How to measure AI return in 90 days in a UK business, with numbers you can defend.

The UK figures are consistent. DSIT’s AI Adoption Research, a 3,500-business study published in February 2026, found that 75 per cent of adopters reported improved workforce productivity, 12 per cent reported higher revenue and 77 per cent had seen no revenue change. PwC’s survey of UK chief executives in January 2026 found 21 per cent saying AI had increased revenue in the previous twelve months, with 74 per cent reporting little or no change. MIT NANDA’s 2025 study of enterprise deployments found 95 per cent of organisations getting no measurable return from generative AI, and Gartner predicted that at least 30 per cent of generative AI projects would be abandoned after proof of concept, naming unclear business value as one of four causes.

Read together, those figures describe a measurement problem as much as a technology one. Productivity gains are real and widely reported; revenue effects are rare because most SME automation removes cost rather than adding sales; and the projects that get abandoned are the ones where nobody wrote down, before the build, what would count as working. The method below exists to write that down.

Weeks one and two are the baseline, and they happen before any supplier is chosen. Pick one process with real volume: reading applications, chasing invoices, answering the enquiry line, keying supplier documents. Count it. How many units a week, how many hours each unit takes across everyone who touches it, how long from arrival to first action, how often it has to be redone, and what it costs per unit at a loaded hourly rate. Write the four numbers down and get two people to agree them.

Weeks three to ten are the automation running, with a person still deciding anything that affects a customer or an individual. Measure the same four numbers every week, and add the fifth: the AI running cost per unit, from the provider invoices, which any competent build reports per feature from day one. Week eleven is the comparison and week twelve is the decision: extend, change or stop. A stop is a legitimate outcome and cheaper than a slow fade.

Define all four before the build. A metric added afterwards is an argument, not a measurement.

A firm receives 400 supplier invoices a month, each keyed and matched by hand at an average of nine minutes across the people who touch it, at a loaded cost of £28 an hour. Baseline: 60 hours a month, £1,680, with a rework rate of one in twenty and an average of three days from arrival to posting. The automation reads each invoice into the ledger, matches it to the order and flags exceptions for a person. After eight weeks: 14 hours a month of human time on exceptions and review, £392, rework one in fifty, same-day posting, and an AI running cost of about £40 a month at published prices for a small model on 400 documents of a few thousand tokens each. Monthly saving £1,248; annualised about £15,000 against a first implementation at £4,000 to £15,000. Payback inside the first year on one process, with the evidence gathered before anyone was asked to believe it.

The numbers in that example are illustrative. The structure is not: the saving is hours times rate, the cost is build plus running cost, and both sides were measured rather than estimated.

Estimating the baseline from memory instead of measuring it, which produces a saving that finance rejects in the first meeting. Counting the time of the person who used to do the task without asking what they do now, which is fine if it is billable work and an argument if it is not. Leaving the running cost out, which is how a saving turns into a bill. Measuring a company-wide effect from a single-process pilot. And declaring the pilot a success at week four because the demo was impressive, which is the road to the abandoned projects in the surveys above.

Extend if the four metrics held and the cost per unit fell by enough to pay for the build within a year. Change if the quality metric moved the wrong way, which usually means the exceptions path needs work rather than the model. Stop if the volume was never there, and write down why, so the next idea is measured against it. Then pick the second process, and run the same twelve weeks. A firm that does this four times in a year has a portfolio of measured returns rather than an AI strategy, and it is a better position to be in.