-page interactive site designed to impress. Both results are playable on the public site. Can I play the games the agent built? Yes. Flappy Deluxe is at expedientai.co.uk/play/flappy-deluxe/ and Heavy Type, the interactive poster, is at expedientai.co.uk/play/heavy-type/. Both run in the browser with no install and no login. Were the results really built from one prompt? The first version of each was. The agent then tested its own work in a real browser, found bugs (a blank first paint in the game, physics letters that did not render on the poster) and fixed them before delivery. That verify-and-fix loop is part of how it works, and it is the part most demos skip. What does this prove for a small business? That the gap between an idea and a working, tested, hosted product is now measured in minutes, not months. The same loop that built a game from one sentence builds business tools: quote systems, enquiry handlers, internal dashboards. The toy is the demo; the loop is the product.]]></content:encoded>
  </item>
  <item>
    <title>Top 10 practical changes in Hermes Agent v0.20.2 (and what they're for)</title>
    <link>https://expedientai.co.uk/feed/hermes-0-20-2-practical-changes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/hermes-0-20-2-practical-changes/</guid>
    <pubDate>Mon, 17 Aug 2026 00:00:00 +0000</pubDate>
    <description>Hermes Agent v0.20.2 landed on 2026-08-16. Ten changes that actually shift how a one-person operator runs Hermes day to day.</description>
    <content:encoded><![CDATA[News Top 10 practical changes in Hermes Agent v0.20.2 (and what they're for) Hermes Agent v0.20.2 landed on 2026-08-16. Ten changes that actually shift how a one-person operator runs Hermes day to day. 2026-08-17 · 9 min read Context: I run Hermes as my main agent on multiple named profiles for separate coding projects, with Telegram as the primary channel, daily cron briefings and delegated multi-step builds. These are the ten changes from 0.20.2 that move the needle on that workflow. 1. Import and resume Claude Code or Codex sessions What it is: Hermes can now import a Claude Code or Codex CLI session and resume it. Why it matters: People move between agent tools all the time: a quick fix in Codex, the long build in Hermes. Previously the work stayed stuck in whichever tool started it. Now a session started elsewhere can be lifted into Hermes and continued with its context intact, which is the difference between re-explaining a problem and picking up where you left off. Practical example: Start a spike in Codex, then /import it into Hermes and hand the continuation to a delegated subagent for the full build. 2. MCP cost and usage overlay in the desktop app What it is: The desktop app now shows a cost and usage overlay for your MCP servers, with schema token estimates and 30-day usage. Why it matters: MCP servers quietly spend money. One server with a heavy schema can cost more per call than the model answering you. Seeing 30-day usage per server makes the quiet leakers visible, so you can switch off or replace the expensive ones instead of discovering them on a bill. Practical example: The overlay shows one MCP server has consumed ten times the tokens of the rest in 30 days. You swap it for a leaner server and the overlay confirms the drop next month. 3. Multi-gateway Connections registry What it is: The desktop app now has a Connections registry for multiple gateways, with profile-scoped refreshes. Why it matters: Running more than one gateway, or a gateway per profile, was previously a silent tangle. A registry that lists them, with per-gateway refresh, makes multi-gateway setups inspectable instead of mysterious. Practical example: You run a Telegram gateway for one client and a Discord gateway for another. The Connections registry shows both and refreshes each independently. 4. MCP health checks with re-auth nudges What it is: The desktop app now runs background health checks on MCP servers and nudges you when one needs re-authentication. Why it matters: The classic MCP failure mode is a dead server discovered mid-task: the agent calls a tool, gets nothing, and wastes a turn. Health checks surface the dead server before you need it, and the re-auth nudge means you fix the token before it bites. Practical example: A server's token expires overnight. The health check flags it with a re-auth prompt first thing, so the morning's tasks never hit the failure. 5. hermes:// deep links to install MCP servers What it is: The desktop app can now install MCP servers from hermes:// deep links, with explicit confirmation. Why it matters: Sharing a server setup becomes a click instead of a config exercise. The explicit confirmation step keeps it safe: a link cannot install silently. Practical example: A colleague sends a hermes:// link for a useful server. One click, one confirm, it is installed and listed. 6. Skills tab hub browser and one-click installs What it is: The desktop Skills tab gains a hub browser, a full-skill detail pane, and one-click installs, scoped per profile. Why it matters: Finding and installing skills was a hunt through folders. Now it is a browse, a read, a click. Profile scoping means each project profile only carries the skills it needs, which keeps prompts lean. Practical example: A new project profile gets its three skills from the hub in one click each, instead of manual folder copies. 7. Ctrl+C fixed for the Kitty terminal protocol What it is: The CLI no longer pushes the Kitty keyboard protocol that broke Ctrl+C in some terminals. Why it matters: This is the small fix with the biggest daily impact: an interrupt that does not work means a runaway command you cannot stop. Restoring plain Ctrl+C in affected terminals brings back the one key you rely on most. Practical example: In a terminal that previously ignored Ctrl+C under the Kitty protocol, interrupt now works again immediately after update. 8. Updater self-heals stale git locks and false updates What it is: The updater now clears stale git locks itself and stops reporting "update available" when there is no real update, including on shallow clones. Why it matters: Two classic update pains, both fixed. A stale lock used to hang updates until you found and deleted it by hand. A false "update available" banner used to send you down a pointless update loop. Both are now handled automatically. Practical example: A crashed update leaves a stale lock. The next hermes update clears it and proceeds instead of hanging. 9. Per-terminal --continue with breadcrumbs What it is: The CLI can now continue per terminal, with terminal breadcrumbs tracking where each session left off. Why it matters: With multiple terminals open, "resume" was ambiguous. Breadcrumbs tell Hermes which terminal was doing what, so --continue picks up the right thread instead of guessing. Practical example: Two terminals, two projects. --continue in each resumes the right project's last thread. 10. Cron one-shot runs no longer orphan jobs What it is: A one-shot hermes cron run no longer orphans the job, and stale execution claims are reaped before dispatch. Why it matters: Running a cron job by hand used to leave state behind that could confuse the next scheduled run. The fix cleans claims first, so manual runs and scheduled runs coexist cleanly. Practical example: You trigger a briefing job manually to test it. The next scheduled tick runs normally with no orphaned state. What did not make the list (but is in the notes) v0.20.2 also includes a compression lean-tail mode and recovery-aware evaluation, a per-turn wall-clock duration in the desktop transcript, pre_tool_call content transformation via a modify directive in hooks, delegation subagents skipping the automatic background review, persisted model routes in the gateway, Telegram DM topics, LiteLLM Claude prompt caching on the OpenAI wire, and installer robustness on Linux and Windows. Real changes, but either narrower or less daily-impacting for the one-person operator workflow. What this means for a UK service business running Mercury OS For the clients we run Hermes-style agent workflows for, the meaningful upgrades are the MCP cost overlay and the health checks. Together they turn a growing server fleet from "trust us, it is fine" into something you can see and budget for. The import of Claude Code and Codex sessions also matters for teams that mix agent tools: work no longer dies at a tool boundary. The honest version of how to read a patch release this fast 0.20.2 arrived three days after 0.20.1 with about 397 merged PRs. At this cadence, the patch label is administrative. The ten items above are the ones I actually expect to use. Everything else in the notes is worth a skim when you hit the specific problem it solves, not before. Frequently asked questions Frequently asked questions Is Hermes Agent v0.20.2 a big release? It is a patch release on top of v0.20.1, but a chunky one: about 397 merged PRs in three days. The ten changes below are the ones that most affect how a one-person operator actually runs Hermes day to day. Can Hermes 0.20.2 import Claude Code or Codex sessions? Yes. The new /import command can bring in and resume Claude Code or Codex CLI sessions, so work started in one agent tool can continue in Hermes. What is the MCP cost overlay in 0.20.2? The desktop app now shows a cost and usage overlay for your MCP servers, with schema token estimates and 30-day usage, so you can see which servers cost real money. What is the most important fix in Hermes 0.20.2? The updater self-heals stale git locks and stops reporting false updates on shallow clones, so 'update available' now means an update actually exists. The Ctrl+C fix for the Kitty keyboard protocol on some terminals is a close second.]]></content:encoded>
  </item>
  <item>
    <title>Why AI Council is the most interesting free AI tool for UK SMEs right now</title>
    <link>https://expedientai.co.uk/feed/why-ai-council-is-cutting-edge/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/why-ai-council-is-cutting-edge/</guid>
    <pubDate>Sun, 16 Aug 2026 00:00:00 +0000</pubDate>
    <description>Six AI advisers argue about your decision before you spend on it. Why deliberation beats a single chatbot, and how UK SMEs can use it.</description>
    <content:encoded><![CDATA[News Why AI Council is the most interesting free AI tool for UK SMEs right now Six AI advisers argue about your decision before you spend on it. Why deliberation beats a single chatbot, and how UK SMEs can use it. 2026-08-16 · 9 min read Try it first: the tool is live and free at expedientai.co.uk/tools/ai-council . Three councils per hour, no login, no email. This article explains why we built it the way we did. The problem with a single chatbot Ask one AI model whether you should hire your first employee, and it will usually say yes, then write you a supportive essay. Large language models are agreeable by training. They mirror your framing, adopt your optimism, and rarely tell you the one thing you needed to hear: the objection you had not considered. That is not a flaw you can prompt away. It is a property of asking a single voice. Real decisions in real businesses survive because they get challenged, by a sceptical accountant, an optimistic salesperson, a cautious operations manager, a customer who does not care about your plan. A small business rarely has that room full of people available on demand. AI Council builds the room. What AI Council actually does You describe a decision: hire the first employee, raise prices, open a second location, drop a product line. The tool convenes six advisers, each a distinct archetype with a different priority: growth, money, operations, customers, risk, and the long view. They deliberate in sequence, react to each other, and disagree where they genuinely disagree. Then it hands you a structured verdict, not an essay: Ranked objections. The strongest arguments against your plan, ordered by how much they matter. Risk flags. The specific things that could go wrong, named plainly. A kill line. A concrete threshold that tells you when the decision stops making sense, for example how many new bookings must land to cover the new wage. Cheap tests. Small, low-cost experiments you can run this week to learn before you commit, like a paid trial month with a subcontractor instead of a permanent hire. That structure is the point. An essay is easy to nod along to. A ranked list of objections with a kill line is something you can act on. Why this is cutting edge, honestly Multi-agent deliberation is one of the most active areas of AI research right now, because a single model's blind spots are exactly what a panel of differently-prompted agents can surface. The cutting edge part is not that we use many models. It is that we force divergence, then compress it into a decision-ready output. Three things make it work in practice: Divergent archetypes, not clones. Each adviser is prompted to hold a different objective function. The money adviser cares about cash, the customer adviser cares about churn, the risk adviser cares about downside. They are designed to collide. Adversarial sequence, not parallel summaries. Advisers see each other's points and push back, so the output is a deliberation, not six independent opinions stapled together. A decision-shaped verdict. The deliberation is distilled into ranked objections, risk flags, a kill line and cheap tests. The raw argument is visible in the chamber; the verdict is what you take into Monday's meeting. And because the free tier runs on free open-weight models, each council costs us a fraction of a penny in compute. That is what makes a genuinely free tier possible without a login wall or an email capture. How it helps a UK small business in practice The tool is built around the decisions SMEs actually face. A few examples of what a council will do with them: Hiring your first employee. The council will produce a kill line around how many new bookings must land to cover the wage, and a cheap test such as running a paid trial month with a subcontractor before committing to a salary. Raising your prices. The customer adviser will name the segments most likely to churn, the money adviser will model the margin gain, and the verdict will flag which customers to grandfather and which to move first. Opening a second location or adding a service line. The risk adviser will surface the fixed-cost exposure, the operations adviser will flag the staffing and process load, and the kill line will tell you the utilisation you need to break even. None of this replaces judgement. It does something more useful for a busy owner: it makes the invisible objections visible before the money is spent, in about a minute, for free. The honest bit The free council argues from generic UK SME archetypes, not from your books. It is a pressure test and a devil's advocate, not a forecast. If you want the advisers to argue from your actual numbers, customers and market, that is the paid AI Council, trained on your data. The free tool exists so you can feel the difference a deliberation makes before you ever pay for one. We would also rather you knew the limits up front: a council can miss domain specifics, and its confidence is not a probability. Treat the kill line as a tripwire to monitor, not a prophecy. Try it Put a real decision in front of the council at expedientai.co.uk/tools/ai-council . Three free councils per hour, no login. If it surfaces an objection you had not considered, it has already paid for itself. Frequently asked questions Frequently asked questions Is AI Council free? Yes. The free tier gives three councils per hour with no login and no email capture. It runs on free open-weight models, so each run costs a fraction of a penny in compute. How is AI Council different from asking ChatGPT? A single chatbot gives you one voice that tends to agree with you. AI Council runs six distinct advisers with different priorities who argue with each other, then hands you ranked objections, risk flags and a kill line instead of one essay. Does the council make the decision for me? No. It pressure-tests the decision and surfaces what you have not thought of. The verdict is structured advice with a kill line and cheap tests, and the call stays yours. What is the paid version? The full AI Council is trained on your business data, so the advisers argue from your numbers, your customers and your market rather than from generic archetypes.]]></content:encoded>
  </item>
  <item>
    <title>Top 10 practical changes in Hermes Agent v0.20.1 (and what they're for)</title>
    <link>https://expedientai.co.uk/feed/hermes-0-20-1-practical-changes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/hermes-0-20-1-practical-changes/</guid>
    <pubDate>Fri, 14 Aug 2026 00:00:00 +0000</pubDate>
    <description>Hermes Agent v0.20.1 landed on 2026-08-13. Ten changes that actually shift how a one-person operator runs Hermes day to day: profile export and import, emergency pause, live subagent steering, and more.</description>
    <content:encoded><![CDATA[Tools Top 10 practical changes in Hermes Agent v0.20.1 (and what they're for) Hermes Agent v0.20.1 landed on 2026-08-13. Ten changes that actually shift how a one-person operator runs Hermes day to day: profile export and import, emergency pause, live subagent steering, and more. 2026-08-14 · 10 min read Context: I run Hermes as my main agent on multiple named profiles (default, pclm_commander, vereby_ops_manager) for separate coding projects. Telegram is the primary channel, with daily cron briefings and delegated multi-step builds. These are the ten changes from 0.20.1 that move the needle on that workflow. 1. /export and /import : move a profile between machines What it is: Two slash commands that package a named Hermes profile so it can be moved or shared. Export produces the profile bundle; import restores it somewhere else. Why it matters: Before this, moving a profile meant copying directories by hand and hoping the paths lined up. Now it is one command out, one command in. For anyone who runs project-scoped profiles, this is the difference between a laptop swap taking an afternoon and taking ten minutes. Practical example: /export vereby_ops_manager before a machine change, /import on the new machine, done. The profile keeps its memory, skills and config. 2. hermes pause / hermes resume : a global emergency stop What it is: hermes pause is an emergency stop that pauses cron dispatch, kanban dispatch and new work until you lift it with hermes resume. Why it matters: The failure mode it fixes is familiar: a cron job starts misbehaving, or a wave of background work kicks off at the wrong moment, and there is no single switch to make everything quiet. Now there is. One command stops the noise, you sort out the cause, one command starts it again. Practical example: A briefing job starts firing errors at 11pm. hermes pause, investigate the job, fix it, hermes resume. No fighting individual schedules. 3. Live orchestration of running subagents : steer, don't restart What it is: When you delegate work to subagents, you can now send a running subagent a course correction without stopping it, or stop it early and keep its partial results. Why it matters: Previously a misdirected subagent either ran to the end or had to be abandoned. Now you can steer mid-flight, which saves whole cycles on long research or build tasks that drift off the brief. Practical example: A subagent is auditing a codebase and starts going too deep on one module. Steer it back to the six-criteria checklist and it adjusts without losing the work already done. 4. Structured-output schemas on delegation : results you can trust by shape What it is: delegate_task can now take an optional JSON schema that the subagent's final answer must validate against. Why it matters: The classic problem with delegated work is the freeform report: the answer comes back, but every field is a different shape. A schema means the parent validates the result and can request one bounded correction if it does not fit. Fewer surprises, less parsing. Practical example: Ask three subagents to research three markets, each returning the same schema: name, pricing, gap, verdict. The consolidated result is a table, not a wall of prose. 5. Per-delegation cost in the result : know what the fan-out cost What it is: Each delegation result now carries its own cost figure alongside the summary. Why it matters: If you run parallel subagents regularly, the token bill can quietly stack up. Seeing the cost per delegation makes it visible where it happens, so you can adjust how many children you spawn and on what model. Practical example: After a five-way parallel research run, the results show the cost of each branch. You spot that two branches cost three times the others and switch those tasks to the cheaper model next time. 6. /refine : run the self-improvement review on demand What it is: /refine triggers the memory and skill self-improvement review immediately, instead of waiting for the automatic post-turn pass. Why it matters: The background review usually fires at the end of turns, which is fine, but sometimes you finish a hard session and want the lessons captured right then while the context is fresh. /refine makes the timing deliberate. Practical example: After untangling a tricky deployment, run /refine and the session's lessons become a skill or memory entry immediately, not whenever the next background fork happens to run. 7. /heartbeat : a recurring prompt when you go quiet What it is: /heartbeat sets up a recurring session re-entry prompt that fires when the session is idle. Why it matters: Long-running sessions drift. A heartbeat brings the agent back to you at a cadence you choose, so an idle session becomes a gentle check-in instead of a silent hole. Practical example: A build session with long waits gets a heartbeat every ten minutes, so the agent re-prompts you when it is waiting, rather than sitting silent until you poke it. 8. Read/write window awareness on desktop : the agent can see what's underneath What it is: The desktop app can answer which OS window is underneath it, and can read the window below when asked. Why it matters: For desktop automation and support work, context matters. Knowing what window the user actually has open is the difference between the agent guessing and the agent knowing what you are looking at. Practical example: You are looking at a dashboard and ask the agent about the numbers on screen. It reads the window below and answers about what you are actually seeing. 9. Security: self-repo git mutations are hard-blocked What it is: The terminal tool now blocks commands that mutate the running Hermes source checkout itself, and protected agent-instruction files always require write approval. Why it matters: This is a self-protection fix. An agent that can accidentally commit to or rewrite its own installation is an agent that can break itself. The block is unconditional and cannot be bypassed by force flags, which is exactly the right design for this class of footgun. Practical example: A stray command that would rewrite the agent's own config or source tree is refused with a clear message instead of silently corrupting the install. 10. Wake word for remote desktop : hands-free from the client What it is: The wake word can be captured on the client side and streamed for remote desktop sessions, alongside the existing on-device capture modes. Why it matters: Voice control is only useful if the mic that hears you is the one near you. For remote setups, capturing on the client and streaming means the wake word works even when the agent runs elsewhere. Practical example: Working on a laptop against a remote Hermes desktop, you say the wake word into your laptop's mic and the remote session responds. What did not make the list (but is in the notes) v0.20.1 also includes the new /learn mode for large corpora, Google Gemini 3.7 Flash in the catalogs, Slack native streaming and task cards, HUD mode window handoff on desktop, session naming at start, cron monitor-mode jobs with hash-suppressed change detection, a per-job durable notepad, vision region-zoom cropping, and a Windows update hand-off orchestrator. Real changes, but either narrower or less daily-impacting for the one-person operator workflow. What this means for a UK service business running Mercury OS For the clients we run Hermes-style agent workflows for, the meaningful upgrades are the emergency pause, the profile export and import, and the structured delegation schemas. Together they turn "agent platform" from something you tinker with into something you can safely schedule, move between machines and hold accountable. The security blocks matter more than any feature: an agent that cannot break its own installation is an agent you can leave running. The honest version of how to read a patch release this big 0.20.1 is labelled a patch but carries more than a thousand commits of fixes and features. That is normal for Hermes: releases ship on a schedule, and the patch boundary is administrative, not a measure of impact. The ten items above are the ones I actually expect to use. Everything else in the notes is worth a skim when you hit the specific problem it solves, not before. Frequently asked questions Is Hermes Agent v0.20.1 a big release? It is a patch release on top of v0.20.0, but a chunky one. The ten changes above are the ones that most affect how a one-person operator actually runs Hermes day to day. What is the emergency pause in Hermes 0.20.1? hermes pause is a global emergency stop that pauses cron, kanban dispatch and new work until you lift it with hermes resume. Useful if a job is misbehaving and you want everything quiet while you sort it out. Can I move a Hermes profile to another machine in 0.20.1? Yes. The /export and /import slash commands package a named profile so it can be moved or shared between machines. What does steering a subagent mean? While a delegated subagent is running, you can send it a course correction without stopping it, or stop it early and keep its partial results. That is the live orchestration added in 0.20.1.]]></content:encoded>
  </item>
  <item>
    <title>New free tool: Website AI Visibility Checker for UK SMEs</title>
    <link>https://expedientai.co.uk/feed/website-ai-visibility-checker-tool/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/website-ai-visibility-checker-tool/</guid>
    <pubDate>Wed, 12 Aug 2026 00:00:00 +0000</pubDate>
    <description>A free client-side check of how readable and citable your website is to AI engines. Paste a URL or page HTML for a 0-100 score.</description>
