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.
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).
