Does a small business actually need AI? A test you can run in an afternoon, and the cases where the answer is no.

The surveys are consistent. The ONS found 35 per cent of UK businesses with ten or more staff using at least one AI technology in June 2026, which means most were not. DSIT’s adoption research found 75 per cent of adopters reporting productivity gains and 12 per cent reporting higher revenue. PwC found 21 per cent of UK chief executives saying AI had increased revenue in the past year. MIT NANDA found 95 per cent of organisations getting no measurable return from generative AI pilots, and Gartner’s four causes of abandonment are poor data, weak risk controls, escalating costs and unclear business value. None of that says AI does not work. It says most firms buy it before they have a process for it to work on, and that is the thing to avoid.

Count one: retyping and reading. For one week, how many times does someone enter the same details into a second system, or read a document to extract a few fields, or write a summary of something already written? Multiply by the minutes each takes. Count two: waiting. How many enquiries, applications or bookings arrive when nobody answers, and how long until someone does? Count three: decisions without data. How many times a month does someone decide a price, a rota, a follow-up or a purchase from memory when the answer is in a spreadsheet or a system nobody opens? Write the three numbers down. If the first is under ten hours a month, the second is a handful, and the third is rare, the answer is no, for now.

Thresholds are rules of thumb; the baseline in a 90-day measurement is what decides.

A firm whose work is mostly physical, with little text and few enquiries. A firm whose volume is small enough that one person comfortably handles it, where the cost of a tool exceeds the hours it saves. A firm whose data is in such a state that any tool would be answering from wrong records, where the first job is tidying, not AI. A firm in the middle of another change, a new system, a merger, a move, where adding AI adds risk without a stable process to attach it to. And a firm being sold AI by a supplier who has not asked what the process is. In each case the right move is to note the three counts, revisit in six months, and spend nothing.

Three things that cost almost nothing. A paid assistant on a business plan, at about £15 to £19 a seat a month, for whoever writes proposals, letters or policies, with the one-page policy so nobody pastes a client’s data into a personal account. A tidy-up of the one system the firm relies on, so the option is open later. And the three counts kept in the diary for a six-month review, because volume grows, and the firm that measured last time will know when the answer changes.

Pick the process with the largest count and baseline it for two weeks. Decide whether an assistant with a template, a workflow tool or a build is the right shape, using the cost guides to price each. Set an acceptance threshold and a 90-day decision date. Then build the one thing, measure it, and decide the second from the result. That is the whole method, and it is the method the surveys say the successful minority follow. An audit does the measurement and the pricing for a firm that would rather buy it than do it, and is credited against any build.