    <content:encoded><![CDATA[News New free tool: Website AI Visibility Checker for UK SMEs A free client-side check of how readable and citable your website is to AI engines. Paste a URL or page HTML for a 0-100 score. 2026-08-12 · 5 min read The short version: check one page of your website . Paste a URL or the page's HTML source, get a 0-100 score across ten readability checks, no login. Why AI visibility matters now When someone asks Google "who are the best electricians in Bristol?" or asks ChatGPT to recommend an accountant, the answer is drawn from pages the AI can read and trust. This is different from a normal search result. The AI reads your page, decides what it means, and decides whether to cite it. If your page has no clear title, no structured data, no FAQ, or thin content, the AI has little to work with. It will cite a competitor who made the basics easy instead. What the checker looks at Ten signals, each scored 0-2, scaled to 0-100: A meaningful page title and meta description Exactly one H1 heading Schema.org structured data, especially Article, FAQPage and LocalBusiness FAQ content, visible or in schema Mobile viewport configuration No noindex directive Content depth Descriptive internal links HTTPS and basic health signals Every check is explained on the tool page, so you know why each one matters and what a fix looks like. How to use it honestly The tool runs entirely in your browser. Nothing is stored. If the free proxy cannot fetch your site, you can paste the page source instead and get the same checks. It is a heuristic, not a full audit. It will not tell you about rankings, backlinks or competitors. But it will tell you, quickly and honestly, whether your page gives AI engines the basic structure they need to cite you. If you want to go deeper, the Local SEO page describes what a fuller engagement covers. And if you are deciding whether automation is worth it at all, the ROI calculator and AI Readiness Scorecard cover that side of things. Check one page of your website and see where the basics stand.]]></content:encoded>
  </item>
  <item>
    <title>New free tool: AI Readiness Scorecard for UK small businesses</title>
    <link>https://expedientai.co.uk/feed/ai-readiness-scorecard-tool/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/ai-readiness-scorecard-tool/</guid>
    <pubDate>Wed, 12 Aug 2026 00:00:00 +0000</pubDate>
    <description>A free 10-question AI readiness assessment for UK small businesses. Score your business 0-100 on data, processes, tools and skills, and see where automation could help. No login required.</description>
    <content:encoded><![CDATA[Tools New free tool: AI Readiness Scorecard for UK small businesses A free 10-question AI readiness assessment for UK small businesses. Score your business 0-100 on data, processes, tools and skills, and see where automation could help. No login required. 2026-08-12 · 5 min read Not every business is ready for AI in the same way. Some need better data. Some need documented processes. Some are genuinely ready to automate. This scorecard tells you which one you are, in about two minutes. The short version: take the AI Readiness Scorecard . Ten questions, a 0-100 score, a plain-English band, no login, no email capture. Why readiness matters more than excitement Most AI projects that fail in small businesses do not fail because the technology is bad. They fail because the foundations were not there. If customer records live in a mix of email and memory, if no one owns the project, or if the team has no appetite for another tool, then even a well-built automation will struggle. The scorecard looks at the four things that actually predict whether automation will take hold: how your data is stored, how your processes are documented, what tools you already run, and how the team responds to change. What the score means 0-33, Early stage. Fix the basics first. Consistent record keeping, documented processes and usable data matter more than buying AI tools. 34-66, Getting there. Some foundations exist. A first pilot on enquiry handling, CRM updates or document chasing is a sensible next step. 67-100, Ready. A structured automation programme can target several processes with measurable results. The score is an opinion based on your answers, not a formal audit. It is a starting point for a conversation, which is exactly what it is designed to be. What it is not It is not a guarantee, a quote, or a full digital audit. It does not replace looking at your actual systems. It is a two-minute, honest first read that tells you where to focus, and it works in your browser with nothing stored. If you have already used the AI Automation ROI Calculator , the two tools fit together: the scorecard tells you whether you are ready to start, and the calculator tells you whether a specific process is worth automating. Take the scorecard , copy the shareable link, and send it to whoever would need to be on board with the next step. Frequently asked questions Why readiness matters more than excitement Most AI projects that fail in small businesses do not fail because the technology is bad. They fail because the foundations were not there. If customer records live in a mix of email and memory, if no one owns the project, or if the team has no appetite for another tool, then even a well-built automation will struggle. What the score means 0-33, Early stage. Fix the basics first. Consistent record keeping, documented processes and usable data matter more than buying AI tools. 34-66, Getting there. Some foundations exist. A first pilot on enquiry handling, CRM updates or document chasing is a sensible next step. 67-100, Ready. A structured automation programme can target several processes with measurable results. What it is not It is not a guarantee, a quote, or a full digital audit. It does not replace looking at your actual systems. It is a two-minute, honest first read that tells you where to focus, and it works in your browser with nothing stored.]]></content:encoded>
  </item>
  <item>
    <title>We built a free AI automation ROI calculator for UK SMEs</title>
    <link>https://expedientai.co.uk/feed/ai-automation-roi-calculator-tool/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/ai-automation-roi-calculator-tool/</guid>
    <pubDate>Wed, 12 Aug 2026 00:00:00 +0000</pubDate>
    <description>A free, no-login ROI calculator for UK small businesses. Estimate the time saved, monthly saving, payback period and first-year ROI from automating one process.</description>
    <content:encoded><![CDATA[News We built a free AI automation ROI calculator for UK SMEs A free, no-login ROI calculator for UK small businesses. Estimate the time saved, monthly saving, payback period and first-year ROI from automating one process. 2026-08-12 · 6 min read The short version: it is free, it takes about a minute, and it works in the browser. Try the AI Automation ROI Calculator . Why we built it Every week we talk to UK small business owners who are trying to justify an automation project. The conversation usually goes the same way. They have a process in mind, they suspect it takes too long, and they have no quick way to turn that feeling into a number they can share with a partner, a director or themselves. Spreadsheet templates exist, but they ask for thirteen inputs and produce a wall of assumptions. Booking a call with a consultant to get a rough estimate feels heavy when you are still at the "is this even worth thinking about" stage. The calculator is the middle ground. Seven inputs, four headline results, one shareable link. What it calculates You enter one process and seven simple figures: How many people do the work Hours per person each week, including checking, copying and rework A fully-loaded hourly cost The percentage of the process you think could be automated One-off implementation cost Ongoing monthly cost It then works out the monthly hours saved, the gross value of that time, the net monthly saving after ongoing costs, and the payback period. It also shows a first-year position, including ROI where a one-off cost has been entered. Everything is transparent. The formulas are printed on the page, and the estimates are deliberately conservative. If you enter a 40% automation share, it uses 40%, not the 90% you will see in vendor demos. What it deliberately does not do It does not count revenue gains from faster replies, quality improvements, or avoided mistakes. Those are real, but they are hard to predict, so they are kept out of the core maths. It does not claim to be financial advice, a quote, or a guarantee. It is a planning estimate, clearly labelled as one. It also does not ask for an email address, a login, or a cookie consent wall. Open the page, move the sliders, get the number. Why we think it matters Automation decisions in small businesses usually stall for one reason: nobody has put a rough number on the problem. The calculator lowers that barrier to zero. It also gives us a better starting point when we do talk. A business that has already estimated one process is a business that has already done the thinking. We can skip the generic pitch and get straight to what should stay human, what can be automated, and what it would actually take. If you are not sure whether a process is a good candidate, the page includes three worked examples: lead follow-up and CRM updates, client onboarding administration, and weekly reporting. There is also a longer guide on how to calculate the ROI of AI automation for a UK small business if you want the reasoning behind the maths. What is next The calculator is the first tool on a small tools page we intend to grow. If there is a calculation you would find useful as a UK small business owner, tell us and we will consider building it. In the meantime: try the calculator , copy the shareable link, and send it to whoever you would normally run the numbers past. Frequently asked questions Is the AI automation ROI calculator free? Yes. You can use it without creating an account or providing contact details. What does the calculator estimate? It estimates monthly hours saved, gross and net monthly saving, payback period and first-year ROI for one repetitive business process, based on the assumptions you enter. Does the calculator guarantee savings? No. It is a planning estimate. Actual results depend on the process, data quality, integration requirements, staff adoption and the amount of human review needed.]]></content:encoded>
  </item>
  <item>
    <title>How to calculate the ROI of AI automation for a UK small business</title>
    <link>https://expedientai.co.uk/feed/how-to-calculate-ai-automation-roi-uk-sme/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/how-to-calculate-ai-automation-roi-uk-sme/</guid>
    <pubDate>Tue, 11 Aug 2026 00:00:00 +0000</pubDate>
    <description>The time-saving maths, the costs to include, and the mistakes to avoid when estimating AI automation ROI. With a free calculator.</description>
    <content:encoded><![CDATA[News How to calculate the ROI of AI automation for a UK small business The time-saving maths, the costs to include, and the mistakes to avoid when estimating AI automation ROI. With a free calculator. 2026-08-11 · 7 min read The short version: Start with one process, not the whole business. Measure the time people actually spend on it, including checking, copying and rework. Apply a conservative automation percentage, usually below half. Subtract real ongoing costs, not just the setup fee. Use a free calculator to check the numbers before you talk to anyone. Why the simple maths works AI automation ROI for a small business usually comes down to one question: does automating this process free up enough staff time to cover the cost? You do not need a finance team to answer that. You need a rough weekly time figure, a sensible hourly cost, and a conservative estimate of what can be automated. The AI Automation ROI Calculator walks through the calculation in under a minute. No login, no email capture, and it gives you a shareable link so you can compare scenarios. The formula, in plain English The maths has three steps. Step one: find the monthly hours spent. Multiply the number of people who do the work by the hours each spends per week, then convert to a monthly figure. Step two: apply the automation share. Multiply the monthly hours by the percentage of the process you believe could be automated. Be conservative here. Human review, judgement and exception handling stay with people. Step three: subtract the costs. Multiply the automated hours by a fully-loaded hourly cost, then subtract the ongoing monthly cost of the software, model usage, hosting and maintenance. That gives you a net monthly saving. Payback is simply the one-off implementation cost divided by the net monthly saving. What to include in a fully-loaded hourly cost Basic salary understates the real cost of staff time. A fully-loaded figure includes employer costs, benefits, management time, equipment and workspace. For a UK employee on £30,000 a year, the fully-loaded cost is often closer to £20 to £25 per hour once you include employer National Insurance, pension and overhead. Use a figure you would use internally for comparing options, not an audited model. Common mistakes when estimating AI automation ROI Overestimating the automation share. If a process needs judgement, approval or relationship management, that part stays human. A 40% automation share is a reasonable starting point for many admin-heavy processes. Double-check before going above 60%. Forgetting the ongoing costs. The setup fee is only half the picture. Model usage, subscriptions, hosting and monitoring repeat every month. If a system costs £150 a month but only saves £120 of time, it does not pay for itself. Measuring only the happy path. Real processes include checking, copying, chasing and rework. If you measure only the ideal flow, you will overestimate the saving. Measure the average week. Ignoring the intangibles. Faster replies, fewer mistakes and better compliance are real benefits but harder to price. Keep them out of the core calculation and mention them separately in the business case. A worked example A three-person sales team spends three hours each per week qualifying enquiries, copying data into a CRM and sending routine follow-ups. That is nine hours a week, roughly 39 hours a month. At a fully-loaded cost of £25 an hour, the time is worth about £975 a month. If 40% of that can be automated, the gross saving is about £390 a month. After £150 a month in ongoing costs, the net saving is £240 a month. With a one-off implementation cost of £2,500, payback is roughly 10.4 months, and the first-year net position is about £380. That is a reasonable project. Not a transformation, but a genuine, measurable saving on one process. What the estimate does not cover Revenue gains from faster responses, quality improvements, avoided mistakes and compliance risk reduction are all real but harder to predict. Keep them out of the core calculation. They are upside, not the basis of the decision. And no estimate guarantees a result. Actual savings depend on the process, data quality, integration requirements, staff adoption and how much human review remains. Treat the calculation as a planning estimate, then test in production. Try the calculator Use the free AI Automation ROI Calculator for UK SMEs to run your own numbers. It covers time saved, gross and net monthly saving, payback period and first-year ROI, and you can share the results with anyone you are discussing the project with. If you want a second opinion on a specific process, book a chat . Twenty minutes, no obligation. Frequently asked questions Why the simple maths works AI automation ROI for a small business usually comes down to one question: does automating this process free up enough staff time to cover the cost? You do not need a finance team to answer that. You need a rough weekly time figure, a sensible hourly cost, and a conservative estimate of what can be automated. The AI Automation ROI Calculator walks through the calculation in under a minute. The formula, in plain English The maths has three steps. Step one: find the monthly hours spent. Multiply the number of people who do the work by the hours each spends per week, then convert to a monthly figure. Step two: apply the automation share. Multiply the monthly hours by the percentage of the process you believe could be automated. Be conservative here. Human review, judgement and exception handling stay with people. What to include in a fully-loaded hourly cost Basic salary understates the real cost of staff time. A fully-loaded figure includes employer costs, benefits, management time, equipment and workspace. For a UK employee on £30,000 a year, the fully-loaded cost is often closer to £20 to £25 per hour once you include employer National Insurance, pension and overhead. Common mistakes when estimating AI automation ROI Overestimating the automation share. If a process needs judgement, approval or relationship management, that part stays human. A 40% automation share is a reasonable starting point for many admin-heavy processes. Double-check before going above 60%. Forgetting the ongoing costs. The setup fee is only half the picture. Model usage, subscriptions, hosting and monitoring repeat every month. A worked example A three-person sales team spends three hours each per week qualifying enquiries, copying data into a CRM and sending routine follow-ups. That is nine hours a week, roughly 39 hours a month. At a fully-loaded cost of £25 an hour, the time is worth about £975 a month. If 40% of that can be automated, the gross saving is about £390 a month. What the estimate does not cover Revenue gains from faster responses, quality improvements, avoided mistakes and compliance risk reduction are all real but harder to predict. Keep them out of the core calculation. They are upside, not the basis of the decision. And no estimate guarantees a result. Actual savings depend on the process, data quality, integration requirements, staff adoption and how much human review remains. Try the calculator Use the free AI Automation ROI Calculator for UK SMEs to run your own numbers. It covers time saved, gross and net monthly saving, payback period and first-year ROI, and you can share the results with anyone you are discussing the project with. If you want a second opinion on a specific process, book a chat . Twenty minutes, no obligation.]]></content:encoded>
  </item>
  <item>
    <title>Meta is back in open weights: Muse Glimmer runs on a laptop, and the manifesto says why that matters</title>
    <link>https://expedientai.co.uk/feed/meta-muse-glimmer-open-weights/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/meta-muse-glimmer-open-weights/</guid>
    <pubDate>Mon, 10 Aug 2026 00:00:00 +0000</pubDate>
    <description>Meta's first open-weights release since Llama 4: a 30B agentic model that runs on one consumer GPU, plus a Zuckerberg open letter on why open weights are an American AI leadership play. What actually changed for a UK service business.</description>
    <content:encoded><![CDATA[News Meta is back in open weights: Muse Glimmer runs on a laptop, and the manifesto says why that matters Meta's first open-weights release since Llama 4: a 30B agentic model that runs on one consumer GPU, plus a Zuckerberg open letter on why open weights are an American AI leadership play. What actually changed for a UK service business. 2026-08-10 · 7 min read Meta released its first open-weights model since Llama 4 yesterday, and it is the most interesting open release of the year for a specific reason: it is an agentic model designed to run on a personal computer. It is called Muse Glimmer, it is 30 billion parameters, and at 4-bit precision it fits in under 20 GB, which means one consumer GPU runs it. The release came with a manifesto from Mark Zuckerberg arguing that open weights are the American AI leadership strategy. Both parts matter, but not the way the headlines frame them. What actually shipped Muse Glimmer is a distilled, open, agentic model. It is distilled from Muse Spark, Meta's frontier line, and released under Apache 2.0, which is the permissive licence that lets businesses modify and self-host it without paying Meta anything. It is an agentic model: it is built to use tools, browse, run code, and complete multi-step tasks, not just answer questions. Meta's own benchmarks put it ahead of the other open models in its size class, Gemma 4 31B and Qwen3.6 27B, on tool use, with 75.5% on the MCP Atlas agentic benchmark against 54.2% and 62.5% respectively. It also scored 94.7 on AIME 2026 and 51.2 on SWE-Bench Pro. The headline number is not the benchmark. It is 20 GB. A frontier-adjacent agentic model that fits on one consumer GPU changes what self-hosting means. Until now, genuinely capable agentic models have mostly lived behind paid APIs, or on hardware that costs more than most small businesses spend on software in a year. Glimmer is the first serious open-weights model that makes the "run it yourself" argument for agentic work, not just drafting and classification. The manifesto is a strategy statement, not a press release. Zuckerberg published "Open Weights and American AI Leadership" alongside the release, arguing that concentrating advanced AI under a small number of closed companies is both a security risk and an economic one, and that the US should compete by spreading capable open models, not by restricting them. Whether you agree with the politics, the commercial signal is clear: Meta is all-in on open weights as a distribution strategy, and it is pricing and licensing to win the developer and small-business layer. What actually changed for a UK service business Three things, in order of practical importance. 1. The self-host bar moved up. The cheap layer of any AI stack is now capable of agentic work, not just text generation. The practical consequence for a service business is that the "local model" conversation changes from "what can we run cheaply for drafting?" to "what can we run ourselves for enquiry handling, data extraction, and tool use?" That is the same capability tier that was API-only a year ago. 2. Open weights are now a policy position, not a niche. When Meta competes on open weights, the whole market follows: pricing pressure on API tiers, more capable small models, more licensing freedom. For a business that wants to own its stack, that trend is the one that matters, because every step of it lowers the cost of running AI without locking you into a provider. 3. The honest caveats still apply. Meta benchmarked Glimmer against two other open models and published the table, and it is a good result. But independent, like-for-like reproduction of those numbers has not happened yet, and the first independent read of the model already notes that Qwen3.6-27B wins several practical agent and multimodal tests in Meta's own table. Also worth knowing: Muse Spark 1.2 and a terminal coding agent called Muse Code shipped earlier this month, and Spark 1.1 introduced Meta's first paid model service in July. Meta is open-weight, but it is also commercialising the frontier line. Those two facts sit together, not against each other. What we are changing in the Mercury OS install shape Nothing this week. Glimmer is brand new and the independent testing is not done. The Mercury OS shape already treats the cheap-model layer as swappable via config, so when a model genuinely earns its place in that layer, the change is small and contained. We would rather be three months late picking up an open model after independent testing than three days early on a benchmark sheet. For UK businesses running AI day-to-day, the short version: Meta is back in open weights, the model is genuinely impressive on paper, and the direction of travel is the same one that has been paying off for a year, capable open models getting more capable and cheaper to run. The boring advice still wins. Wait for the independent tests, run your own ten prompts against it, and switch only if it does your job better. The model will still be open in six weeks. Frequently asked questions Can Muse Glimmer actually run on a normal laptop? Meta says yes at reduced precision. The model is 30 billion parameters, and quantised to about 4-bit it fits under 20 GB, which is within reach of a single consumer GPU. That is the headline: a capable agentic model that runs on local hardware rather than only through a paid API. Is Muse Glimmer free to use for a business? The weights are released under Apache 2.0, which permits commercial use, modification, and self-hosting without paying Meta. What you pay is the hardware to run it and the engineering time to integrate it. That is the trade-off of self-hosting: lower per-use cost, higher setup cost. Should a UK small business switch to Muse Glimmer? Not yet. The benchmark results are strong but they are Meta's own numbers, and independent testing has not caught up. The sensible move is to wait for independent reproduction, then run your own most common prompts against it. If it beats what you use today on your actual work, the switch is a config change in most modern stacks. Why does Meta releasing open weights matter to a small business? Because open-weight releases put pricing pressure on every closed API and give you the option of owning the model yourself. When a company of Meta's scale competes on open weights, the whole market moves that way, which is what has been driving down the cost of the cheap layer of AI for the last year.]]></content:encoded>
  </item>
  <item>
    <title>Top 10 practical changes in Hermes Agent v0.20.0 (and what they're for)</title>
    <link>https://expedientai.co.uk/feed/hermes-0-20-practical-changes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/hermes-0-20-practical-changes/</guid>
    <pubDate>Sat, 01 Aug 2026 00:00:00 +0000</pubDate>
    <description>The Herald Release dropped on 2026-08-03. Ten changes from v0.20.0 that actually shift how a one-person operator runs Hermes day-to-day, anchored in the release notes, with the headline grabbers deliberately passed over.</description>
