AI agents for small business UK: what to actually expect from a supplier
Most small businesses looking at AI agents are being sold a demo, not a working system. The fair test is whether a supplier can show the thing running on a real business, with real faults on record and real fixes against them, not just a script. This page uses one such business, an independent practice in Liverpool, as the worked example throughout.
- Stays exactly where it is
- Never copied out, never trained on
- Voice receptionist and call summaries
- Patient SMS and WhatsApp replies, bookings
- Orders, lab tracking, delivery notes
- Recall queue and reactivation
- Daily brief and complaints digest
The boundary, drawn. Clinical records on the left never cross to the right.
What an AI agent should mean for a small business
An AI agent takes on one job with clear edges, answering the phone, replying to a customer, moving a job along a pipeline, and does it the same way every time, inside rules that don't bend under pressure. For any business handling customer data, the first thing that should be said, early and without being asked, is what stays in the business's existing systems and what the AI actually touches. At Winstanley & Son, an independent opticians, that boundary is stated up front on every sales call: clients keep their existing practice management software; clinical records never leave it; the AI runs only the operational layer, messages, bookings, orders, and reporting, around it.
The honest-AI checklist
Before taking any AI agents supplier at their word, a small business owner can reasonably ask for evidence of each of these, not just a claim:
- A real install, not just a demo. Has this run on an actual, trading business, with real customers, for long enough to have hit real edge cases?
- A visible incident log. Every system fails sometimes. A supplier who can show you what broke, when, and exactly what changed permanently as a result is a safer bet than one who claims nothing has ever gone wrong.
- Refuse-rather-than-guess behaviour. Does the system say "I'm not sure" and stop, or does it confidently make something up? Two customers with the same surname should get a clarifying question, not a coin-flip.
- A confirm-before-write step on anything costly to get wrong. An order, a booking, a payment: does the system read it back and wait, or does it act on a best guess?
- A human override that actually works. Can a staff member take over a single customer's conversation instantly, and is there a single practice-wide stop for a genuine emergency?
- A data boundary stated in writing, not just in a sales pitch. What exactly does the AI read and write, and what does it never touch?
- Evidence over approval badges. No UK body licenses the use of AI on customer or patient data, so a supplier claiming an official "approval" should be questioned; what a serious supplier can show instead is a data protection impact assessment, a clear sub-processor list, and a human confirmed to be in the loop.
The medical-device line, and why it matters beyond healthcare
Software that transcribes, drafts or summarises for a person who then approves it is generally not a regulated medical device. Software that goes further, and makes the decision itself, such as triaging symptoms or recommending a clinical course of action, can cross into regulated territory. The general lesson travels beyond healthcare: the safer, and often the more useful, design keeps a person making the judgement call and has the AI do the preparation and the admin around it, not the decision itself.
Why one practice per area, not a subscription free-for-all
Crewley takes on one practice per area at a time, on a renewing exclusivity period, rather than selling the same install to every competing business in a town. The model is a head office without a franchise: the build, the guardrails and the fixes are supplied centrally, but ownership of the customer relationship, and the profit from it, stays with the business itself.
Who this is not built for
This is a done-for-you build with a person driving it into your business alongside you, not a self-serve product you configure alone: a practice manager or business owner reviews every workflow change before it goes live, and the staff interface runs in Slack, so a team with no appetite for a Slack-based workflow, or no willingness to sit through that review, is not the right fit yet. [FILL: any additional stated criteria for who Crewley will and won't take on, beyond the exclusivity and review-in-the-loop points already on record]
Questions small business owners ask
Is there a UK body that approves AI systems for handling customer data?
No single body licenses this, whatever a vendor's marketing implies. The real question to ask a supplier is what evidence they can show: a signed data protection impact assessment, a clear list of who else touches the data, and a documented human-in-the-loop step.
How do I know an AI agent won't just make things up to my customers?
Ask to see it refuse. A system built properly will say it isn't sure and ask a clarifying question rather than guess, and it should read back anything costly, like an order or a booking, before creating it.
What happens when the AI gets something wrong?
Ask for the incident log, not a claim of perfection. A supplier who can point to a real fault, what it changed permanently as a result, is more trustworthy than one who says nothing has ever gone wrong.
Does this replace my existing software?
No, in a well-built system it should not. It should sit alongside what you already use and run only the operational layer around it, leaving your core records exactly where they are.
Why does Crewley only take one business per area?
So the practice that installs it isn't competing against a neighbour running the exact same system. Ownership of the customer relationship and the results stays with the business, not shared with a competitor down the road.
Where can I see this working, not just described?
See the agents overview for the full picture of what runs at Winstanley & Son today, the AI receptionist page and the patient recall page for two of the agents in detail, or the case study for the worked example in full.
See how this checklist stacks up against your own business before you commit to any AI supplier.
Book a free recall and response audit