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.
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.
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.
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.
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.
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.
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