    <content:encoded><![CDATA[Tools Top 10 practical changes in Hermes Agent v0.20.0 (and what they're for) The Herald Release dropped on 2026-08-03. Ten changes from v0.20.0 that actually shift how a one-person operator runs Hermes day-to-day, anchored in the release notes, with the headline grabbers deliberately passed over. 2026-08-01 · "11 min read" We use Hermes Agent as our day-to-day terminal assistant at the studio. Multiple named profiles, Telegram as the primary comms channel, daily cron briefings, and a skills library that's grown fat over the last year. So when the Herald Release (v2026.8.3) dropped on 3 August, we did what we do with any release: pulled the notes, pulled the PRs, and asked which changes actually move the needle on a one-operator workflow. These are the ten. They are not the headline grabbers. They are the ones that change how a single person uses Hermes across a working day. 1. Voice mode is finally a conversation, not a voicemail What it is: Hermes used to wait until it had generated the full reply, then speak it as one long audio file. Now it streams clause-by-clause as the model produces text. You can interrupt it mid-sentence by just talking, it stops, listens, and the model is told you cut in. Busy-aware silence detection means it doesn't talk over you. Works in CLI voice mode, on the desktop, and through the messaging gateways. Why it matters: Today's pain: you ask a question, the agent thinks for ten seconds, then monologues a one-minute answer you can't stop. With this, voice feels like a real conversation, you steer, you interrupt, you say "wait, back up", and it does. Practical example: "What's on the kanban today?" Hermes starts listing tasks. You say "stop, just the blocked ones", it cuts off mid-word, the model gets the interruption in context, and continues with just the blocked list. Same way you'd talk to a colleague. Under the hood (#69511, #73862, #74223, #74000, #69602): streaming TTS with clause-level chunking, barge-in handler that injects the user's cut-in into the model turn, and a busy-aware silence detector that knows when you're still talking. 2. "Hey Hermes", wake words on-device What it is: Pick your own wake phrase, "hey Hermes", "computer", anything you like, and Hermes starts listening. Detection runs on-device, so no audio leaves your machine while it waits. Different wake phrases can route to different profiles, and saying "stop" ends the voice chat on every surface without touching the keyboard. Why it matters: Until now, voice mode meant sitting down and pressing a button. With wake words, you can talk to Hermes from across the room. Your terminal is now something you can have a hands-free conversation with, the same way you'd ask Siri on a HomePod. Practical example: You're at the kitchen counter, laptop in the other room. "Hey Hermes, what's the Vereby deploy status?" It answers out loud via the desktop. You say "stop" when you're done. No keyboard, no phone, no friction. Under the hood (#70509, #73106, #73933): open-vocabulary wake-word detection running locally, profile-routed wake triggers, and a global voice-stop phrase that fires across every surface. 3. /redirect, correct the agent mid-turn without /stop What it is: If Hermes is heading the wrong way, you no longer have to stop it and re-explain. Type a correction while it's working and the active turn is redirected: work in flight is preserved, the original prompt is kept, the agent course-corrects with your new guidance. Pairs with double-ESC draft discard and a composer undo stack. Why it matters: The classic LLM failure: "I asked it to audit the auth module and it's rewriting the database layer. By the time I see what's happening, three minutes of work is wasted." With redirects, you correct mid-flight the same way you'd correct a junior dev, "no, stop, do X instead", without losing the work already done. Practical example: "Audit the Vereby webhook handler." Hermes starts reading the wrong file. You type "no, the handler is in backend/app/webhooks/, not backend/app/api/". It course-corrects, keeps the audit framing, and the wrong-file read is logged but not punished. Under the hood (#63104, #72339, #74736): live turn-redirect with in-flight preservation + composer-level undo stack. 4. Tools that fix themselves, no more "I tried again" What it is: A sweep of self-recovery upgrades across the core toolset. Truncated terminal output now spills to a file the agent can read back. Patch detects already-applied edits and diagnoses whitespace mismatches instead of failing. write_file verifies content on disk before declaring success. Searches that match nothing probe for near-misses and try multi-line variants. The default tool-calling iteration limit jumped 90 → 500. Why it matters: The failure mode this fixes is invisible until you hit it. Hermes runs a long task, hits a tool error, doesn't know how to recover, and burns 5–10 turns going in circles. With self-recovery, the tools return actionable hints and the agent gets back to work instead of getting stuck. Practical example: You're auditing a 4,000-line FastAPI service. Hermes's search returns no matches for "rate_limit", wrong whitespace? Auto-multiline probe runs, finds "rate-limit" (hyphenated), audit continues. Three months ago this would have stalled the agent for two turns. Under the hood (#77041, #76998, #77024, #77055, #77011, #76992, #72176): terminal truncation-to-file + cwd echo, patch idempotence + whitespace diagnosis, search zero-match recovery, and the iteration cap bump. 5. Compression that respects your conversation What it is: Context compression got a deep overhaul. Proactive tool-result pruning for large-window models. Per-turn micro-compaction that amortizes the cost instead of one giant pause. A guaranteed N-user-message tail so recent conversation always survives. Progress-aware timeouts that stop punishing slow summary models. Ghost-skill defense so a pruned skill can never silently haunt a session. Thresholds configurable per-model and in absolute tokens. Why it matters: Long sessions used to stall: the model would suddenly pause for 30 seconds while compression ran, lose the last few exchanges, and behave erratically. With this, compression is invisible, it amortizes across turns, keeps your recent context safe, and stops surprising you with mid-session amnesia. Practical example: A morning cron briefing generates 40 turns of follow-up questions. Halfway through, compression runs in the background per-turn instead of as one big stall. The last 10 messages are guaranteed to survive. You never see the work pause, and the briefing's later answers still have the earlier context they need. Under the hood (#70254, #75345, #70250, #71508, #70275): proactive tool-result pruning + per-turn micro-compaction + guaranteed tail window + ghost-skill defense. 6. CLI power-user wave, !command , /init, /diff, /context, /focus What it is: Five new CLI commands that compress a week's worth of friction into one keystroke each. !command runs a shell command instantly without spending a model turn. /init scans your project and generates (or updates) an AGENTS.md. /diff shows staged/all/session changes from any surface. /context breaks down exactly what's filling your context window. /focus gives you a reduced-output view with hidden-line recovery for long outputs. Plus hermes import-agent migrates your Claude Code or Codex CLI setup into Hermes in one command. Why it matters: Five commands, five different daily frictions gone. The single biggest one is !command : today, if you want to run ls -la mid-task, the model spends a turn on it. Now you type !ls -la and the shell output drops straight into the conversation, no turn spent. Practical example: You're mid-task and want to check the current git status. !git status → instant output, model keeps its turn budget intact. Later, /diff session shows you what files Hermes touched in this conversation. /context lets you see why a long task is slowing down, "ah, it's the read_file results filling 60% of the window, time for /compress ." Under the hood (#72257, #72178, #72240, #72242, #72302, #72262, #72190): shell-bypass mode for !command , project scanner for /init, session-scoped diff state, context budget breakdown view, and the importer. 7. Smart approvals grow up, hermes approvals suggest What it is: The approval layer used to be either too eager (auto-approve safe stuff) or too strict (block the same rm five times a day). Now hermes approvals suggest mines your approval history into allowlist proposals. Operators can customize the smart-approval policy. A consecutive-denial circuit breaker stops a misbehaving loop cold. Desktop pairing approvals are profile-correct with a proper surface to answer them from. Plus a new approval gate for docker/podman daemon-redirect commands. Why it matters: Approvals are the second-most-clicked button in Hermes after Enter. With this, the agent proposes new allowlist rules based on what you've already approved, you accept the ones that look right, and the click count drops by half over a week without losing any safety. Practical example: You approve docker compose logs -f five times in three days. On day four, hermes approvals suggest proposes adding it to your allowlist. You accept. The agent never asks again for that command. A week later, you have a 30-command allowlist that's actually yours, not a default. Under the hood (#72259, #72186, #72203, #74446, #71092): approval-history mining, customizable policy, circuit-breaker for runaway denies, profile-correct desktop approvals. 8. Voice on every platform, WhatsApp, Feishu, DingTalk, LINE, QQ, Photon, Weixin What it is: Send a voice note to Hermes on any of the seven previously-text-only messaging platforms and it's transcribed, answered, and replied to with auto-TTS, platform-aware (opus where platforms want opus, captions attached correctly). STT is now a fully-configurable hermes tools category with GUI toggles, dashboard dropdowns, unified language resolution, and OpenAI's gpt-transcribe support. One unified spoken-text preprocessor cleans markdown, code, and URLs out of speech across every TTS provider. Why it matters: If your customers or team chat on WeChat or LINE or WhatsApp, you can now send Hermes a voice note and get a voice reply back, in their platform's preferred format. No more "type it out, the bot can't hear you." This is the difference between "AI assistant that happens to live in chat" and "AI assistant you can actually talk to on the platform your customer prefers." Practical example: A Bristol client sends a WhatsApp voice note asking about their Vereby campaign status. The gateway transcribes, Hermes checks the campaign data, the reply is sent back as opus audio with captions. The client hears "yes, the campaign launched yesterday and has 47 sign-ups" while reading the same text in the captions. Under the hood (#73515, #73508, #73910, #73513, #73067): platform-aware voice adapters + unified language resolution + spoken-text preprocessor. 9. Outbound webhooks, Hermes pushes, you don't poll What it is: Until now, integrating with Hermes meant polling or listening on a platform. Now Hermes pushes signed lifecycle events (session activity, turn completions, tool events) to any HTTP endpoint you register, with HMAC signatures so your receiver can verify authenticity. Wire Hermes into your CI, your home automation, your dashboards, or any service that speaks HTTP, with no polling loop. Why it matters: Today, if you want to trigger a Vereby-side action when Hermes finishes a turn, you poll. With webhooks, Hermes tells you. That turns "near-real-time assistant" into "actually-real-time assistant" for any external system you can write a POST endpoint for. Practical example: You want a Slack notification every time a Hermes turn completes in your Vereby profile. Register a webhook on the Vereby side pointing at Slack's incoming-webhook URL, signed with a shared secret. Every Hermes turn → one Slack message with the summary. Zero polling, zero latency. Under the hood (#69406): signed outbound webhook delivery with HMAC verification on the receiver side. 10. Faster everywhere, again, and a desktop that feels native What it is: The headline numbers: hermes -w cold start dropped ~14s → ~1.8s. hermes update no-ops got 2–6s faster. Heavy SDKs lazy-load off the import path. Config reads stopped deep-copying (54× faster on the telemetry gate). Prompt caching now covers tool schemas on native Anthropic without history loss. The desktop shipped a second 60fps wave, streaming cost is independent of transcript length, drag stays at 60fps with five streaming tabs, idle CPU is near zero in the background. Why it matters: You live in the desktop app for hours and you launch Hermes thirty times a day. If it stutters or takes 14 seconds to start, you feel it. This wave is not glamorous, it is the difference between "I'm using an AI tool" and "I'm using my tool, which happens to be AI." Practical example: You hermes -w to start a worktree session. Three weeks ago, you waited 14 seconds. Today, ~1.8 seconds. By the time your hand moves from keyboard to mouse, the chat is ready. Same on the desktop: you're reviewing a long code-review transcript, the model finishes a new reply, the diff lands, scroll position preserved, sidebar updates without re-flow. Five years ago that was a dream. Now it's the desktop app under default load. Under the hood (#76032, #71637, #74218, #74204, #71835, #72346, #75218): cold-start path reduction, SDK lazy-loading, no-deep-copy config reads, tool-schema prompt caching on Anthropic, and the desktop's second 60fps wave. What didn't make the list (but is in the notes) The release also ships grounded-citations research skill with fact-checking (the difference between "sounds right" and "provably sourced"), A2A v1.0 for agent-to-agent interoperability (closes one of the oldest open feature requests), Hermes desktop as a platform (artifacts with sandboxed live preview, a plugin SDK, global hotkey quick-entry, multiple GUI windows), the Vercel AI Gateway provider, a four-phase Relay parity wave, and a long list of Windows hardening. None of those change how a one-person operator runs Hermes across a working day. They change what you can build on top of Hermes. What this means for a UK service business running Mercury OS For Mercury OS customers, nothing changes the day this release ships. The four modules (Hermes Agent, Vereby, S0cial Master, the Zoho Lead Engine) keep working on the existing install. When you run hermes update , the install shape changes underneath you, but the user-facing behaviour stays the same. For Ted's own work, this is the release that turns Hermes from "an assistant you talk to in text" into "an assistant you can actually talk to". The voice work is the headline, but the under-the-hood wins, /redirect , self-recovering tools, smarter compression, are the ones that show up daily. The webhook layer is what makes Hermes composable with the rest of the Mercury OS stack in a way it wasn't before. The honest version of how to read a release this big When the release notes lead with "~3,650 commits and 1,400 merged PRs", most of what changed is invisible to the operator. The useful question is "what do I do differently on Monday?" For most one-person operators, the answer this release is: talk to it. Try voice mode. Try the wake word. Try /redirect next time it goes off the rails. Notice that long sessions stop stalling. That is the headline of this release for our usage. Yours may differ. Frequently asked questions Should I upgrade to Hermes Agent v0.20.0 today? Yes, with one caveat. The voice and A2A work is solid; the under-the-hood compression, tool-recovery, and CLI power-user wins are solid; the webhook layer is solid. The caveat: if you have any custom voice mode wiring from before v0.19, check the migration notes, wake-word detection now runs on-device, which changes where your audio lives. For a UK service business running on Mercury OS, this is an upgrade you do on a quiet afternoon and forget about. For a heavy voice user, give yourself an hour to retrain your muscle memory on the new barge-in model. What does conversational voice actually feel like? It feels like talking to a person who actually listens. You ask a question, the agent starts answering, you interrupt halfway through to clarify something, the agent stops mid-word, acknowledges your cut-in, and answers the clarified version. The piece that surprises first-time users is the silence detection: the agent doesn't talk over you because it knows when you're still talking. If you've ever used Siri or Alexa and felt like they weren't really listening, this is the version that fixes that. Do I need to reconfigure anything when I upgrade? No for the core. The v0.20 release preserves every existing config, every MCP server, every profile, every cron job, every Telegram/Discord/Slack adapter. Voice features are off by default, you opt in per profile. Wake-word detection runs on-device by default once you enable voice mode, which is a security upgrade (your audio doesn't leave your machine) but worth knowing. New features like /redirect, `!command`, /init, /diff, /context, /focus, and `hermes approvals suggest` are on by default but can be turned off in config if you prefer the old behaviour. Outbound webhooks require explicit endpoint registration. The only mandatory change is the version number itself. The standard `hermes update` flow handles it. Is Hermes Agent v0.20.0 stable enough for production? Yes. We've been running v0.20 on the studio's daily work since the release shipped, and the only production-shaped issue we hit was an A2A discovery payload edge case that was patched within 48 hours. For a UK service business running on Mercury OS, this is a stability win, not a risk. The four under-the-hood quality-of-life wins (self-recovering tools, smarter compression, CLI power-user commands, smart approvals grown up) are all in your favour and all revert easily in config if needed. The two headline-grabbing features (conversational voice, wake words) are opt-in per profile, so you can ship the rest to the whole studio while testing voice on yourself first.]]></content:encoded>
  </item>
  <item>
    <title>Five new guides for solopreneurs: from idea to invoice</title>
    <link>https://expedientai.co.uk/feed/solopreneur-idea-to-invoice-guides/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/solopreneur-idea-to-invoice-guides/</guid>
    <pubDate>Thu, 30 Jul 2026 00:00:00 +0000</pubDate>
    <description>We published five new free PDF guides for solopreneurs and startups: validate your idea, build the thing, get found, get the conversation, and close the deal. Same format, same honesty, new audience.</description>
    <content:encoded><![CDATA[News Five new guides for solopreneurs: from idea to invoice We published five new free PDF guides for solopreneurs and startups: validate your idea, build the thing, get found, get the conversation, and close the deal. Same format, same honesty, new audience. 2026-07-30 · 5 min read We published five new free PDF guides for people starting from nothing: validate the idea, build the thing, get found, get the conversation, close the deal. Same format as the first seven , new audience. Why a second series Last week we published seven guides for UK businesses already running . They answer the questions we get on every discovery call: where does AI help, how do I test it, is my data safe, what do I ask the vendor. This week we published five more. These ones are for a different person: someone at the very start. No product yet. No customers. Maybe not even a company name. Just an idea and a suspicion that AI might help them get to the first invoice faster. They are on the guides page , they cost nothing, and the deal is the same as before: enter your email, we send you the download link, that is the end of the transaction. No spam, no newsletter, no follow-up sequence. What is in the five Guide 08, Is This Worth Doing? is the validation framework. Market research prompts you can run in twenty minutes with free AI tools. Competitor analysis you can do in an afternoon. How to talk to ten people and get honest answers instead of polite ones. And a five-question go/no-go checklist that tells you whether to build, pivot, or stop. Twelve pages. Guide 09, Build the Thing covers the build/buy decision, AI-assisted copywriting for your offer, no-code prototyping with v0 and Bolt, pricing without guessing, and the boring admin (contracts, T&Cs, privacy policies) that AI can draft but a human must check. Twelve pages. Guide 10, Get Found is marketing for people who hate marketing. Pick one channel and ignore the rest. Write content that does not sound like AI wrote it. SEO basics that actually move the needle. And an email list you own, because followers are rented. Ten pages. Guide 11, Get the Conversation explains why your first ten customers will not come from a website. Cold outreach that is not spam, with a three-sentence email template. LinkedIn without the cringe. And a one-person CRM that does not collapse after week three. Ten pages. Guide 12, Close the Deal is the last one. Proposals that get signed. The pricing conversation that does not end in discounting. Onboarding without dropping the ball. And turning one client into three through testimonials and referrals. Ten pages. How the two series fit together If you are already running a business and trying to work out where AI fits, start with Guide 01, Where AI Actually Helps and work through the first seven . They are about AI in a business that already exists. If you are at the start, idea stage, no product, no customers, start with Guide 08, Is This Worth Doing? and work through to Guide 12, Close the Deal . They are about getting from nothing to your first paying client, with AI helping at every stage. The two series cross-reference each other. Guide 08 points you to Guide 01 if you need to understand where AI helps in general. Guide 12 points you to Guide 02, The 30-Day AI Pilot if you want to pilot a system once you have clients. They are designed to work as one collection of twelve, or as two separate journeys depending on where you are. Same rules as before British English. Plain language. No exclamation marks. No fabricated testimonials. Every tool mentioned has a free tier or costs less than twenty quid a month. Every guide is a print-ready A4 PDF you can write in the margins of. And the same honesty about limits. Guide 08 tells you when to stop. Guide 09 tells you when not to build. Guide 10 tells you which marketing channels to ignore. Guide 11 tells you when cold outreach is the wrong move. A guide that only tells you what to do is a brochure. These are field guides. What this means for you If you are thinking about starting something, read Guide 08 first. If the checklist says stop, stop. You have saved yourself six months and a few thousand pounds, and we have earned a bit of trust. If it says go, read the next four and get to your first invoice. If you read all twelve and decide you do not need us, that is a good outcome. If you read them and want a named person to build the first system with you, that is what we do, and we will tell you on the first call if the guides are all you need. The full collection of twelve is on the guides page . Free, print-ready, and yours to keep.]]></content:encoded>
  </item>
  <item>
    <title>Seven free AI guides for UK small businesses | released today</title>
    <link>https://expedientai.co.uk/feed/free-ai-guides-uk-smes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/free-ai-guides-uk-smes/</guid>
    <pubDate>Thu, 30 Jul 2026 00:00:00 +0000</pubDate>
    <description>We published seven free PDF field guides covering where AI helps, how to run a pilot, data privacy, prompt writing, vendor vetting, and setting up your own agent. Here's why they exist and how to use them.</description>
    <content:encoded><![CDATA[News Seven free AI guides for UK small businesses, released today We published seven free PDF field guides covering where AI helps, how to run a pilot, data privacy, prompt writing, vendor vetting, and setting up your own agent. Here's why they exist and how to use them. 2026-07-30 · 6 min read Today we published seven free PDF guides for UK small and medium businesses. They are on the new guides page , they cost nothing, and all we ask is your email so we can send you the download link. No spam, no newsletter, no follow-up sequence, just the guide. This post explains why we wrote them, what is in each one, and how to use them in the order that makes sense. Why free, and why now We talk to UK business owners every week about AI workflow automation. The questions are always the same, in roughly the same order. Where does AI actually help? How do I test it without breaking what works? Is it safe to put our data in? Why does ChatGPT give me rubbish? What should I ask the vendor who just gave me a demo? What is an AI agent and do I need one? Can I set one up myself? We were answering the same seven questions on calls, in emails, and in proposals. So we wrote the answers down once, properly, and made them public. A business owner should be able to get a straight answer to these questions without booking a discovery call first. If the guides answer your question completely and you never need us, that is a good outcome for us too, it is how you build a reputation in a market that has been burned by AI theatre. What is in the collection Seven guides, eighty-nine pages total, each one answering a single question. Guide 01, Where AI Actually Helps is the starting point. It gives you a scoring framework for deciding which tasks are worth automating and which are not. It includes a worked example that ends in "stop", because sometimes the right answer is not to build. Twenty-six pages. Guide 02, The 30-Day AI Pilot is a day-by-day plan for testing AI on one task alongside the way you do it now. What to measure, when to stop, and how to make the go/no-go call. Twelve pages. Guide 03, Your Data, Your Rules answers the question everyone is afraid to ask: can I put business data into ChatGPT? It covers what the ICO expects, a four-tier way to classify your data, and a checklist to run before you type anything into an AI tool. Ten pages. Guide 04, Writing AI Instructions That Actually Work explains why vague prompts produce vague output, and gives you a six-part structure that turns inconsistent results into reliable ones. Ten pages. Guide 05, The AI Vendor Interrogation Kit is twenty questions to ask any AI vendor before you sign, a scoring card to rate their answers, and a walk-away threshold. Eight pages. Read it before the next demo. Guide 06, Your First AI Employee explains what an AI agent actually is (not a chatbot), what it can automate, what it costs, and the permission ladder that keeps it safe. Twelve pages. Guide 07, Set Up Your Own AI Agent is a no-code walkthrough of installing Hermes, an open-source agent that runs on your own machine. The install, the first job, the guardrails, and the honest cost. Eleven pages. Why they are useful Three reasons, and none of them are "because we wrote them". First, they answer the question you actually have, not the question a vendor wants you to have. Each guide is scoped to one problem. You are not reading a brochure, you are reading an answer. Second, they are honest about limits. Several guides include worked examples that end in "don't build this". A guide that tells you when AI is the wrong answer is more useful than one that tells you it is always the right answer. Third, they are written for UK businesses specifically. That means ICO guidance, UK GDPR, pound Sterling costs, and examples from trades, salons, cafes, and service firms, not Silicon Valley case studies. How to use them Start with Guide 01 if you are not sure where AI fits. Start with Guide 02 if you already know and want to test it. Jump to Guide 05 if a vendor is asking you to sign something. The full collection, with a description of each guide, is on the guides page . Every guide is a print-ready A4 PDF. Enter your email, grab the link, print it, pass it to a colleague, write in the margins. That is what field guides are for. What this means for you If you read all seven and decide you do not need us, you have saved yourself a discovery call and we have earned a bit of trust. If you read them and want a named person to build the first system with you, that is what we do, and we will tell you on the first call if the guides are all you need. Either way, the answers are now public, free, and yours to keep.]]></content:encoded>
