How to identify AI use cases in your business

The Office for National Statistics published its first full account of AI use in UK businesses in July, covering late 2023 to June 2026.¹ Use among firms with ten or more employees has roughly tripled, from around 12% to around 35%, while depth has barely moved: the average adopting business used around 1.4 AI technologies at the start of that period and uses around 1.6 now. And when the ONS asked what prevents or delays adoption, the most commonly reported factors were difficulty identifying business use cases, then cost, then a lack of expertise.

That first barrier is a different kind of problem from the other two, because it gives way to method rather than to budget. We run AI audits for UK businesses, and finding and ranking use cases is the first week of every one. What follows is how we do that work, written so you can run it on your own business.

Use cases hide in the work nobody counts.

Nobody’s job description says re-key the same order into a second system, or chase three people for a status update, yet whole afternoons go to exactly that. The candidates worth finding are almost never a department; they are the connective tasks inside one, and four verbs locate most of them: the work people re-key, the work they hand off, the work they chase and the work they summarise.

Our audit findings are mostly examples of those four verbs. At House of Hackney, 20 hours a week of finance reconciliation is now automated, and reconciliation is re-keying with a check attached. One firm we audited was sending 41 sales invoices a month as 41 individually written emails. At Stow Brothers, the admin came to 172 hours a month once someone actually counted it.

There is a fifth place to look that no verb quite covers: the things people search for and fail to find. In our audits, the request to ask a question in plain English and have it answered from the firm’s own documents comes up unprompted more than anything else staff ask for. The tally across seven engagements is in what seven UK AI audits kept finding.

Rank candidates on four numbers.

A found candidate is not yet a use case. The first two numbers to put against it are frequency and time: how often the work happens, and how long one instance takes when somebody times it rather than remembers it. Together they set the ceiling on the return, and a twenty-minute job done every day gives back more than a three-hour job done once a quarter. The daily job is also easier to supervise, because whoever checks the output sees a hundred examples a month instead of four.

The third number is the cost of a wrong answer. A wrong line in an internal summary gets caught at the next meeting, while a wrong figure in a customer contract travels. Error cost decides how much human checking the output will need, and that checking is part of the price of the use case, so high-stakes work has to clear a higher bar before it earns a place on the shortlist.

The fourth is barely a number: does the data the work needs exist, and is anyone allowed to let a system read it? In our audits this factor decides recommendations more often than any model choice. In at least four of the seven engagements, where the data lived or what state it was in shaped the core recommendation, and at one consultancy the honest advice was a document migration before any AI at all; one live programme folder alone held 129 gigabytes across about 45,000 files.

What disqualifies a candidate.

Some candidates leave the table for reasons no score changes. If the process runs differently every time, there is nothing stable to automate, and the fix is process work before it is AI work. If nobody can own the output, meaning no one is prepared to answer for it when it is wrong, the candidate is not ready however strong its numbers look. And if the work is the kind your team would never hand over even where a machine could do it, write that down as a decision rather than leaving it as an instinct.

We put the refusals in writing. In at least two of the seven audits we have run, the deliverable included a list of workflows where we advised against AI, one of those lists eight items long. A shortlist gets more believable when something visibly failed to make it.

What the shortlist looks like before anyone buys.

By the end you want two or three candidates, not ten. Most of what the first pass finds should die in the ranking, and what survives is usually the small set where the case has stopped being a leap of faith. Each survivor should carry a measured hours number and a named owner for the output. It should say where its data lives and what shape that data is in. And the reasoning should be written down, so a colleague, or later a vendor, can challenge it line by line.

That order matters commercially. A shortlist built before the demos means each demo gets judged against your ranked work instead of writing your shortlist for you, and if an off-the-shelf tool clears one of the shortlisted candidates cheaply, the shortlist has done its job. If you have not yet counted the AI your staff already use, or asked them what they would hand to a machine, that comes first, and where to start with AI in your business covers the month before this one.

This is the audit’s first week, done properly.

Everything above is what an AI audit does in its first week, with structured interviews in place of corridor guesses and someone across the table who has counted this work in other firms. The rest of the audit is what makes the shortlist safe to act on: scoring each survivor on payback, data, systems and accountability, then pricing a route for each. Ours is independent, with no stake in recommending a build, and it completes within four weeks. You leave with a written report and a prioritised 12 month roadmap, and the AI Audit and Transformation page sets out the engagement. If you want a reading on where you stand before you talk to anyone, the free AI readiness check asks ten questions and gives you a score out of 100 with suggested next steps.

References

1. Office for National Statistics. (2026). Artificial intelligence in UK businesses: 2023 to 2026, released 20 July 2026, accessed on 31 August 2026, https://www.ons.gov.uk/businessindustryandtrade/business/businessservices/articles/artificialintelligenceinukbusinesses/2023to2026