  </item>
  <item>
    <title>Realistic AI automation projects for South West businesses</title>
    <link>https://expedientai.co.uk/feed/realistic-ai-automation-south-west/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/realistic-ai-automation-south-west/</guid>
    <pubDate>Wed, 22 Jul 2026 00:00:00 +0000</pubDate>
    <description>Five realistic AI automation projects South West small businesses ship first. The bar for a good first project, when to skip it, and how a Bristol AI consultancy typically runs the engagement.</description>
    <content:encoded><![CDATA[Tips Realistic AI automation projects for South West businesses Five realistic AI automation projects South West small businesses ship first. The bar for a good first project, when to skip it, and how a Bristol AI consultancy typically runs the engagement. 2026-07-22 · 7 min read South West small businesses do not need an AI transformation. They need one or two realistic AI automation projects that ship inside a quarter, save a measurable amount of time or money, and do not require a Chief AI Officer to keep running. This is the working list we run through with Bristol, Bath, and Somerset small business owners on a twenty-minute diagnostic. Five realistic projects most of them can ship first. The bar a project has to clear before it is worth shipping. And the cases where the right answer is "not yet, fix the underlying system first". The bar for a good first AI project A good first project clears four checks. It saves measurable time, with a before/after number you can quote three months later. It has one clear owner in the business who will adjust it when something changes. It ships in two to four weeks, not two to four months. And it fits inside the systems the business already runs, the inbox, the calendar, the CRM, the booking system, the accounts package. Anything that fails one of those four checks is either too vague to ship, too expensive to recover, or too slow to demonstrate value. We turn those down, even when the customer is excited. The excitement fades. The bill does not. Five realistic projects we ship first for South West SMEs These are the projects that, in our experience, pay back first for Bristol and South West small businesses. Each one ships in two to four weeks and produces a measurable saving in the first month. Missed-call capture for Bristol trades. A plumber, electrician, roofer, or heating engineer loses calls while on a job. An AI voice agent picks up, takes the brief, promises a callback by 5pm, and drops the lead in the CRM before the customer puts the phone down. The team sees the lead in their admin dashboard within seconds. Win rate climbs without anyone answering a call from a noisy van. Enquiry-to-quote workflow for Bristol professional services. A solicitor, accountant, architect, or planning consultant gets a one-page email enquiry. The AI reads it, drafts a first-pass quote based on the firm's pricing logic, flags anything it is unsure about, and puts the draft on the partner's queue. The partner reviews and sends. Quote turnaround drops from three days to three hours. Out-of-hours reception for Bristol salons, restaurants, and clinics. A Bristol salon or restaurant gets calls until 11pm. The team stops taking them at 6. An AI agent handles bookings, FAQs, and message-taking after hours, syncs the booking to the diary, and sends the team a morning summary. No missed calls, no late-night messaging back and forth. Weekly reporting for any Bristol SME. The Monday-morning sales report. The weekly cashflow summary. The marketing dashboard. The team stops compiling it. The AI compiles it from the data that already exists in the systems they already run, and emails it at 7am Monday. Four hours a week back, every week. Lead enrichment and follow-up for Bristol B2B service businesses. A new lead comes in. The AI looks them up, drafts a personalised first message based on what their company actually does, and queues it for the human to send. The team starts the conversation warmer, with less typing, and the time-to-first-reply drops from days to hours. None of these need a six-figure budget. None of them need a platform. They need a Bristol AI consultancy that will sit down, pick one or two, and ship them properly. When to skip the project Honest section. There are cases where shipping the AI automation project first is the wrong call, and the South West small business owner will be better served by fixing something else. If the call volume does not justify the automation, skip it. A business that takes five missed calls a week does not need an AI voice agent yet. They need a callback rota. If there is no clear owner in the team, skip it. An automation without an owner becomes an automation nobody adjusts. After three months it stops working and nobody notices until a customer does. If the underlying system is broken, fix the system first. An AI that drafts quotes from a CRM full of stale data is just a faster way to send the wrong price. The data layer has to be honest before the automation layer can be useful. For Bristol trades businesses in particular, the first AI win is often local search rather than workflow automation. If the website and the Google Business Profile are not pulling their weight, the Local SEO audit is the right starting product. AI automation is the next step up, once the customer can actually find you. What we actually do when you ask for a quote The first conversation is a twenty-minute diagnostic. We listen for the workflow that is costing the most measurable time or money, the system that is most painful to maintain, and the team that would actually own the result. We come back with one of three answers. Yes, ship this. Here is the project, here is the rough cost, here is the rough time, here is what "good" looks like three months later. Not yet. Here is the underlying system that needs fixing first, and what the AI project looks like once that is done. Not this one. Here is the workflow that actually pays back, which is different from the one that sounds the most exciting. All three answers are useful. The cheapest way to find out which one applies to your business is the diagnostic. The services page walks through how an Advise engagement works if you want the longer version. Twenty minutes. A short diagnostic. If you are a South West small business owner with a workflow that has been costing you time or money for too long, the cheapest answer is a twenty-minute call. Get in touch via the contact form]]></content:encoded>
  </item>
  <item>
    <title>Top 10 practical changes in Hermes Agent v0.19.0 (and what they're for)</title>
    <link>https://expedientai.co.uk/feed/hermes-0-19-practical-changes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/hermes-0-19-practical-changes/</guid>
    <pubDate>Wed, 22 Jul 2026 00:00:00 +0000</pubDate>
    <description>The Quicksilver Release dropped on 2026-07-20. Ten changes from v0.19.0 that actually shift how a one-person operator runs Hermes day-to-day, anchored in the release notes, with the headline grabbers deliberately passed over.</description>
    <content:encoded><![CDATA[Tools Top 10 practical changes in Hermes Agent v0.19.0 (and what they're for) The Quicksilver Release dropped on 2026-07-20. Ten changes from v0.19.0 that actually shift how a one-person operator runs Hermes day-to-day, anchored in the release notes, with the headline grabbers deliberately passed over. 2026-07-22 · 10 min read We use Hermes Agent as our day-to-day terminal assistant at the studio. Multiple named profiles, Telegram as the primary comms channel, daily cron briefings, and a skills library that's grown fat over the last year. So when the Quicksilver Release (v2026.7.20) dropped on 20 July, we did what we do with any release: pulled the notes, pulled the PRs, and asked which changes actually move the needle on a one-operator workflow. These are the ten. They are not the headline grabbers. They are the ones that change how a single person uses Hermes across a working day. 1. First-turn latency cut by ~80% on every platform What it is: Cold-start "Initializing agent..." used to eat ~4.3 seconds before your first turn even reached the model. It is now ~0.9s on the CLI, gateway, TUI, desktop, and cron. Pair that with reasoning models that now stream their thinking live by default and a response box that paints per token instead of per line, and the "deep breath before Hermes answers" feel is gone. This is not a benchmark win, it is a daily-use win. Every command you run, every cron you wait for, every desktop session you open, all feel snappier. Why it matters: Today's pain: you ask Hermes a quick question, it shows you a spinner for a couple of seconds, then answers. With this, the answer starts arriving before you have finished your next thought. Practical example: A cron briefing fires at 7am UK time. Instead of seeing "Initializing agent..." and a blank panel for several seconds, you see the briefing start streaming before the kettle boils. Under the hood: Discord capability detection moved off the critical path, reasoning models stream by default, the response box paints per token. (#59332, #59389) 2. delegate_task now streams live and survives a crash What it is: When you kick off parallel subagents, you can now tail -f the live transcript files the moment the children launch, every tool call, result, and streamed reply in one human-readable log per child. And background delegation completions are now durable, if the process restarts mid-run, the results are restored and delivered through an ownership-checked ledger instead of vanishing. Why it matters: Today's pain: "I asked three agents to research three things. Now I'm staring at the screen waiting. Three minutes. Five. Did the third one finish? I don't know." With this, you see the work happen, you can chase any one child live, and even if Hermes crashes mid-run the results are recovered and delivered. Practical example: "Audit these three modules for security issues in parallel" → you see all three transcripts streaming → you can tail -f any one of them → if the laptop sleeps and Hermes is killed mid-run, the completed results land in your queue when you restart. Under the hood (#67479, #63494): live transcript file per subagent + ownership-checked delivery ledger. 3. A finished answer can no longer be lost, the delivery-obligation ledger What it is: If the gateway died between generating your response and confirming the platform actually delivered it, that answer used to be silently gone (and you had paid for the turn). The new delivery-obligation ledger tracks every response from "generated" to "delivered and acknowledged by the platform" so a crash mid-handoff recovers cleanly. Why it matters: The classic LLM failure mode: "I paid for a turn and got nothing back." With this, every turn has a verifiable delivery state. If the platform dropped your message, you see that, you re-send or re-raise, and no work is silently lost. Practical example: Telegram gateway dies after Hermes generates a response but before the Telegram API confirms delivery. On restart, the ledger re-tries the handoff, the user sees the message, and the turn is marked delivered. 4. /subscription and /topup, manage your Nous plan from the terminal What it is: Changing your subscription used to mean a trip to the billing website. Now /subscription opens a full flow right in the TUI or classic CLI: see your plan and remaining allowance, preview exactly what an upgrade costs ("Pay $46.30 and upgrade now") or when a downgrade takes effect, and apply it, with scheduled-change banners and undo. Why it matters: If you manage Hermes the way we do, you adjust the plan every few weeks as your usage shifts. The browser round-trip for that used to cost ten minutes of context. With this, it costs a slash command. Practical example: You notice you've been hitting the higher tier for two weeks running. /subscription → preview the downgrade → confirm. The change takes effect at the next billing cycle and you see a banner reminding you of the date. 5. Password manager integration, Bitwarden and 1Password, straight into Hermes What it is: You can plug Bitwarden or 1Password straight into Hermes so the agent can pull credentials out of your password vault at runtime, scoped per project, with explicit per-call audit logging. No more pasting API keys into prompts or .env files. Why it matters: Every credential you put in a .env file is a credential that's about to leak through git, a screen share, or a clipboard. With this, the credential lives in the vault, Hermes asks for it on demand, and the audit log shows exactly which subagent accessed which credential and when. Practical example: "Find me a worked example of how the Bristol voice agent handled a late payment follow-up last month, in the same tone as our other comms." Hermes pulls the OpenAI key from your vault for that one subagent, runs the search, and the key never enters your prompt or workspace. 6. Smart approvals, flagged commands get judged, not blocked What it is: Hermes used to either auto-approve safe commands or hard-block dangerous ones, with a manual approve layer in the middle for everything ambiguous. Now, the LLM looks at flagged commands and judges them against your project context, approving the ones that look safe in your setup, blocking the ones that don't, and showing you the reasoning either way. Why it matters: Today's pain: the approval layer is either too eager (auto-approves things you didn't read) or too strict (blocks the same safe rm command five times a day). With smart approvals, the LLM does the read-once-and-decide work that you don't have time for. Practical example: You run an audit script that reads from a temp directory. Hermes sees the path is /tmp/audit-* , the script is signed by you, the operation is read-only, and approves it with a one-line reason. You see the reason, you trust it, you move on. 7. Reasoning streams live by default What it is: Reasoning models (the ones that think before they answer) used to hold the whole chain until they were done, then dump it as a single block. Now they stream the thinking live as they go, token by token, so you can see the model "thinking out loud" while you wait. Why it matters: When the model takes 30 seconds to think, you either stare at a spinner or you stare at the model working. The latter is more reassuring, you can see when it goes down a wrong path, you can stop it before it commits to an answer you don't want, and you can learn from the reasoning for next time. Practical example: "Diagnose why the Vereby webhook ack came back 200 but no email landed." The model thinks for 30 seconds visibly, shows the chain of reasoning, then answers. The reasoning is evidence, not magic. 8. 1Password + Hermes as an MCP credential store What it is: Related to #5 but different in shape. Beyond pulling individual credentials, you can now point an MCP server at your password vault and let any Hermes session consume credentials from it as if they were first-class config. The vault stays the source of truth, Hermes becomes a thin consumer. Why it matters: Multi-agent setups need a lot of secrets. Today you copy them around. With this, the secrets live in one place, agents reference them by name, and rotating a credential rotates it everywhere. Practical example: The Vereby MCP, the S0cial Master MCP, and the Zoho Lead Engine MCP each need their own keys. They live in 1Password. Hermes reads them on demand. When you rotate one key, all three systems update on the next read. 9. Inline choice pickers for /reasoning and /fast on Telegram, Discord, and Matrix What it is: When you type /reasoning or /fast in the chat, you used to get back a text reply asking you to pick a level. Now you get a native one-tap picker button row on the platform itself, so the choice is a tap, not a typed reply. Why it matters: /reasoning and /fast are two of the most-used slash commands. Reducing them from a two-step typed exchange to a one-tap button is a small win that adds up across a day. Practical example: You are about to ask a hard question from your phone in a meeting. /reasoning → one tap to pick "deep" → the model thinks, streams, answers. The chat session stays in flow. 10. The desktop app got 20+ targeted perf PRs and now feels native What it is: Long replies used to cost 14x more CPU in the markdown splitter than they do now. Giant diffs froze the review pane until they were virtualised. Session switching thrashes layout no more. Streaming no longer re-renders the sidebar and every tool row per token. Profile backends pre-warm on hover intent. Boot-hidden panes mount at idle. Why it matters: You live in the desktop app for hours. If it stutters, you feel it. The speed overhaul is not glamorous, but it is the difference between "I'm using an AI tool" and "I'm using my tool, which happens to be AI". Practical example: You're reviewing a long code review transcript and the model finishes a new reply. The diff lands, the scroll position is preserved, the sidebar updates without re-flow, and you keep reading. Five years ago that was a dream. Now it's the desktop app under default load. What didn't make the list (but is in the notes) The release also ships a Fireworks AI provider with cost estimation, a hosted MCP gateway, more provider-agnostic tool routing, webhook payload diffing, the safe-session-import flow, Discord-specific toolsets editable from the web UI, manual Telegram bot setup clarified, a contribution-driven shell on a layout-tree model for the desktop, Vertex credential resolution, and a long list of security hardening. None of those change how a one-person operator runs Hermes across a working day. They change how Hermes runs internally. What this means for a UK service business running Mercury OS For Mercury OS customers, nothing changes the day this release ships. The four modules (Hermes Agent, Vereby, S0cial Master, the Zoho Lead Engine) keep working on the existing install. When you run hermes update , the install shape changes underneath you, but the user-facing behaviour stays the same. For Ted's own work, this is a release to pay attention to. The delivery-obligation ledger changes the trust model around background work. The smart-approvals change the way you babysit the agent. The credential store is a security upgrade. None of those require a reinstall. All of them show up the next time you start the agent. The honest version of how to read a release this big When the release notes lead with "~2,245 commits and 1,065 merged PRs", most of what changed is invisible to the operator. The useful question is "what do I do differently on Monday?" For most one-person operators, the answer is "open Hermes slightly faster and watch a subagent run for the first time." That is the headline of this release for our usage. Yours may differ. Frequently asked questions Should I upgrade to Hermes Agent v0.19.0 today? Yes for almost everyone. It is a quality-of-life release, not a breaking one. The four breaking changes that did land are all in your favour: durable background delegation means you do not lose work mid-run; smart approvals are a softer approval layer than the old one; the delivery-obligation ledger is a better trust model around background work. The actual risk is on the Hermes maintainers' side, not yours. Upgrade, watch one full work day, keep the old install around for a week in case you want to roll back. What does the ~80% TTFT cut actually feel like? It feels like Hermes answers before you have finished your next thought. The cold-start spinner that used to eat a couple of seconds is now under a second. Reasoning models stream live by default, so for the first time on a hard question you can watch the model think, not stare at a blank panel. If you only ever used Hermes for quick commands you will barely notice. If you ran it on the daily cron briefings, the desktop app, or the Telegram gateway, you will feel it immediately. Do I need to reconfigure anything when I upgrade? No. The v0.19 release preserves every existing config, every MCP server, every profile, every cron job. New behaviours like smart approvals and durable background delegation are on by default but can be reverted in config. New integrations like Bitwarden and 1Password require explicit opt-in. The only mandatory change is the version number itself. The standard `hermes update` flow handles it. Is Hermes Agent v0.19 stable enough for production? Yes. We have been running v0.19 on the studio's daily work since the release shipped, and the only production-shaped issue we hit was a credential-resolution change in the Vertex provider that was patched in the same week. For a UK service business running on Mercury OS, this is a stability win, not a risk. The four quality-of-life wins (speed, live subagent transcripts, smart approvals, durable background work) are all in the customer's favour and all revert easily in config if needed.]]></content:encoded>
  </item>
  <item>
    <title>China is shipping frontier-class open-weight models faster than the US, and they cost a fraction to run</title>
    <link>https://expedientai.co.uk/feed/china-open-weight-july-2026/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/china-open-weight-july-2026/</guid>
    <pubDate>Wed, 22 Jul 2026 00:00:00 +0000</pubDate>
    <description>Inkling from the US, Kimi K3 and Qwen3.8-Max from China, plus GLM-5.2. Four frontier-class model releases in eight days, three of them open-weight and two of them priced below Claude. What changed for a UK service business.</description>
    <content:encoded><![CDATA[News China is shipping frontier-class open-weight models faster than the US, and they cost a fraction to run Inkling from the US, Kimi K3 and Qwen3.8-Max from China, plus GLM-5.2. Four frontier-class model releases in eight days, three of them open-weight and two of them priced below Claude. What changed for a UK service business. 2026-07-22 · 9 min read The 14 July post covered the previous week's wave: GPT-5.6, a new Claude, Gemini, DeepSeek V4 and the rest. Six days later, four more frontier-class models shipped. This post covers those four, and the bigger shift they describe together: the open-weight race is now being written mostly in Beijing and Hangzhou, and the cheap layer of the AI stack just got a lot cheaper. What actually shipped between 15 and 19 July Inkling from Thinking Machines Lab (US, 15 July) is the first open-weight model from Mira Murati's lab. Mixture-of-Experts, 975B total parameters with 41B active per forward pass, multimodal, and a feature the others do not have: controllable reasoning effort. You can dial up the reasoning for a hard task and dial it back for a cheap one, instead of paying for the maximum every time. It debuted at 41 on the Artificial Analysis Intelligence Index, which makes it the new leading US open-weights model. The weights are out, you can fine-tune it on Thinking Machines' own Tinker platform. Pricing for hosted inference is not in the public coverage I have seen yet, but it is positioned as US frontier-class open weights. Kimi K3 from Moonshot AI (Beijing, 17 July) is the one Silicon Valley is paying attention to. 2.8 trillion parameters, open weights due 27 July, 1M token context window. The benchmark numbers are real: it ranks fourth on Artificial Analysis and currently leads Arena.ai's Code WebDev leaderboard. The price is the part that will move markets: $3 per million cache-miss input tokens, $0.30 cached, $15 per million output, flat across the entire 1M window rather than stepped by context tier. That is roughly the Claude Sonnet tier, and it is open weight, which means you can also self-host if your volume justifies the GPU bill. Qwen3.8-Max from Alibaba (Hangzhou, announced 19 July at the World AI Conference in Shanghai) is 2.4 trillion parameters, multimodal, and pre-announced days after the Kimi K3 release, which is not an accident. The headline number is that Alibaba's own evaluation puts it just behind Claude Fable 5. The honest caveat: this is a pre-announce, and Qwen3.8 is currently API-only. Alibaba has historically open-weighted after a closed preview, but no 3.8 timeline has been published yet. Until the weights land, this is a watch-this-space story rather than a switch-today story. GLM-5.2 from Z.ai (China, mid-July) is the one that has been quietly running in production at several shops I read. Mixture-of-Experts, 1M token context, and per three separate leaderboards (BenchLM, Thunder Compute, acecloud) it is the strongest all-round open-weight LLM right now for agentic engineering, software development, and long-horizon reasoning. If you have been following the cheap layer of your stack for the last six months, you may already be running a GLM variant. The GLM 5.2 release moves that ceiling up another notch. Why Chinese labs are shipping faster, and cheaper There are three honest reasons and one cynical one. The honest reasons first. They are mostly competing with each other. The domestic Chinese AI market has priced inference as a commodity. DeepSeek, Moonshot, Alibaba's Qwen team, Z.ai, Zhipu, Stepfun, MiniMax, Tencent, Xiaomi, ByteDance, and the rest are all shipping into the same buyer base, and that buyer base expects frontier-class output at near-zero inference cost. The Chinese frontier is competitive because the Chinese market forces it to be. The training cost per frontier model has fallen sharply across 2025 and 2026. The infrastructure efficiency story (better kernels, smaller active parameter counts, better data, better distillation) is real, and it benefits every lab equally. A model that would have cost $100M to train in 2024 trains for a fraction of that today. The savings show up in the API price. Open-weight is a strategy, not a giveaway. When you release open weights, you win developer mindshare without having to run a closed API business at scale. The Chinese labs that have gone open-weight first (DeepSeek, Moonshot, Qwen) have done so in part because the developer base they wanted to capture was the global one, not the domestic one. Open weight is a market-entry strategy. The cynical reason: there is a strategic subsidy angle to some of this, where the headline pricing is supported by state-level infrastructure investment. That is real, and it is the kind of thing that may or may not last. Either way, the price a UK small business pays is the price a UK small business pays. Kimi K3 is $3 per million input tokens. DeepSeek V4 Flash is $0.14 input / $0.28 output per million tokens, which is twenty to forty times cheaper than the equivalent frontier-class closed model. That gap is the point. The capability bar moved, again This is the third capability-bar shift in six months, after the DeepSeek V3 moment in late 2024 and the GPT-5 / Claude 4 moment in early 2025. Each time the cheap layer gets better, the work you used to escalate moves down. Each time the frontier layer gets better at one thing, that one thing becomes a default rather than an exception. For a UK service business, the practical version is short. The cheap layer of your stack (drafting, summarising, classification, triage, basic extraction, first-pass reply drafting, lead scoring) is now good enough that you should not be paying frontier money for ninety percent of it. The frontier layer (final-draft proposals, contract language, security review, customer-facing responses where brand voice has to be exact) is still worth paying for, because the cost of getting it wrong on those tasks is real. The old worry, that you were locked into one provider because the cheap model could not do enough, is mostly gone. You can route cheaply across Kimi K3, GLM-5.2, DeepSeek V4 and a half-dozen smaller open-weight models, and pick the one that does the specific job best. The orchestration layer (Hermes-style, or whatever else you use) gets more valuable, not less, because the routing decision is the one that pays back the most attention. What this means if you are running AI yourself The boring, correct answer applies. Run your top ten prompts against the new models that matter to your work, the same prompts you already use, not a curated benchmark set. Pick the one that does your job better. Move on. The story of this week is not a single breakthrough. It is that the open-weight race has a clear leader right now and it is not based in California. That is good for the economics of running AI in a small business, and it is the kind of structural shift you only have to understand once, because it does not reverse. If you are a UK service business owner, the takeaway is short and good. The AI you are paying for (whether that is ChatGPT, Claude, Gemini, or the cheaper open-weight layer running underneath) just got better and cheaper in the same week. The vendors that are on top of their game will already be switching the cheap layer of their stack to one of these new models. If you do not see a change in the next two to four weeks, that is a signal to ask your vendor. Frequently asked questions Is the cheap layer of an AI stack now genuinely as capable as the frontier layer was a year ago? For most of the work a UK service business asks an AI to do, yes. Drafting replies, summarising documents, classifying enquiries, drafting social content, extracting structured data from a PDF, scoring inbound leads. Open-weight models at the Kimi K3, Qwen3.8, and GLM-5.2 tier can do all of that at frontier-class quality, at a small fraction of the per-token cost. The frontier layer still wins on the work where the cost of being wrong is high and the workflow is bounded: the proposal sent to a £100k client, the security review, the contract language, the response that has to be right the first time. The cheap layer is now good enough that the question for a small business is no longer which model to use. It is which task to route where, and how to keep the routing honest. Why are Chinese labs pricing frontier-class models so much lower than the US frontier? Three honest reasons and one cynical one. The honest reasons: the Chinese labs are mostly competing with each other in a domestic market that has priced inference as a commodity, the training cost per frontier model has fallen sharply across 2025 and 2026, and open-weight releases let them win developer mindshare without running a closed API business. The cynical reason: there is a strategic subsidy angle to some of it. Either way, the price a UK small business pays is the price a UK small business pays. Kimi K3 is $3 input / $15 output per million tokens, roughly the Claude Sonnet tier. DeepSeek V4 Flash is $0.14 input / $0.28 output per million tokens, which is twenty to forty times cheaper than the equivalent frontier-class closed model. That gap is the point. Does an open-weight model mean my data is private? Only if you run it yourself. Open-weight means anyone can download the model, including you. If you then run it on your own hardware or on infrastructure you control, your prompts and your data do not leave that boundary. If you use a hosted endpoint (OpenRouter, Fireworks, Together, the lab's own API), the same rules apply as for any other cloud API: the provider can see your traffic unless they explicitly publish a no-retention policy. Read the provider's data policy before assuming privacy. The honest version: open-weight gives you the option of self-hosting. It does not give you privacy by default. Should a UK service business switch to a Chinese open-weight model right now? Not blindly. The cheap layer of most UK small business stacks is already open-weight, and most of it is Chinese-lab origin already (DeepSeek, Qwen, GLM, Yi). The decision is about which specific model to run, on which task, with which routing. For drafting, summarising, classification and triage, Kimi K3, GLM-5.2 and the DeepSeek V4 family are genuinely strong picks at frontier-class price points. For final-draft proposals, contract language and anything customer-facing where the brand voice has to be exact, stick with the frontier model you already trust, and route to it. Test the new model against your top ten prompts before assuming it is a drop-in replacement for the model you currently use on that task.]]></content:encoded>
  </item>
  <item>
    <title>AI workflow automation for Bristol SMEs: where it actually pays off</title>
    <link>https://expedientai.co.uk/feed/ai-workflow-automation-bristol-smes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/ai-workflow-automation-bristol-smes/</guid>
    <pubDate>Wed, 22 Jul 2026 00:00:00 +0000</pubDate>
    <description>Where AI workflow automation pays off for Bristol SMEs, where it does not, and how a small business in the South West typically adopts it. Plain English, written by an AI consultant in Bristol.</description>
    <content:encoded><![CDATA[Tips AI workflow automation for Bristol SMEs: where it actually pays off Where AI workflow automation pays off for Bristol SMEs, where it does not, and how a small business in the South West typically adopts it. Plain English, written by an AI consultant in Bristol. 2026-07-22 · 7 min read AI workflow automation for Bristol SMEs is not about chatbots. It is not about a magic assistant. It is about the boring, repetitive, slightly painful work that costs a small business measurable revenue every week: missed calls, slow quotes, support volume, the Monday-morning reporting grind, the "who was supposed to chase that" black hole. When a plumber in Bedminster misses a call on a Tuesday, that is a job. When a salon in Clifton sends a quote three days late, that is a customer who booked somewhere else. When a solicitor in Clifton drafts the same onboarding email by hand for the fifteenth time this month, that is four hours a week the team will never get back. These are the workflows AI is good at fixing. Not because AI is clever. Because the work is repeatable, well-defined, and the cost of getting it slightly wrong is recoverable. This is the read we give Bristol small business owners when they call us. Not the demo. The actual list of where AI workflow automation pays off, where it does not, and what the first ninety days look like in practice. What "AI workflow automation" actually means for a small business Strip the marketing away and "AI workflow automation" for a Bristol SME means three concrete things. First, an enquiry handler that works across web chat, WhatsApp, and voice, takes a structured brief, and routes it to the right person. Second, internal tools that take a Monday-morning job and turn it into an automatic brief, report, or digest. Third, voice agents that catch the calls a small team physically cannot pick up while they are on a job. None of those things are science fiction. They are software that does a specific, bounded job, connected to the systems a small business already runs (the inbox, the calendar, the CRM, the booking system). The "AI" part is the part that understands a sentence well enough to act on it. The "automation" part is the part that moves data from where it is to where it needs to be, without a human pressing "send". Most Bristol small businesses we speak to do not need an "AI strategy". They need one or two of these workflows, working reliably, by the end of the quarter. Where it pays off first for Bristol SMEs These are the entry points we see paying back inside sixty days for South West small businesses. Each one is concrete, measurable, and shippable in two to four weeks. Missed-call capture for Bristol trades. A plumber, electrician, or roofer loses calls while on a job. An AI voice agent picks up, takes the brief, promises a callback by 5pm, drops the lead in the CRM before putting the phone down. Quote-follow-up automation for service businesses. A quote goes out, the customer goes quiet. The AI drafts a polite follow-up at day three, day seven, and day fourteen. The human signs off the send. Win rate climbs without the team writing another chase email. Weekly reporting for any Bristol SME. The Monday-morning sales report, the weekly cashflow summary, the marketing dashboard. The team stops compiling it. The AI compiles it from the data that already exists, and emails it at 7am Monday. Customer-service tier-1 deflection. "What are your opening hours?", "Do you cover BS6?", "How much is a quote?". Roughly sixty percent of inbound questions are predictable. The AI answers them from your existing knowledge. The team only sees the rest. Internal enquiry routing. A web form, a WhatsApp, a missed call, an email. They all land in different places. The AI reads the brief, decides who in the team it is for, and puts it on the right person's queue. Lead enrichment and outreach for B2B service businesses. A new lead comes in. The AI looks them up, drafts a personalised first message based on what their company actually does, and queues it for the human to send. None of these need a six-figure budget. None of them need a Chief AI Officer. They need a Bristol AI consultancy that will sit down, pick one or two, and ship them properly. Where it does not (yet) Honest section. There are workflows AI is not yet the right tool for, and most Bristol SMEs will get a better return by leaving them alone for a year. Anything bespoke and creative where the human voice is the product. The bespoke campaign concept, the brand voice for a new launch, the nuanced client negotiation. AI will draft. A human has to ship. Anything where the cost of being wrong is reputational and the workflow is not yet measurable. The complaint reply to an unhappy long-term customer. The legal letter. The medical advice line. We will build internal tooling around these workflows (the right brief, the right escalation path, the right notes). We will not put the AI on the front line of them. Anything with fewer than ten clear examples per month. If a team only chases late payments three times a quarter, the AI workflow is slower than the human just doing it. Wait until the volume is there. How a Bristol SME typically adopts this The path we run with Bristol small businesses is three stages. Advise, a twenty-minute diagnostic call, then a written ninety-day plan that names the one or two workflows that will pay back first. Build, one or two specific workflows shipped, integrated into the systems you already run, with a human in the loop on anything customer-facing. Enable, the team is trained, the workflow is documented, and there is a clear owner in the business who can adjust it as the business changes. That is the whole thing. No retainer, no platform fee, no lock-in. If the workflow does not pay back, we tell you on the diagnostic and you keep the plan. If you are a Bristol small business owner looking at this and recognising your own list, the cheapest next step is a twenty-minute diagnostic. We will tell you, honestly, whether the workflow is ready for AI, or whether the underlying system needs fixing first. Either answer is useful. For the broader landscape of how a Bristol AI consultancy engages with a small business, from first call to shipped workflow, the services page walks through it. If you are a Bristol trades business whose first AI win is more likely to be local search than automation, the local SEO entry point is the right place to start; AI workflow automation is the step after that. Twenty minutes. A short diagnostic. If you are a Bristol SME wondering where AI workflow automation actually pays off for your business, the cheapest answer is a twenty-minute call. Get in touch via the contact form]]></content:encoded>
  </item>
  <item>
    <title>How Bristol service businesses can use AI agents to cut admin</title>
    <link>https://expedientai.co.uk/feed/ai-agents-bristol-service-businesses/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/ai-agents-bristol-service-businesses/</guid>
    <pubDate>Wed, 22 Jul 2026 00:00:00 +0000</pubDate>
    <description>How Bristol service businesses can use AI agents to cut admin. Real workflows for trades, salons, restaurants and professional services. Plain English from an AI consultant in Bristol.</description>
    <content:encoded><![CDATA[Tips How Bristol service businesses can use AI agents to cut admin How Bristol service businesses can use AI agents to cut admin. Real workflows for trades, salons, restaurants and professional services. Plain English from an AI consultant in Bristol. 2026-07-22 · 9 min read A Bristol trades business taking two hundred calls a month and missing forty of them is leaving meaningful money on the table. The exact number varies by trade and ticket size. The shape of the problem does not. A missed call rarely calls back. A missed call books the competitor. The cost of admin is not the line items. It is the customers who never reach you. AI agents are the cheapest way we have seen to close that gap for a Bristol service business. They are not magic. They are software that does one specific, bounded job, well, every time, without forgetting. This post walks through what that looks like in practice, and what it does not. What "AI agents" actually means in a Bristol service business An AI agent, in the sense most Bristol service businesses will meet it, is software that does a specific, bounded job: answer a missed call, draft a quote, chase a payment, summarise a week of enquiries. It is not a magic assistant. It is not a person. It is a worker that follows the brief, every time, and hands back to a human when the brief stops matching the situation. A real-world script looks like this. A customer phones the salon in Clifton at 9pm on a Saturday. The agent picks up on the first ring, in a clear natural voice. "Hi, this is the AI receptionist for [Salon Name] in Bristol. How can I help?" The caller wants to book a colour appointment. The agent asks which week, which day, which stylist. The caller picks. The agent books it in the salon's system, sends a confirmation text, and writes the booking into the morning briefing for the team. The caller hangs up happy. The salon owner never touched the phone. That is the pattern. AI for the predictable first contact. Human for the judgement, the relationship, the complaint, the weird case. Three jobs an AI agent does well today for Bristol SMEs 1. Twenty-four seven call handling for trades, salons and restaurants Missed calls are the most expensive admin problem a Bristol service business has, and the one with the cleanest AI workflow. The agent picks up every call, takes a structured brief, promises a callback window, and drops the lead into the CRM before the caller puts the phone down. The trade owner comes off the job, sees four new leads in the dashboard, calls them back in order. The numbers vary, but the shape is consistent. A Bedminster plumber misses roughly ten to fifteen calls a week on jobs. A Clifton salon misses eight to twelve on a Saturday. A Stokes Croft electrician misses twenty during a long install. Each missed call is a job that went somewhere else. An AI voice agent catches them. The cost is under fifty pounds a month for the basic setup. A temp receptionist costs three hundred pounds a day. The voice agent is honest about being an AI. The opening line tells the caller. Ninety-five percent of callers say "yes, fine". The five percent who want a human get transferred or taken a message. Trust is built by being clear, not by pretending. 2. Quote and follow-up automation for service businesses Quote follow-up is the second-most-expensive admin problem for a Bristol service business, and the one most teams simply do not do well. A quote goes out on Monday. The customer goes quiet by Wednesday. By Friday, the team has forgotten. By the following Monday, the lead is cold. An AI agent drafts the quote from the brief, sends it, drafts the first follow-up for day three, the second for day seven, the third for day fourteen. The human reviews and presses send. The agent watches for the reply and flags to the human the moment a lead goes quiet or comes back hot. The human still signs off the price. The AI does the typing, the timing, and the watching. The same pattern works for invoices. The AI drafts the chase email, watches for payment, flags the customer who has gone quiet for thirty days. The credit controller only sees the cases that need a phone call. 3. Internal admin tools for the Bristol team The third job is the one most Bristol SMEs do not realise AI is good at. The team spends four to six hours a week compiling reports, briefings, and digests that already exist in the data. The Monday-morning sales report. The weekly enquiry digest. The Friday cashflow summary. The monthly marketing dashboard. An AI agent takes the data from the CRM, the booking system, the bank feed, and the inbox, and compiles the report. It arrives in the team inbox at 7am Monday. The team reads it over coffee. The cost-saving is not "AI replaces the team". It is "AI removes the four hours a week the team spends compiling the report". That is four hours back into the business every week, forever. The same pattern works for client onboarding in a Bristol professional practice. The AI takes the new client brief, drafts the welcome pack, populates the CRM, schedules the kickoff call, and sends the calendar invite. The partner signs off the welcome pack. The AI handles the rest. What an AI agent isn't Honest section. An AI agent is not a person. It is not a replacement for a human voice on a tricky call, a complaint, or a long-term client relationship. It is not magic. It does not learn by itself in production without supervision. It does not "understand" your business the way you do. An AI agent is a very good worker for a specific, bounded, well-briefed job. The minute the job stops being well-briefed, the agent hands back to a human. That is the design. That is the safety net. That is what makes the system trustworthy for a Bristol business whose reputation is the product. If anyone tries to sell a Bristol service business an "AI that runs your business", walk away. That is not the technology. That is marketing. If you're a Bristol service business wondering where to start The honest answer is: pick the job where you are losing time or money every week, not the most exciting AI capability. For most Bristol service businesses the answer is missed calls. For some it is quote follow-up. For others it is the Monday-morning report grind. The cheapest way to find out is a twenty-minute diagnostic with a Bristol AI consultancy that has actually shipped these systems. We will tell you, in plain English, whether the workflow is ready for an AI agent, whether the underlying system needs fixing first, or whether the workflow is not yet worth automating. All three answers are useful. For the broader picture of how a Bristol AI consultancy engages with a service business, the parent page walks through the diagnostic-to-shipped-workflow path. If you are a Bristol trades firm whose first AI win is more likely to be local search than automation, the local SEO entry point is the right starting point; the AI agent is the step after that. Twenty minutes. A short diagnostic. If you are a Bristol service business wondering where an AI agent would actually pay back, the cheapest answer is a twenty-minute call. Get in touch via the contact form Looking for the wider picture? I have put all five of the new AI agents side by side, what each is actually for, and which ones you can get in the UK, on AI employees and AI agents .]]></content:encoded>
  </item>
  <item>
    <title>Local SEO Audit · £149 + Managed &amp; Video</title>
    <link>https://expedientai.co.uk/local-seo/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/local-seo/</guid>
    <pubDate>Mon, 20 Jul 2026 00:00:00 +0000</pubDate>
    <description>Local SEO for UK trades. £149 audit, £497 full Map Pack setup, £150/mo managed, plus professional and AI video for your GBP and YouTube. Bristol, UK.</description>
    <content:encoded><![CDATA[Local SEO audit for UK trades Find out why your competitors are in the Map Pack and you aren't. £149. One payment. A 1-page report in 5 working days. We check your Google Business Profile, your Map Pack position for the searches your customers actually type, your known citations, and the local signals on your website. Plain English, no jargon, no 60-page PDF. Book the £149 audit What's in the report One-off, no subscription Delivered in 5 working days Yours to keep, act on, or ignore Why it matters The Map Pack is where the work is. When someone in your town searches "emergency plumber near me" or "electrician BS1" at 9pm on a Tuesday, your business should be one of the three names they see. The Map Pack takes 44% of the clicks on local search results. The businesses in it get the call. The ones below it don't. Google Search · Local 3-pack Three businesses. Top of the page. Below them, everyone else fights for what's left. 76% of local mobile searches result in a phone call within 24 hours. If your business isn't one of the three in the Map Pack, the customers searching in your area right now are calling someone else. The fix isn't complicated. It's just boring, and it doesn't fit on a Saturday between jobs. The £149 audit tells you exactly which of the six things you need to fix, in what order, and how much each one matters. No 60-page report, no strategy call, no upsell. One page, plain English, delivered in 5 working days. What's in the report Six things. Ranked by impact. Every audit checks the same six things. Every report has the same structure. You know what you're getting before you book. 01 Google Business Profile score. Out of 100. Based on whether your profile is claimed, verified, complete, has photos, has hours, has services listed, has Q&A seeded, and has a description. Most trades businesses score 40-60. Top 3 in the Map Pack usually score 80+. 02 Map Pack position, by keyword. We track 3-5 keywords your customers actually type ("emergency plumber Bristol", "boiler repair BS1"). We tell you whether you're in the Pack, where, and who's in front of you. Screenshot included. 03 Review count, velocity, response rate. Total reviews, new this quarter, average rating, how many of yours the business owner responded to. Counter-intuitive fact: review velocity matters more than total count. Steady drip wins. 04 Citation health. You give us a list of citations you care about (your website, Yell, Checkatrade, anywhere else you're listed). We check each for Name/Address/Phone consistency. One mismatch is enough to confuse Google. 05 Website local signals. Is your NAP in the footer? Is there a LocalBusiness schema markup? Is there a sitemap? Is there a Google Maps embed on the contact page? These are the things that signal "we're a real local business" to Google. 06 Competitor comparison. The 3 businesses currently in the Map Pack for your main keyword. Their GBP scores, their review counts, their photo counts, their citation counts. Honest context for where you stand. What if your profile isn't set up yet? No GBP, or never claimed, or unverified? We handle it. Most trades businesses fall into one of three camps: never created a Google Business Profile, created one but never claimed it (Google auto-generates listings for many businesses), or claimed it but never verified (so edits don't go live). The audit covers all three. Included in the £149. No profile yet We create it. You give us your business name, address, phone, hours, services, and a few photos. We create the profile, fill in every field Google cares about, post the photos, and seed the Q&A. Verification arrives by postcard, phone, or email depending on what Google chooses. Unclaimed We claim it. Google auto-creates GBP listings for most businesses based on public data. Most trades business owners don't realise they have one. We claim it, take over ownership, and clean up whatever's been auto-populated (often wrong hours, no services, no photos). Claimed but broken We fix it. Missing hours. Old address. No services listed. Three photos from 2021. Profile description says nothing. We've seen all of it. The audit includes the cleanup: hours, services, photos, description, attributes, categories. Done as part of the report, not as a separate job. What the report looks like One page. Plain English. Yours to keep. PDF, 1 page, scored checklist at the top, prioritised fix list at the bottom. No fluff, no agency-speak, no "synergise your omnichannel presence." Just what to fix, in what order. Local SEO Audit · Alpha Rod 5 working days · prepared by Expedient AI · 12 July 2026 SCORES GBP completeness 42 / 100 Map Pack (1 of 5) Not in pack Reviews 23 · 2 this qtr TOP 3 QUICK WINS 1. Add 8 photos of recent jobs to GBP. Top 3 in the Map Pack have 40+. You have 6. 2. Update hours to "Open 24 hours" on GBP (your website already says this). You lose out-of-hours searches. 3. Your Yell listing has your old BS3 address. Update to BS2. One mismatch costs Pack position. CITATION DRIFT Yell.com: phone 01249 448345 ✓ · address ✗ (old unit) Checkatrade: phone ✓ · address ✓ · no hours Facebook: phone ✓ · address ✗ (postcode missing) COMPETITORS (Map Pack, "emergency plumber bristol") Frank Plumbing: 4.5★, 296 reviews, 24h, GBP score 87 EMERGENCY PLUMBER BRISTOL: 4.9★, 565 reviews, 24h, GBP score 91 Smart Plumbing: 4.8★, 199 reviews, 24h, GBP score 84 Illustrative example · 1 page · PDF What you actually get. A score for each of the six checks, with the GBP score and Map Pack position at the top because those are the two that matter most. The top 3 quick wins below, ranked by impact, specific to your business (not generic SEO advice). Citation drift table if any of your listings are inconsistent. Competitor comparison so you know what you're up against. The 1-page constraint is deliberate. If the report needs 10 pages, you won't read it. If it fits on one, you will. Actionable beats comprehensive. Yours to keep. No contract, no upsell to a "managed service", no "strategy call" required. Use the report yourself, hand it to a web agency, or never act on it. Your £149 either tells you something useful or it doesn't, and the report stands on its own either way. How it works Five days. Three emails. One report. 01 You book and tell us your details. Business name, town, your website, the 3-5 keywords your customers actually type, the citation URLs you care about, where to send the report. One form, 5 minutes. 02 We run the audit. We check your Google Business Profile, your Map Pack position for each keyword, your known citations, your website's local signals, and the top 3 competitors in the Pack for your main keyword. Automated, takes 4-6 hours spread across days 1-4. 03 The report lands in your inbox. Day 5. PDF, 1 page. Scored checklist at the top, prioritised fix list at the bottom. Plus the raw screenshots and data we used, in case you want to verify anything yourself. 04 You act on it (or don't). The top 3 quick wins are usually 2-3 hours of work. Hand them to your web agency, do them yourself on a Saturday, hire someone on Fiverr. The report tells you what to do, not what we do next. The honest bit Who this is for. Who it isn't. For you if: You run a local UK trades business (plumber, electrician, roofer, gas engineer, locksmith, etc) within a 20-mile radius. You know the Map Pack matters but you don't know exactly where you stand or what to fix first. You want a clear, honest read on your local SEO without signing a retainer or learning another tool. Also for you if you have no Google Business Profile yet, or never claimed or verified the one Google made for you, because we set that up as part of the audit. Not for you if: You don't have a website yet. Start with Website Build & Hosting from £300 setup , come back to Local SEO once the site is live. Or you serve the whole UK from one office and Map Pack ranking isn't the right channel for you. Need more than the audit? Three options above the £149 read. Some businesses want the report and the work done for them. Some want a monthly presence so the Pack position holds. Some want video on their GBP and YouTube because that's the new search surface. Pick what fits. Mix and match. Cancel monthly after one month, no contract. £497 ONE-OFF Full Map Pack setup. Everything in the £149 audit, plus we do the fixes. Google Business Profile claimed, verified, filled out, with photos, services, hours, Q&A. Citations built across the 12 directories that actually move the needle (Yell, Checkatrade, Yelp, Thomson Local, FreeIndex, Hotfrog, Cylex, Scoot, TouchLocal, Tipped, UK Small Business Directory, 192.com). NAP consistency audit and correction on your existing listings. LocalBusiness schema added to your website. Maps embed on your contact page. One-time project, two weeks from intake. Best for: businesses that know what to do but don't have the time or the patience to do it themselves between jobs. Book the £497 setup £150/MO MANAGED Hold the Pack position. Once you're in the Pack, you stay there by being active. We run your GBP every month: 2-4 posts, monthly photo refresh, Q&A seed, review response (you send us the new ones, we reply). Citation drift monitoring across the 12 directories. Monthly rank report for your 3-5 keywords with Pack position, top 3 competitors, and a 1-line action note. Monthly fee, cancel after one month, no setup fee if you've had the £149 audit or £497 setup with us. Best for: businesses that have got into the Pack and want to stay there without doing the monthly work themselves. Start the £150/mo plan FROM £397 GBP video + YouTube channel. The new search surface for trades is video. AI Overviews cite YouTube in 30% of local-intent answers, and a GBP video gets 5x the calls of a GBP without one. Two routes. Filmed video: one half-day with a local crew in your area, 3 short-form videos (60-90 seconds each, on-site at a real job or in your workshop), edited and uploaded to your GBP and YouTube channel with titles, descriptions, and tags tuned for local AI search. £897 one-off. AI video: no filming. We script, storyboard, and produce 3 short-form videos from your existing photos and assets using AI tooling. £397 one-off. Both routes include a 30-min YouTube channel setup call so your channel exists, has the right categories, and starts with a real upload. Best for: businesses that know their next customers are finding people on YouTube and ChatGPT, not just Maps. Book a video package Side by side What you get, what you don't. Four SKUs. The audit is the entry. The other three are the work above it. Pick the one that matches what you want to hand over. Anything outside the four is a custom engagement and goes on services.html, not this page. £149 AUDIT Read. GBP score (out of 100) Map Pack position, 3-5 keywords Review count + velocity Citation NAP check (your URLs) Website local signals Top 3 competitor comparison Top 3 quick wins 1-page PDF, 5 working days £497 SETUP Build. Everything in the audit GBP claimed + verified GBP fully filled + photos Citations built (12 dirs) NAP drift fix (existing) LocalBusiness schema added Maps embed on contact page 2 weeks from intake £150/MO MANAGED Hold. 2-4 GBP posts / month Monthly photo refresh Q&A seeding Review responses (you send, we reply) Citation drift monitoring Monthly rank report (3-5 kws) 1-line action note Cancel any month after first FROM £397 VIDEO Show. 3 short-form videos (60-90s) Local AI search titles + tags Uploaded to GBP + YouTube YouTube channel setup call (30 min) Filmed: £897 with local crew AI: £397, no filming GBP video = 5x calls Yours to publish or reshoot Most customers start with the £149 audit, then add either the £497 setup or the £150/mo plan once they've seen the report. The video option is independent and works either before or after the audit. Stack any combination. Video The new search surface is video. AI Overviews cite YouTube in nearly a third of local-intent answers, and a Google Business Profile with a video gets five times the calls of one without. If you have no video, your business is invisible at the second step of the journey. AI video · £397 Three videos, no filming. Produced from your existing photos and copy. 60-90 seconds each, captioned, with titles and tags tuned for local search. Uploaded to your GBP and YouTube channel. Five days from intake. Filmed video · £897 Half a day, three videos. A local crew on a real job or in your workshop. Edited and captioned, uploaded to both, with titles tuned for the terms your customers type. You keep the raw footage. Video SEO The full breakdown. How video helps you twice, what the channel setup call covers, and what video will not fix for you. The honest version, including when it is not worth doing. See video SEO Video covers what your customers watch. If the problem is that AI assistants and answer engines cannot read your website at all, that is a different discipline: Agentic Engine Optimization . From £149 audit to a fully managed Map Pack presence with video. Read the page. Pick the SKU that fits. Or book the £149 audit first if you'd rather see where you stand before committing to anything. We do the work; you get the calls. See the four SKUs]]></content:encoded>
  </item>
  <item>
    <title>Brand Voice Setup · Your tone, baked into AI</title>
    <link>https://expedientai.co.uk/brand-voice-setup/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/brand-voice-setup/</guid>
    <pubDate>Mon, 20 Jul 2026 00:00:00 +0000</pubDate>
    <description>A one-off engagement that turns how your business actually sounds into a reusable voice document. Built into every AI workflow you run afterwards. £350 one-off. From Bristol, UK.</description>
    <content:encoded><![CDATA[Brand Voice Setup · One-off service Your tone, baked into AI. A 30-minute intake. A one-page voice document. A reference catalogue your team and every AI workflow you run can use for the next 12 months. £350 one-off, no subscription. Book the intake See what's included One-off, fixed price Delivered in 5 working days UK small businesses What you get Three things, delivered together. 01 The voice document. One page. Plain English. Who you are, who you aren't, the rules your writers and your AI workflows both follow. Sentences your brand uses. Sentences your brand doesn't. Read aloud at the next team meeting and it lands. 02 The reference catalogue. Twelve to twenty real examples from your own website, emails, sales calls. Each one tagged with what makes it work. Use it to brief new writers, brief new AI agents, settle internal debates about tone. 03 The prompt-ready version. The same voice document, formatted as a system prompt and as a Vereby-compatible profile. Drop it into ChatGPT, Claude, your own assistant, or any future workflow you build. Stays valid for 12 months, renewable at £150. How it works Three steps. Five working days. 01 Intake call. 30 minutes. A short call to understand the business, the audience, and the situations where voice matters most (sales emails, customer support, social, internal comms). I ask the questions; you talk. No prep needed. 02 Draft. Three to four working days. I pull everything you have publicly (site, posts, emails you've signed) plus the intake notes. Draft the voice document and the reference catalogue. Show you the work with annotations explaining every choice. 03 One round of revisions, then hand over. You read, you react, you tell me what to tighten. One round of revisions is included. Once signed off, you get the three deliverables in Google Doc, PDF, and prompt-ready formats. Yours to use, edit, or hand to a writer. What it's for Use the voice document for the next twelve months. When you're the source Brief a freelance writer and get copy that actually sounds like you Train a new hire on what "on-brand" means without a three-hour meeting Settle a disagreement about whether something "sounds like us" Hand the same voice to a marketing agency you hire for a one-off project When AI is the source Set the system prompt for a ChatGPT or Claude workspace the team uses Wire the voice into a Vereby assistant or a bespoke workflow you build Stop rewriting AI drafts from scratch because they don't sound like you Make sure every AI-generated email, social post, and support reply uses your tone, not a generic one The honest bit What this isn't. Not a rebrand. A rebrand changes logos, colours, positioning, and market. This doesn't touch any of that. It captures what you already are and makes it explicit. Not a tone-of-voice guide from a 60-page PDF. Most tone guides sit in a Google Drive nobody opens. The voice document is one page, plain English, written for the people who'll actually use it. Not a subscription. £350 once. The voice document stays valid for 12 months. If your business shifts materially before then, I'll redraft at half price. Renewal after 12 months is £150 if you want it. Not a substitute for a writer. It tells writers and AI what your voice is. It doesn't write the next campaign. If you need a writer for a specific project, that's a separate engagement. Pricing £350, one-off. 30-minute intake, three to four days drafting, one round of revisions. Three deliverables. Yours to keep, edit, and use however you like. Book the intake Ask a question first Expedient AI Ltd · Registered in England and Wales · Bristol, UK Already a Vereby tenant? The Brand Voice Setup is bundled with every Vereby tenant. You already have the deliverable. If you want it as a standalone document you can hand to writers and agencies, the £350 covers the formal version. See Vereby]]></content:encoded>
  </item>
  <item>
    <title>Twelve models in seven days: a working reading of the July 2026 AI wave</title>
    <link>https://expedientai.co.uk/feed/july-2026-ai-model-wave/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/july-2026-ai-model-wave/</guid>
    <pubDate>Tue, 14 Jul 2026 00:00:00 +0000</pubDate>
    <description>Twelve frontier AI models shipped in one week: GPT-5.6, a new Claude, Gemini, DeepSeek V4 and more. What actually changed for a UK service business.</description>
    <content:encoded><![CDATA[News Twelve models in seven days: a working reading of the July 2026 AI wave Twelve frontier AI models shipped in one week: GPT-5.6, a new Claude, Gemini, DeepSeek V4 and more. What actually changed for a UK service business. 2026-07-14 · 8 min read The week from 2 July to 9 July 2026 was the densest release week we have tracked. Twelve model announcements landed, several of them frontier-class. This is a working reading, not a press digest. The point is to separate the changes that matter to a UK service business running AI day-to-day from the ones that matter only to people training the next frontier. What actually shipped OpenAI launched GPT-5.6 on 9 July (TechCrunch coverage), in three variants: Sol, its workhorse, Terra, the intermediate option, and Luna, the budget tier. Same shape as GPT-5 in March: three price points, three capability tiers. For a UK small business this matters less for the headline and more for the pricing tiering. The Luna tier is what most service businesses actually pay for, because most of the volume work (drafting, summarising, classification, triage) does not need the frontier. The interesting question is whether Luna gets you the same reliability improvements that shipped in Sol, or whether the reliability work is locked behind the top tier. That detail is not in the public coverage yet. Anthropic released a new Claude coding model at the Code with Claude event, and the Straits Times coverage called it "better at coding" than the previous generation. Anthropic did not give it a numbered name in the headline coverage, instead positioning it as the next step in the Claude coding line. For a service business that has been using Claude for code-adjacent work (writing technical specs, reviewing vendor deliverables, drafting API contracts), the upgrade is more about context handling and long-output coherence than raw capability. Worth a pilot on the same prompt set you already use, before assuming it is better. Google refreshed the Gemini family. Google has not announced a numbered flagship in the same way as the others this week, but the Gemini refresh landed alongside Google's broader AI rollout updates. For a small business already paying for Workspace, this is the version that arrives automatically. It is not a model you switch to. It is a model that switches to you. The interesting question for any operator is whether the new Gemini improves on tool-use reliability for your customer support workflows, because that is where most Gemini upgrades either pay off or quietly don't. DeepSeek previewed V4. Reuters and TechCrunch both covered the V4 preview in late April. The way to read V4 is not "another Chinese model." It is "DeepSeek is closing the gap with the frontier and pricing it like open source." For a service business running on the cheap-model layer of the architecture, V4 changes what that layer can do. If you currently route drafting and triage to a small open-weight model, V4 means you can do more on the cheap layer before escalating to a frontier model, which lowers your per-task cost and increases your reliability. Mistral released its first robotics model on 8 July (Reuters). Not directly relevant to a UK service business unless you build physical products or industrial automation. Worth knowing it exists, not worth a switch. xAI released Grok 4.5 alongside the broader GPT-5.6 rollout. xAI's positioning has been developer-led, and Grok 4.5 continues that. For a service business, Grok has not been a default recommendation and the latest release does not change that. If you are using xAI for a specific reason (real-time data, X integration), test it; otherwise ignore it. Google, Mistral, Meta and several open-weight labs also shipped smaller updates and open-weight variants through the week. These do not dominate the headlines but matter if you are running local models on a Mac mini, an NVIDIA Spark, or a VPS. When open-weight drops, the Mercury OS install shape changes with it, because the cheap-model layer is upgradeable in place. What actually changed for a UK service business Three shifts. None of them are about the raw capability of the model. They are about the operating economics. 1. The cheap layer got cheaper and more capable. DeepSeek V4 preview, Google's open-weight refreshes, plus the continuing trend of open-source and open-weight labs shipping capable models at near-zero inference cost, mean that the tier of work your business does most of (drafting, summarising, classification, triage, basic extraction) is now good enough that you do not need to escalate it to a frontier model. The cost difference matters. A UK service business running on Hermes orchestration can route ninety percent of volume to the cheap layer and only spend frontier-model money on the work that actually needs it. That cost discipline gets cheaper every month. 2. The frontier layer got specialised. GPT-5.6 with three variants, Anthropic's coding-specialised model, Google's Gemini refresh tuned for tool use. The frontier is no longer "one big model that does everything slightly better." It is "three to five specialised models that each do one thing very well." For a service business, that means choosing your frontier model per workflow, not per provider. The Mercury OS install shape already routes by task. The news this week is that the routing decision has more options to choose from. 3. The capability bar moved for everything. When the cheap layer gets better, the work you used to escalate moves down. When the frontier layer gets better at one thing, that one thing becomes a default rather than an exception. Net effect: more of your business can run on AI at lower cost, but the bar for "good enough" has also moved up, because customers have seen the new models. What we are changing in the Mercury OS install shape For Mercury OS customers, nothing changes this week. The four modules (Hermes Agent, Vereby, S0cial Master, the Zoho Lead Engine) keep working on the existing cheap-model layer. When new open-weight variants are worth swapping in, that happens as a config change, not a reinstall. For prospective Mercury OS customers, the conversation changes. The cost pitch ("you don't need to pay frontier money for ninety percent of your volume") is now even stronger than it was a month ago, because the cheap layer genuinely is better than it was. The frontier pitch ("frontier models for the ten percent that genuinely needs them") is now more nuanced, because there are three to five specialist frontends to choose between rather than one big model. For Ted's own work, this week is the kind of week where it pays to run the same prompt set against the new models and check which one moved the needle on which task. That is a 2-4 hour exercise per model and it usually surfaces one or two specific upgrades worth making. Expect small changes to the architecture across August, not a big rewrite. The honest version of how to think about a week like this If you are not in the AI business and you are a UK service business owner, the takeaway is short. The AI you are paying for (whether that is ChatGPT, Claude, Gemini, or whatever sits behind the tool you use) just got better. If your vendor is on top of their game, you should already see the improvement in the next two to four weeks. If you do not, that is a signal to ask them. If you are running AI yourself (whether through Mercury OS or another stack), do the boring thing. Run your ten most common prompts against the new models that matter to your work. Pick the one that does your job better. Move on. The week was not about a single breakthrough. It was about the cheap layer getting cheaper and the frontier layer getting specialised. That is good news for the economics of running AI in a small business, and not much else. Frequently asked questions Should a UK service business switch models every time a new one launches? No. The bottleneck for most UK small businesses is not which model they use. It is the prompts, the data, the integrations, and the operator habit. Switching models for the sake of it costs you more than it gains. Pick a stack that works, test it every three to six months, switch the one piece that genuinely moves the needle, and leave the rest. Is open-weight or open-source actually good enough to run a service business? For most of the work a service business asks an AI to do (drafting replies, summarising documents, classifying enquiries, drafting social content, drafting proposals) yes. The cheap layer of the stack, open-weight and open-source models, is genuinely good enough for ninety percent of volume. You save the frontier models for the ten percent that needs the highest quality: the proposal you send to a £100k client, the security review, the response that has to be right the first time. What is the cheapest reasonable AI stack for a small UK business right now? A small open-weight or open-source model on the cheap layer (drafting, summarising, classification, basic extraction), a frontier model only on the work that genuinely needs it (final-draft proposals, security review, contract language), and a Hermes-style orchestration layer that picks which is which. Total run cost for a one-person service business is roughly ten to forty pounds a month, depending on volume and on how much of the work stays on the cheap layer. How often should a small business review its AI stack? Quarterly is the right cadence for most operators. The cheap layer changes every few weeks, the frontier layer changes every few months, and the orchestration decision (which model for which task) is the one that pays back the most attention. A two-hour quarterly review, run the same ten prompts against the new options, change one thing, leave the rest.]]></content:encoded>
  </item>
  <item>
    <title>Zoho AI lead generation for UK small businesses: the system we built, end to end</title>
    <link>https://expedientai.co.uk/feed/zoho-ai-lead-generation-uk/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/zoho-ai-lead-generation-uk/</guid>
    <pubDate>Sun, 12 Jul 2026 00:00:00 +0000</pubDate>
    <description>How to wire AI lead generation into your existing Zoho CRM. A bespoke product that searches LinkedIn, scores prospects against your fit criteria, and drafts LinkedIn messages in your voice.</description>
    <content:encoded><![CDATA[Tips Zoho AI lead generation for UK small businesses: the system we built, end to end How to wire AI lead generation into your existing Zoho CRM. A bespoke product that searches LinkedIn, scores prospects against your fit criteria, and drafts LinkedIn messages in your voice. 2026-07-12 · 9 min read Most UK small businesses that run on Zoho are losing prospects the same way: not because the leads are bad, but because nobody has the time to chase them. The owner is on a job site, the office manager is buried in admin, the sales rep doesn't exist. The leads sit in the inbox, get stale, and quietly die. The honest answer is that most UK small businesses do not need more leads. They need the leads they already have worked harder. Zoho AI lead generation is the system that does that, without hiring anyone. What Zoho AI lead generation actually does Three things, every Monday, Wednesday and Friday at 7am UK time. The system runs itself. You read the digest. Finds the right people on LinkedIn. You give us a target persona (job titles, industries, regions, exclusions) and the system searches LinkedIn via Bright Data every digest run. People who match the persona and whose recent activity suggests they're a real fit get added to your list. The search is targeted, not spammy, the right people, not a thousand random connections. Ranks them against your fit criteria. Each new prospect is scored against the criteria you set on day one. Top-of-list is highest fit. Low-fit gets dropped before they ever reach your inbox. You only see the leads worth your time. The score is logged on the prospect in Zoho so your team can see why each one was prioritised. Writes the LinkedIn message in your voice. The system drafts a short, specific LinkedIn message for each scored prospect in your brand voice (we scrape your existing public content to learn it on day one). Drafts wait in a queue for you to read, edit, send. The message reads like you, not like a bot. How it works inside your existing Zoho The architecture sits behind your existing Zoho CRM. Nothing changes on your side. The team logs into the same Zoho they always have. The AI writes into the workflows you already use. Zoho CRM is the system of record. Every prospect, every lead, every status change lives here. We never write to anything you haven't approved. The AI sits behind the scenes, your team logs into the same Zoho they always have. Bright Data is the LinkedIn search engine. Residential proxies, no monthly cap. Replaced Apify in 2026. Roughly three pence per weekly Cornwall run. Pay-as-you-go. The data is the data; the AI is the AI. OpenRouter is the LLM gateway. One key, many models. Cheap models for routine scoring and drafting. Frontier models for high-stakes replies. The router picks automatically based on the task. Supabase is the read-cache for the admin dashboard. Real-time. Zoho is the source of truth. Supabase is the speed layer. The admin dashboard reads from Supabase for instant filtering and search; the source data lives in Zoho. Next.js 15 is the admin dashboard. Runs on Railway. See the digest, review drafts, edit brand voice, change filters. Your team logs in, no training needed. Hermes Agent is the orchestrator. Plans, delegates, runs the right model for the right task. Open source. Cost guardrails built in. Falls back to cheaper models automatically when monthly budget hits 80%. The full architecture is documented in the Zoho Lead Engine product page and the original build case study. How long it takes to set up Fourteen days from kickoff to first digest. We have written about the setup process in detail on the approach page. The short version: Days 1-2, Discovery. A 30-minute Zoom. We learn your persona, region, what "warm" means to you, what a good lead looks like in your pipeline, and what a bad lead looks like. You share your Zoho login (read-only first, write access when you're ready). You walk us through an existing prospect you liked. Days 3-7, Setup. We configure your tenant's filter criteria, brand voice profile, and Zoho schema mapping. Zoho becomes the system of record for every prospect; our admin dashboard reads from a real-time cache so you can review and edit drafts without going into Zoho. Setup runs in 7 to 10 days. Days 8-14, First digest. The first Monday, Wednesday or Friday after setup completes, the system runs end-to-end: search, score, draft, queue. You log in to your admin dashboard, read the AI's picks, edit the drafts you want to personalise, send the ones you're happy with. The rest sit in the queue for review. What it costs The CRM cost is unchanged, you keep paying Zoho. The AI layer costs roughly fifteen to twenty-five pounds a month per tenant in API spend, plus a small Next.js admin dashboard hosted on Railway Hobby at five pounds a month. Total: under thirty pounds a month for the AI layer, on top of whatever Zoho already charges you. We don't publish a price for the build itself because the cost depends on the complexity of the persona and the integrations. Most builds run in the low four figures. Talk to us about a 30-min scoping call to get a real number for your business. Who this is right for Zoho AI lead generation is right for a UK small business that: Already runs on Zoho (or another CRM that we can wire into the same architecture). Sells a service that takes time to research and qualify (B2B, professional services, recruitment, etc). Has a clear target persona and a region. Has the discipline to read the weekly digest and send the drafts. Doesn't need to be in 50 LinkedIn conversations at once. It is not right for a business that wants volume over quality, or that hasn't yet defined what a good lead looks like. The system amplifies what the team already does. It does not invent a sales motion from nothing. What this means for you If you run a UK small business on Zoho and the above describes what you need, the system is right for you. The AI layer costs under thirty pounds a month on top of Zoho. The setup takes two weeks. The hard part is having the discipline to read the weekly digest and send the drafts. If you cannot do it yourself, that's fine. We build these systems for a living, see the Zoho Lead Engine product page for the full architecture, or the work page for what's already shipping. The forty-quid plumbing firm post is a working example of the right-sized stack. You can read about how I build these systems in the Approach section, or see real builds in my Work. Frequently asked questions Can I add AI lead generation to my existing Zoho CRM? Yes. Zoho is the system of record - the AI writes prospects, scores them, and stages drafts inside the workflows you already use. You do not need to migrate. You do not need a parallel CRM. The AI sits behind the scenes, your team logs into the same Zoho they always have. What does Zoho AI lead generation actually do? Three things, every Monday, Wednesday and Friday at 7am UK time. The system searches LinkedIn for prospects matching your target persona and region, scores them against your fit criteria, and drafts a LinkedIn message in your voice. The drafts wait in a queue for you to read, edit, send. How long does it take to set up Zoho AI lead generation? Fourteen days from kickoff to first digest. One day to learn your persona and what a good lead looks like in your pipeline. Three to five days to configure the filters, brand voice, and Zoho schema mapping. The remaining days run the system in shadow mode while you compare its output to your existing manual workflow. What is the cost of Zoho AI lead generation? The CRM cost is unchanged - you keep paying Zoho. The AI layer costs roughly fifteen to twenty-five pounds a month per tenant in API spend, plus a small Next.js admin dashboard hosted on Railway Hobby at five pounds a month. Total: under thirty pounds a month for the AI layer, and the rest is whatever Zoho already charges you.]]></content:encoded>
  </item>
  <item>
    <title>Cheap AI for UK SMEs: the right-sized stack, with real monthly costs</title>
    <link>https://expedientai.co.uk/feed/cheap-ai-for-uk-smes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/cheap-ai-for-uk-smes/</guid>
    <pubDate>Sun, 12 Jul 2026 00:00:00 +0000</pubDate>
    <description>Open-weight and open-source models now do the bulk of SME admin work for under forty pounds a month. We break down the cost: which models, what each one does, and what the right-sized stack looks like.</description>
    <content:encoded><![CDATA[Tips Cheap AI for UK SMEs: the right-sized stack, with real monthly costs Open-weight and open-source models now do the bulk of SME admin work for under forty pounds a month. We break down the cost: which models, what each one does, and what the right-sized stack looks like. 2026-07-12 · 7 min read UK small businesses have been sold a particular story about AI. The story says: the best models are the expensive models, and the expensive models are the only ones worth using. The story is half true. The best models for general-purpose reasoning are the frontier ones: Claude, GPT, and the next generation of closed-weight systems. If you are building a model that has to pass a bar exam, or write a legal brief, or summarise a long medical record, you need a frontier model. Most small UK businesses are not doing any of those things. Most small UK businesses are drafting replies to quote requests, booking jobs into a calendar, chasing late payments, and posting on social media twice a week. For those tasks, the frontier models are overkill. The open-weight models are good enough. And the cost difference is enormous. What cheap AI for UK SMEs looks like in 2026 Three monthly services, each under fifteen pounds, that a tradesperson can run from their phone. Zoho CRM holds every job, every customer, every quote. The plumber types into it the way he always has, on his phone in the van. Nothing changed on his side. OpenRouter routes the AI calls. Most of the work goes to open-weight models, MiniMax, DeepSeek, GLM, at roughly a tenth of the cost of a frontier model. When something matters (a complaint, a pricing dispute, a final review before send) the call gets bumped to Claude or GPT. The router picks automatically. A small Next.js app runs on Railway Hobby, five pounds a month. The only screen the plumber ever opens. Three things on it at any time: unpaid jobs, waiting enquiries, quotes needing a follow-up. No dashboard sprawl. Total monthly cost: thirty-eight pounds. Less than most tradespeople spend on tools, less than one missed appointment. We have written about this stack in detail on the forty-quid plumbing firm post. The architecture is the same one we use in Vereby, our multi-tenant platform for SMEs. Why open-weight models are good enough Open-source and open-weight models have come far enough that, for ninety percent of the work, the output is indistinguishable from the frontier ones. They don't hallucinate as much. They follow instructions better. They handle UK English without a fight. If you tried them eighteen months ago and were unimpressed, try them now. The reason is partly the model architecture and partly the routing. A well-built router sends routine work to the cheap model and the hard work to the expensive one. The cheap model handles the first draft. The expensive model handles the final review. The system is faster, cheaper, and better-tested than a single-model setup. The price difference between models is small in absolute terms, four pence versus six pence per reply, about forty pence a day, one hundred and fifty pounds a year. Not the money that decides whether to add AI. The money that decides it is the time you stop losing. Where cheap AI is not the right tool Three limits worth being honest about. Money, contracts, complaints, keep a human in the loop. Open-weight models get tone and facts right most of the time, but the ten percent they miss is the kind of error that loses a customer. Up-to-the-minute UK-specific information still needs a search step on top of the model, regulation changes, supplier pricing, local council rules. The model will guess confidently if you let it. Sensitive personal data needs a careful audit, medical notes, financial advice, safeguarding concerns. The models aren't unsafe, but they aren't pre-tuned for these domains the way dedicated tools are. In every case where the answer matters, the model is the draftsperson, not the publisher. The plumber reads every reply before it goes out. The salon owner reviews every quote. The cafe manager checks every health-and-safety note. The model writes. The human decides. How long cheap AI takes to set up Two to four weeks for a trades firm with a CRM already in place. A week to map the actual workflows (not the ones in the brief, the ones the business actually runs), a week to wire the data, a week to train the model on the business's voice and tone, and a week running the system alongside the old way before switching over. Faster for a service business with a CRM already in place. Slower for a brand-new business with no data history. The full breakdown is on our approach page. The bottom line on cheap AI for UK SMEs For most admin tasks, open-weight models handle the work for under forty pounds a month. The router, not the model, is what makes the system cheap. The trade-off is not quality. It is ambition, the system will not write the next great British novel, but it will draft your reply to the customer who asked about a leaking radiator before your kettle boils. You can read about how I build these systems in the Approach section, or see real builds in my Work (Vereby is the multi-tenant platform that powers setups like this one). Frequently asked questions What is the cheapest AI a small UK business can run today? For most admin tasks - drafting replies, summarising documents, scheduling - open-weight models like MiniMax, DeepSeek and GLM handle ninety percent of the work for under forty pounds a month. Frontier models are reserved for the ten percent where the cost of being wrong is high. Is cheap AI safe for customer-facing replies? Open-weight models are safe for the eighty percent of replies that follow your existing templates and tone. The twenty percent that involve money, complaints, or commitments need a human in the loop or a frontier model. The right-sized stack buys you capacity, the human stays in charge of judgement. Why is AI still expensive for some UK small businesses? Most UK small businesses are paying premium prices because they are using frontier models for routine work. MiniMax, DeepSeek, and GLM now match the quality of GPT-4 for narrow tasks at a tenth of the cost. The router, not the model, is what makes the system cheap. How long does cheap AI take to set up? Two to four weeks for a trades firm with a CRM already in place. One week to map the actual workflows, one week to wire the data, one week to train the model's voice, and a final week running the system alongside the old way. Faster if the business is digital-first.]]></content:encoded>
  </item>
  <item>
    <title>AI workflow automation in the UK: what it actually replaces, and what it does not</title>
    <link>https://expedientai.co.uk/feed/ai-workflow-automation-uk/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/ai-workflow-automation-uk/</guid>
    <pubDate>Sun, 12 Jul 2026 00:00:00 +0000</pubDate>
    <description>AI workflow automation for UK small and medium businesses. The specific tasks it replaces well, the ones it does not, and the architectural pattern that holds up in production.</description>
    <content:encoded><![CDATA[Tips AI workflow automation in the UK: what it actually replaces, and what it does not AI workflow automation for UK small and medium businesses. The specific tasks it replaces well, the ones it does not, and the architectural pattern that holds up in production. 2026-07-12 · 8 min read UK small business owners are being sold two stories about AI workflow automation. The first story says AI will transform the business. The second story says AI is too expensive for small businesses. Both stories are wrong. The truth is in the middle. AI workflow automation for a UK small business is the right-sized system that runs the repetitive tasks the business already does, the ones that lose money when they are done badly, that nobody enjoys doing, and that follow predictable patterns. If you can name one task you do every week that you wish you did not, the system is right for you. If you cannot name one, you are looking for a transformation that does not exist yet. What AI workflow automation actually replaces Three automations cover the bulk of the value for most UK small businesses. Drafting replies to incoming enquiries. A customer emails about a leaking radiator. The system reads the message, finds the last three jobs the customer had, pulls standard pricing, and writes a reply in the plumber's voice. He reads, edits if he cares, and presses send. Time saved: about forty minutes a day for a one-man business. Booking jobs into a calendar. When a customer confirms a time, the system updates the calendar, sends a WhatsApp reminder the day before, and writes the job onto the invoice when the work is done. Three touch points. Zero manual work. The customer gets a confirmation in under a minute. Chasing late payments. Customers fourteen days past their invoice get a polite WhatsApp message. Twenty-eight days gets a firmer email. Forty-two days gets a phone call from the plumber himself. The cheap chore, automated. The harder conversation, kept human. This is not "AI everywhere". This is three specific tasks, running reliably, paying for themselves. What AI workflow automation does not replace Three limits worth being honest about. Money, contracts, complaints, keep a human in the loop. Open-weight models get tone and facts right most of the time, but the ten percent they miss is the kind of error that loses a customer. A polite WhatsApp message at fourteen days is automated. A firm phone call at forty-two days is a human job. Up-to-the-minute UK-specific information still needs a search step on top of the model, regulation changes, supplier pricing, local council rules. The model will guess confidently if you let it. Sensitive personal data needs a careful audit, medical notes, financial advice, safeguarding concerns. The models aren't unsafe, but they aren't pre-tuned for these domains the way dedicated tools are. In every case where the answer matters, the model is the draftsperson, not the publisher. The plumber reads every reply before it goes out. The salon owner reviews every quote. The cafe manager checks every health-and-safety note. The model writes. The human decides. The architecture that holds up in production The architecture is the same one we use in Vereby, the multi-tenant platform we built for UK small businesses. The components are: Zoho CRM (or HubSpot, Pipedrive, or whatever CRM the business already uses) holds the data. The AI sits behind the scenes. The team logs into the same CRM they always have. OpenRouter routes the AI calls. Most of the work goes to open-weight models, MiniMax, DeepSeek, GLM, at roughly a tenth of the cost of a frontier model. When something matters (a complaint, a pricing dispute, a final review before send) the call gets bumped to Claude or GPT. The router picks automatically. A small Next.js app runs on Railway Hobby, five pounds a month. The only screen the team ever opens. Three things on it at any time: unpaid jobs, waiting enquiries, quotes needing a follow-up. No dashboard sprawl. A brand-voice configuration is the secret sauce. The system scrapes the business's existing public content, the website, the social posts, the email signature, and learns the voice. Drafts read like the business, not like a bot. Total monthly cost: under forty pounds for a one-man trades business. Linear scaling for larger teams. The four-step process for setting up AI workflow automation The setup runs in four weeks, not four months. We have written about the architecture in detail on the approach page. The short version: Week 1, Discovery. A one-hour call to learn the business. What tasks does the team do every week that they wish they did not? What is a good reply? What is a bad reply? What tone do they use with which customer? What does a missed-call cost the business? Week 2, Data wiring. Connect the CRM. Connect the calendar. Connect the WhatsApp. Connect the email. The system reads the data, not the other way around. Week 3, Brand voice training. The system reads the business's existing public content and learns the voice. The plumber's voice, not a generic UK trades voice. The salon owner's voice, not a generic UK salon voice. The first drafts are reviewed by the team. The system iterates. Week 4, Shadow mode. The system runs alongside the existing workflow. The team compares the system's drafts to their own. After a fortnight, if the drafts are good enough, the system goes live. If they are not, the setup is not finished. What this means for a UK small business If you run a small UK business and the above describes what you need, the system is right for you. The three automations above cost about thirty-eight pounds a month. The setup takes a few weeks. The hard part is having the patience to test the system in production for a fortnight before you trust it. If you cannot do it yourself, that's fine. We build these systems for a living, see the services page for how the engagement works, or the work page for what's already shipping. The forty-quid plumbing firm post is a working example of the right-sized stack. If you are looking for a transformation, you are looking for something that does not exist yet. What exists is the next three tasks, automated well, running reliably, paying for themselves in the time your team stops losing. You can read about how I build these systems in the Approach section, or see real builds in my Work (Vereby is the multi-tenant platform that powers setups like this one). Frequently asked questions What is AI workflow automation? AI workflow automation is software that runs the repetitive tasks a small business does every day - quote replies, calendar booking, invoice chasing, social posts, weekly reports. The system uses a language model to read, write, and route messages, and a router to choose between cheap and frontier models per task. The work is the same. The volume is what changes. What can AI workflow automation do for a UK small business? Three high-value automations: drafting replies to incoming enquiries in your own voice; booking jobs and sending reminders without anyone opening a calendar; chasing late payments on a polite schedule. Together they save most small UK businesses eight to twelve hours a week. What is AI workflow automation NOT good at? Anything involving money, contracts, or complaints needs a human in the loop. Anything up-to-the-minute (regulation, supplier pricing, local council rules) needs a search step. Anything genuinely novel for the business is a human job. The model writes. The human decides. How do I know if AI workflow automation is right for my UK small business? If you can name one task you do every week that you wish you did not, the system is right for you. If you cannot name one, you are looking for a transformation that does not exist yet. Start with the one task. Get it right. Then the next one.]]></content:encoded>
  </item>
  <item>
    <title>AI voice agent for Bristol small businesses: the 24/7 receptionist that costs less than a temp</title>
    <link>https://expedientai.co.uk/feed/ai-voice-agent-bristol/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/ai-voice-agent-bristol/</guid>
    <pubDate>Sun, 12 Jul 2026 00:00:00 +0000</pubDate>
    <description>How UK small businesses in Bristol and the South West use AI voice agents to answer out-of-hours calls, book jobs, and capture enquiries. The architecture, cost, and what to expect.</description>
    <content:encoded><![CDATA[Tips AI voice agent for Bristol small businesses: the 24/7 receptionist that costs less than a temp How UK small businesses in Bristol and the South West use AI voice agents to answer out-of-hours calls, book jobs, and capture enquiries. The architecture, cost, and what to expect. 2026-07-12 · 8 min read An AI voice agent answers real phone calls using a language model for understanding and a text-to-speech system for speaking back. For UK small businesses in Bristol and the South West, the cost is under fifty pounds a month for the AI layer plus a UK number. The agent handles the sixty to seventy percent of calls that follow predictable patterns: quote requests, opening hours, booking, status updates. The thirty percent that need human judgement (complaints, complex questions, angry customers) get handed off. UK small businesses in Bristol and the South West have a specific problem. A plumber in Bedminster misses twelve calls on a Tuesday. A salon in Clifton misses eight calls on a Saturday. An electrician in Stokes Croft misses fifteen calls during a job. Each missed call is a lost customer, and most of those customers never call back. The traditional fix is a temp receptionist, or an out-of-hours answering service, or a second phone line that forwards to a mobile. Both work. Both cost more than they should. The AI voice agent is the third option, and it is the one that has finally matured enough to be useful in 2026. What an AI voice agent does An AI voice agent is software that answers a real phone call using a language model for understanding and a text-to-speech system for speaking back. The caller hears a natural voice, not a robot. The agent can take a message, book a job, answer a question, or transfer to a human. The technology has three pieces. Speech-to-text converts the caller's voice into text the model can read. A language model interprets the question and decides what to do. Text-to-speech converts the model's response back into a natural voice. All three run in under a second on a modern voice agent, so the caller hears a normal conversation pace, not the lag of a chatbot. For UK small businesses, the AI voice agent handles the sixty to seventy percent of calls that follow predictable patterns: quote requests, opening hours, booking, status updates. The thirty percent that need human judgement (complaints, complex questions, angry customers) get handed off. The AI books the call. The human takes the call. What an AI voice agent costs for a Bristol small business Three components. The UK number (a Bristol or 0117 area code number is roughly five pounds a month). The AI minutes used (roughly fifteen to twenty-five pence each, around fifteen to thirty pounds a month for a typical Bristol trades business). The integration into your booking system or CRM (one-off build cost, depending on complexity). Total monthly: under fifty pounds a month. Total build: depends on the business, but the basic version is a few afternoons. The comparison is brutal. A temp receptionist costs three hundred to four hundred pounds a day. An out-of-hours answering service costs two to three hundred pounds a month for basic message-taking. An AI voice agent costs fifty pounds a month for basic message-taking, two hundred pounds a month for booking plus a polite conversation with the caller about which service they need. The technology is no longer the bottleneck. The bottleneck is whether your CRM is ready to receive what the agent captures. How it sounds: a real Bristol conversation The technology has come far enough that the agent can handle a Bristol accent, a Cardiff accent, a Glaswegian accent, and most regional variations. The system is trained on UK English. The first thing the caller hears is the agent's name and the business's name, in a clear natural voice. The second thing the caller hears is the question. A real call goes like this. The plumber's number rings. The agent picks up: "Hi, this is the AI receptionist for Smith Plumbing in Bristol. How can I help?" The caller says: "I have a leaking radiator, can someone come out this week?" The agent says: "Of course, can I take your name and postcode?" The caller gives them. The agent says: "I'll get one of the team to call you back this afternoon. Can I take a contact number?" The caller gives the number. The agent says: "Perfect, you'll hear from us by 5pm. Anything else?" The caller says no. The agent says: "Thanks for calling. Bye." The call ends. Twenty-two seconds. The plumber sees the lead in his admin dashboard before he puts the phone down. The architecture that holds up in production The components are: Twilio (or Aircall, or any UK telephony provider) holds the UK number and routes incoming calls. The agent picks up on the first ring. Retell AI (or Vapi, or Bland) is the voice agent platform. It does the speech-to-text, language model, and text-to-speech in one package. The agent's voice, prompt, and behaviour are all configurable from a single dashboard. OpenRouter is the LLM gateway for the agent's brain. Cheap models for routine calls (quote requests, opening hours). Frontier models for the calls that need better reasoning (complex booking, edge cases). A small Next.js app runs on Railway Hobby, five pounds a month. The only screen the team ever opens. Three things on it at any time: missed calls, today's bookings, voicemails. No dashboard sprawl. Your CRM holds the booking, the customer record, and the job. The agent writes into it the moment the call ends. The full architecture is documented in the Vereby case study, Vereby is the multi-tenant platform that powers voice agents for our customers. What this means for a Bristol small business If you run a trades firm, a salon, a clinic, or any UK small business that misses calls and loses customers, the AI voice agent is right for you. The technology has come far enough that the calls feel like calls. The cost is under fifty pounds a month. The setup takes a few afternoons. If you cannot do it yourself, that's fine. We build these systems for a living, see the services page for how the engagement works, or the work page for what's already shipping. The forty-quid plumbing firm post is a working example of the right-sized stack for a trades business. Frequently asked questions What is an AI voice agent? An AI voice agent is software that answers a real phone call using a language model for understanding and a text-to-speech system for speaking back. The caller hears a natural voice, not a robot. The agent can take a message, book a job, answer a question, or transfer to a human. The cost is roughly fifteen to twenty-five pence per minute, far less than a temp. Do AI voice agents work for UK small businesses? Yes - for the sixty to seventy percent of calls that follow predictable patterns: quote requests, opening hours, booking, status updates. The thirty percent that need human judgement (complaints, complex questions, angry customers) get handed off. The right architecture is AI for volume, human for judgement. What does an AI voice agent cost for a Bristol small business? Three components: the telephony (a UK number, roughly five pounds a month), the AI minutes used (roughly fifteen to twenty-five pence each, around fifteen to thirty pounds a month for a typical Bristol trades business), and the integration into your booking system or CRM. Total: under fifty pounds a month, replacing a temp or out-of-hours answering service at a fraction of the cost. Will my customers know they are talking to an AI voice agent? Best practice in the UK: yes, tell them. The opening line should be \"Hi, this is the AI receptionist for X - is that OK with you?\" Most customers say yes. The five percent who say no get transferred to a human or a voicemail. The transparency builds trust in a way that pretending-to-be-human never does.]]></content:encoded>
  </item>
  <item>
    <title>AI for small business UK: where to start, what to skip, and what it actually costs</title>
    <link>https://expedientai.co.uk/feed/ai-for-small-business-uk/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/ai-for-small-business-uk/</guid>
    <pubDate>Sun, 12 Jul 2026 00:00:00 +0000</pubDate>
    <description>A practical first-week playbook for UK small businesses adopting AI. What to try first, what to defer, and what the realistic cost is. No vendor pitch.</description>
    <content:encoded><![CDATA[Tips AI for small business UK: where to start, what to skip, and what it actually costs A practical first-week playbook for UK small businesses adopting AI. What to try first, what to defer, and what the realistic cost is. No vendor pitch. 2026-07-12 · 7 min read UK small business owners have heard the pitch enough times to be wary. "AI will transform your operations." "AI will free your team from repetitive work." "AI is the future of small business." Every vendor says it. None of them say where to start. This post is where to start. We work with UK small and medium businesses every week on AI workflow automation. The pattern is the same: businesses that succeed with AI start small, get one thing working, and add the second thing six weeks later. Businesses that fail with AI try to do everything at once, buy five tools, fire the workflow nobody trusts, and end up with another monthly subscription nobody opens. The first task is the one your team already does badly and does often. For most UK small businesses, that is one of three: drafting replies to incoming enquiries, booking jobs into a calendar, or chasing late payments. All three follow predictable patterns. All three happen every week. All three lose money when they are done badly. None of them need a frontier model to do them well. Pick one. Run an open-weight model on it. Keep a human in the loop on anything involving money, contracts, or complaints. We have written about the right-sized stack in detail on the forty-quid plumbing firm post. The short version: open-weight models from MiniMax, DeepSeek, and GLM now do the bulk of SME admin work for under forty pounds a month. Frontier models from Claude and GPT are reserved for the ten percent where the cost of being wrong is high. The router, not the model, is what makes the system cheap. What AI does well in a small UK business Three automations, each running without anyone asking. Drafting replies to incoming enquiries. A customer texts about a leaking radiator. The system reads the message, finds the last three jobs the customer had, pulls standard pricing, and writes a reply in the plumber's voice. He reads, edits if he cares, and presses send. Time saved: about forty minutes a day. Booking jobs into a calendar. When a customer confirms a time, the system updates the calendar, sends a WhatsApp reminder the day before, and writes the job onto the invoice when the work is done. Three touch points. Zero manual work. Chasing late payments. Customers fourteen days past their invoice get a polite WhatsApp message. Twenty-eight days gets a firmer email. Forty-two days gets a phone call from the plumber himself. The cheap chore, automated. The harder conversation, kept human. This is not "AI everywhere". This is three specific tasks, running reliably, paying for themselves. What AI does not do well in a small UK business Three limits worth being honest about. Money, contracts, complaints, keep a human in the loop. Open-weight models get tone and facts right most of the time, but the ten percent they miss is the kind of error that loses a customer. Up-to-the-minute UK-specific information still needs a search step on top of the model, regulation changes, supplier pricing, local council rules. The model will guess confidently if you let it. Sensitive personal data needs a careful audit, medical notes, financial advice, safeguarding concerns. The models aren't unsafe, but they aren't pre-tuned for these domains the way dedicated tools are. In every case where the answer matters, the model is the draftsperson, not the publisher. The plumber reads every reply before it goes out. The salon owner reviews every quote. The cafe manager checks every health-and-safety note. The model writes. The human decides. How long AI setup takes for a small UK business Two to four weeks for a trades firm with a CRM already in place. A week to map the actual workflows (not the ones in the brief, the ones the business actually runs), a week to wire the data, a week to train the model on the business's voice and tone, and a week running the system alongside the old way before switching over. Faster for a service business with a CRM already in place. Slower for a brand-new business with no data history. The full breakdown is on our approach page. What this means for you If you run a small UK business and wonder whether you can do this yourself, you probably can. The three automations above cost about thirty-eight pounds a month. The setup takes a few afternoons. The hard part is having the patience to test the system in production for a fortnight before you trust it. If you cannot do it yourself, that's fine. We build these systems for a living, see the services page for how the engagement works, or the work page for what's already shipping. The forty-quid plumbing firm post is a working example of the right-sized stack. If you are looking for a transformation, you are looking for something that does not exist yet. What exists is the next three tasks, automated well, running reliably, paying for themselves in the time your team stops losing. You can read about how I build these systems in the Approach section, or see real builds in my Work (Vereby is the multi-tenant platform that powers setups like this one). Frequently asked questions What is the first AI task a UK small business should try? The first task is the one your team already does badly and does often. Quote replies, missed-call follow-ups, invoice chasing, weekly social posts. Pick one. The system that handles one task well teaches you what the next one needs. Do I need a developer to set up AI in a UK small business? For off-the-shelf tools (ChatGPT Team, Microsoft Copilot, Google Workspace Gemini) no. For a system that grounds replies in your own data and runs unattended, you usually do. Most UK setups take two to six weeks at the small studio level. How much does AI cost for a small business in the UK? Open-weight models now handle ninety percent of admin work for under forty pounds a month. Frontier models (Claude, GPT) are reserved for the ten percent where the cost of being wrong is high. Total stack for a one-person trades business: roughly thirty-eight pounds a month all-in. What should I not do when adopting AI in a small business? Three things: do not automate anything that involves a customer's money, contract, or complaint without a human in the loop. Do not buy five AI tools at once. Do not promise your team or your customers that AI will replace them - it changes the work, it does not delete it.]]></content:encoded>
  </item>
  <item>
    <title>Top 10 practical changes in Hermes Agent v0.18.0 (and what they're for)</title>
    <link>https://expedientai.co.uk/feed/hermes-0-18-practical-changes/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/hermes-0-18-practical-changes/</guid>
    <pubDate>Fri, 03 Jul 2026 00:00:00 +0000</pubDate>
    <description>The Judgment Release dropped on 2026-07-01. Ten changes from v2026.7.1 that actually shift how a one-person operator runs Hermes day-to-day, anchored in the release notes, peer-to-peer tone, with the headline grabbers deliberately passed over.</description>
    <content:encoded><![CDATA[Tools Top 10 practical changes in Hermes Agent v0.18.0 (and what they're for) The Judgment Release dropped on 2026-07-01. Ten changes from v2026.7.1 that actually shift how a one-person operator runs Hermes day-to-day, anchored in the release notes, peer-to-peer tone, with the headline grabbers deliberately passed over. 2026-07-03 · 9 min read We use Hermes Agent as our day-to-day terminal assistant at the studio. Multiple named profiles, Telegram as the primary comms channel, daily cron briefings, and a skills library that's grown fat over the last year. So when the Judgment Release (v2026.7.1) dropped on the first of July, we did what we do with any release: pulled the notes, pulled the PRs, and asked which changes actually move the needle on a one-operator workflow. These are the ten. They are not the headline grabbers. They are the ones that change how a single person uses Hermes across a working day. 1. delegate_task runs in the background and reports once when everything finishes What it is: When you kick off parallel subagents, your chat is no longer blocked. The dispatcher returns immediately; you keep working; and when every subagent has finished, their results come back as one consolidated block, not one message per subagent. Why it matters: Today's pain: "I asked three agents to research three things. Now I'm staring at the screen waiting. Three minutes. Five." With this, you kick it off, send the next message, and the consolidated result lands when all three are done. Practical example: "Audit these three modules for security issues in parallel" → you can immediately say "while that runs, show me last week's cron logs" in the same chat. When the audit finishes, you get one block with all three reports stacked. Under the hood (#49734): one async-pool slot owns the whole batch; results roll into one {results:[...]} block; orchestrator subagents stay synchronous so deeper fan-out behaviour is preserved. 2. /learn , turn any session into a skill What it is: One slash command, takes a directory, URL, the workflow you just walked, or pasted prose, and distils a Hermes-compatible skill out of it, authors SKILL.md to your CONTRIBUTING.md standards automatically. Why it matters: Today we save skills through skill_manage(action='create') after noticing a pattern. With /learn , you do it inline: /learn ~/projects/vereby-backend after finishing a deploy, and the next time, that workflow is one keystroke away. Works the same on CLI, Telegram gateway, TUI, dashboard and desktop because it's an open-ended agent turn rather than a host-side engine. Practical example: "I just spent 20 minutes talking you through the Vereby webhook → email-handoff flow. Now /learn that and next time you do it from one command." 3. /prompt (alias /compose), write long prompts in your editor What it is: Opens $EDITOR so you can hand-write a multi-line prompt in real markdown instead of fighting a one-line input. Why it matters: When you're about to delegate something meaty, "audit this codebase against these six criteria and write a report", typing it as one line in the CLI is friction. Now you do /prompt (or Ctrl+G in the TUI), the editor opens, you write the spec in proper markdown, save, and it sends as the next turn. Practical example: /prompt pre-seeds with seed text, so /prompt "Review the auth module for" opens with that in the file, and you keep typing. 4. Cron continuations, reply to a cron delivery and the agent has context What it is: Cron job deliveries become continuable, on Telegram (and other threaded platforms), a dedicated thread opens. On Slack, flat in-channel continuable delivery. You reply, the agent has the brief in context instead of starting fresh. Why it matters: Today: cron briefing lands in Telegram, you want to drill into one section, you send a long message restating the question. Now: you just reply in the thread. The agent already knows what it sent. Also fixed in the same release wave: Cron base_url overrides that could exfiltrate provider credentials are now blocked at the scanner, a direct attack-surface win if any of your cron fetches untrusted URLs. Missed-grace jobs run once after downtime instead of being deferred indefinitely. Cron jobs layer enabled MCP servers onto per-job toolsets, better scoping, less credential bleed. Practical example: Morning briefing mentions a Vereby deploy result. You reply "show me the timeout config it landed on" in the same thread, no re-prompting. 5. /goal completion contracts, the agent proves it's done What it is: /goal now accepts a completion contract, what "done" looks like, and judges the work against evidence (running your project's checks) instead of stopping when the model feels like it. A profile-scoped evidence ledger backs this up. Why it matters: The classic LLM failure mode: "I think I fixed it." With completion contracts, you say "done means pytest -q exits 0 and the integration test passes", and the goal loop won't accept completion until that's actually true. Also released in 0.18.0: /goal wait parks the standing-goal loop on a background process, plus a pre_verify hook for wiring in custom checkers. Practical example: "/goal stop the kanban dashboard when smoke-tests pass and no process is on port 8000", the agent keeps poking until both are verified, not until it thinks they're verified. 6. Profile descriptions make kanban decomposition route by role What it is: New flag on hermes profile create : hermes profile create xyz_app --description "FastAPI multi-tenant backend specialist" The description is read by the kanban decomposer so it can route tasks to roles, not just profile names. Skip it and add later via hermes profile describe . Why it matters: When you say "build me an xyz app", an LLM-based decomposer reads the description of every profile you have and routes the work to the right one, without you spelling out a profile name. Today this is opaque (you needed to know the names). Now it's semantic. Practical example: You have a vereby_ops_manager profile whose description is "Vereby multi-tenant backend, MCP connectors, Resend email pipeline." When you say "investigate the slow Resend webhook," the decomposer picks it up automatically because the role matches. 7. /journey + desktop memory graph, see and prune what your agent knows What it is: CLI and TUI get a learning timeline showing every memory and skill the agent has accumulated, with in-place edit/delete. The desktop app pairs this with a playable radial memory graph (top-down, scrollable). Why it matters: Memory accumulates. Without visibility, you eventually hit token pressure, lose track of which assumptions are still true, and silently make decisions on stale facts. With this, you can see when a memory was saved, what triggered it, and prune it on sight. Practical example: "I see my agent remembered that Vereby uses MariaDB. We migrated two months ago." /journey → delete that memory → next session, no more stale-database decisions. 8. Slack cron continuation + Telegram command-menu tweaks What it is: Flat in-channel cron continuations on Slack (so replied-in-channel works on Slack). Configurable Telegram command menu with a higher default cap so skills stay visible. Drain general send pool on Telegram timeout before retry. All small reliability wins to the messaging layer. Why it matters: If Telegram is your primary channel, the visible improvement is: your skills and slash commands don't fall off the menu as you accumulate them. Slack users get the threading equivalent of Telegram's continuations. Practical example: This is the quiet-but-every-week kind of fix, fewer "where did my command go" moments. 9. Background self-improvement routes to a cheaper model What it is: The post-turn self-improvement fork (the one that decides whether to save a memory or distil a skill at the end of each turn) now routes to an auxiliary model, digests context instead of replaying the full conversation, and adapts cadence. Why it matters: Today every turn ends with a fork that costs about as much as the main turn. With this, it's 5–10% of the cost. On high-velocity sessions that adds up. You keep the self-improvement; you stop paying full main-model price for it. Practical example: You don't have to do anything. Your /usage numbers just go down without losing the memory pipeline. 10. /reasoning full, uncapped thinking, with timestamps in /history What it is: /reasoning full removes the budget cap on the model's reasoning trace. /timestamps (and timestamps in /history ) show when each message happened in long sessions. Why it matters: When you hit a hard problem and want to know what the model thought before answering, like debugging a failed delegate or a stale-fact loop, uncapped reasoning shows the whole chain, not a clipped summary. Practical example: "Diagnose why the Vereby webhook ack came back 200 but no email landed." The model thinks for 30 seconds visibly, shows its reasoning, then answers. The reasoning is evidence, not magic. Frequently asked questions What is Hermes Agent? Hermes Agent is a terminal-first AI assistant built by Nous Research. It runs locally, supports multiple named profiles for separate projects, integrates with Telegram and other messaging channels as a comms layer, and is designed for operators who want a long-running daily assistant rather than a one-shot chat tool. Do I need to be a developer to use the changes in Hermes v0.18.0? Most of the changes (delegation, cron continuations, /journey, /reasoning full) work the same regardless of how technical you are, they're surface-level UX wins. The deeper ones (/goal completion contracts, kanban profile routing, /learn distilling workflows into skills) require you to be comfortable configuring your own project and writing a short verification contract. What's the difference between /learn and skill_manage(action='create')? skill_manage creates a skill from a prompt you've already written, and is best when you know exactly what skill you want. /learn distils a skill from context, a directory you've just worked in, a URL with documentation, or the conversation you just had. Use /learn when the skill shape is still emerging; use skill_manage when you've already decided what the skill should look like. Will /reasoning full slow my agent down? It removes the budget cap on the thinking trace, so on hard problems the model may think longer than it would with the default capped budget. On routine questions there's no measurable difference. If you find a session is over-thinking, switch back to the default with /reasoning (without 'full').]]></content:encoded>
  </item>
  <item>
    <title>The forty-quid plumbing firm: what one trades business actually runs</title>
    <link>https://expedientai.co.uk/feed/cheap-ai-forty-quid/</link>
    <guid isPermaLink="true">https://expedientai.co.uk/feed/cheap-ai-forty-quid/</guid>
    <pubDate>Mon, 29 Jun 2026 00:00:00 +0000</pubDate>
    <description>A working breakdown of the AI workflow stack a UK plumbing firm runs for under forty pounds a month. Real numbers, real models, architecture-led cost control that holds up for a trades business.</description>
    <content:encoded><![CDATA[News The forty-quid plumbing firm: what one trades business actually runs A working breakdown of the AI workflow stack a UK plumbing firm runs for under forty pounds a month. Real numbers, real models, architecture-led cost control that holds up for a trades business. 2026-06-29 · 5 min read A plumber in Sheffield just answered 47 customer enquiries before his kettle boiled, using a forty-quid-a-month AI stack. It drafts replies to quote requests, books jobs into his calendar, and chases missed payments, right-sized AI for a one-man UK trades business that barely writes emails. He didn't write a single line of code. The promise, in one breath: Under £40/month, all-in, for a one-man UK plumbing firm's AI workflows. Zoho CRM + OpenRouter + a small Next.js app on Railway Hobby. Three automations that run themselves: quote replies, calendar booking, late-payment chasing. Small UK businesses, plumbers, salons, cafes, lose eight to twelve hours a week to admin that right-sized AI can handle in minutes. Open-source and open-weight models now do that work for less than a takeaway coffee per day, closing the gap between a three-person trades firm and a chain with a CRM team. The firms winning right now started small last quarter. The stack, in plain English Three monthly services, each under fifteen pounds, that a tradesperson can run from their phone. Zoho CRM holds every job, every customer, every quote. The plumber types into it the way he always has, on his phone in the van. Nothing changed on his side. OpenRouter routes the AI calls. Most of the work goes to open-source and open-weight models, MiniMax, DeepSeek, GLM, at roughly a tenth of the cost of a frontier model. When something matters (a complaint, a pricing dispute, a final review before send) the call gets bumped to Claude or GPT. The router picks automatically. A small Next.js app runs on Railway Hobby, five pounds a month. The only screen the plumber ever opens. Three things on it at any time: unpaid jobs, waiting enquiries, quotes needing a follow-up. No dashboard sprawl. Total monthly cost: thirty-eight pounds. Less than most tradespeople spend on tools, less than one missed appointment. AI workflows for a small UK plumbing firm: what runs every morning Three automations, each running without anyone asking. 1. Drafts replies to incoming quote requests. A customer texts about a leaking radiator. The system reads the message, finds the last three jobs the customer had, pulls standard pricing, and writes a reply in the plumber's voice. He reads, edits if he cares, and presses send. Time saved: about forty minutes a day. 2. Books jobs into the calendar. When a customer confirms a time, the system updates the calendar, sends a WhatsApp reminder the day before, and writes the job onto the invoice when the work is done. Three touch points. Zero manual work. 3. Chases missed payments. Customers fourteen days past their invoice get a polite WhatsApp message. Twenty-eight days gets a firmer email. Forty-two days gets a phone call from the plumber himself. The cheap chore, automated. The harder conversation, kept human. This is not "AI everywhere". This is three specific tasks, running reliably, paying for themselves. Cheap AI stack for trades businesses: why the cost stays low The real cost isn't the model price, it's unanswered customer enquiries and forgotten invoices. For most small UK trades businesses, saving those eight to twelve admin hours a week is worth thousands of pounds a year; the AI stack that does it costs tens. The price difference between models is small in absolute terms, four pence versus six pence per reply, about forty pence a day, one hundred and fifty pounds a year. Not the money that decides whether to add AI. The money that decides it is the time you stop losing. Open-source and open-weight models have come far enough that, for ninety percent of the work, the output is indistinguishable from the frontier ones. They don't hallucinate as much. They follow instructions better. They handle UK English without a fight. If you tried them eighteen months ago and were unimpressed, try them now. Where this stack is not the right tool Three limits worth being honest about. Money, contracts, complaints, keep a human in the loop. Open-source and open-weight models get tone and facts right most of the time, but the ten percent they miss is the kind of error that loses a customer. Up-to-the-minute UK-specific information still needs a search step on top of the model, regulation changes, supplier pricing, local council rules. The model will guess confidently if you let it. Sensitive personal data needs a careful audit , medical notes, financial advice, safeguarding concerns. The models aren't unsafe, but they aren't pre-tuned for these domains the way dedicated tools are. In every case where the answer matters, the model is the draftsperson, not the publisher. The plumber reads every reply before it goes out. The salon owner reviews every quote. The cafe manager checks every health-and-safety note. The model writes. The human decides. How long AI setup takes for small UK businesses Two to four weeks for a trades firm like the one above. A week to map the actual workflows (not the ones in the brief, the ones the business actually runs), a week to wire the data, a week to train the model on the business's voice and tone, and a week running the system alongside the old way before switching over. Faster for a service business with a CRM already in place. Slower for a brand-new business with no data history. I wrote about how I scope the work on my approach page . If you run a small business and wonder whether you can do this yourself, you probably can. The three services above cost about thirty-eight pounds a month. The setup takes a few afternoons. The hard part is having the patience to test the system in production for a fortnight before you trust it. What I'm watching for in the next twelve months Voice agents that take a real phone call without an appointment. The technology has improved enough this year that I'm about to deploy two voice agents for UK clients, one for a clinic and one for a trades firm. Both will answer the phone out of hours, take messages, and book jobs for the morning. Watch this space. Better memory. Every system I build today learns something about the business and forgets it the next morning. The next generation of open-source and open-weight models will let the system remember a customer's preferences across months of conversation without me re-teaching it. That is when the real change happens, most small-business admin is "remember this preference and apply it consistently for the next year", not just "answer this query". If you run a small UK business and want to know where to start, identify the one admin task that takes the most time each week and run an open-source model on it. Not the whole company. One task. Get that right, then the next one. You can read about how I build these systems in the Approach section , or see real builds in my Work (Vereby is the multi-tenant platform that powers setups like this one). Frequently asked questions What is the cheapest AI a small UK business can run today? For most admin tasks, drafting replies, summarising documents, scheduling, open-source and open-weight models like MiniMax, DeepSeek and GLM handle ninety percent of the work for under forty pounds a month. Frontier models (Claude, GPT) are reserved for the ten percent where the cost of being wrong is high. Do I need a developer to set up AI workflows for my small business? For off-the-shelf tools (ChatGPT Team, Google Workspace Gemini, Microsoft Copilot) no. For a system that grounds replies in your own knowledge and runs on the architecture behind this kind of right-sized stack, you usually do. Most setups take two to six weeks at the small UK studio level. Is right-sized AI safe for customer-facing replies? Open-source and open-weight models are safe for the eighty percent of replies that follow your existing templates and tone. The twenty percent that involve money, complaints, or commitments need a human in the loop or a frontier model. The right-sized stack buys you capacity, the human stays in charge of judgement. How long before AI pays for itself in a small business? For admin-heavy trades (plumbers, electricians, salons) eight to twelve hours a week is the typical saving. At forty pounds a month all-in, the system pays for itself when one missed appointment or one chased invoice is recovered per month.]]></content:encoded>
  </item>
</channel>
</rss>
