Choosing your first AI use case

The first AI project in a business is never just a technical choice. It's a credibility decision. Get it right and the organisation leans in — people see the point, trust grows, and the second project is easier to fund. Get it wrong and the opposite happens: the pilot drags, the savings never show up, and "AI" quietly becomes a thing that didn't work here. That memory lingers for years.

The pattern is well documented. S&P Global Market Intelligence found that in 2025 some 42% of companies abandoned most of their AI initiatives, up sharply from 17% the year before, with roughly 46% of proofs of concept scrapped before they ever reached production. McKinsey's global survey tells the same story from the other end: 88% of organisations now use AI somewhere, yet only 7% have it genuinely scaled across the business. A great many first projects are stalling. Choosing well is how you avoid being one of them.

Start where the work is repetitive and the stakes are low

The best first use case is almost boring. You want high-volume, rules-based work where the cost of an occasional error is small and easily caught. Repetition gives AI something to actually save — a task done five times a day is worth automating; a task done twice a year is not. Low stakes mean you can let the system run, learn where it slips, and fix it without anything important breaking.

This is also where the UK's own evidence points. Government research by DSIT found that the single most-cited barrier to adoption wasn't cost or technology — 71% of firms simply hadn't identified a clear use for AI, with a shortage of skills the next blocker at 60%. The constraint is direction, not capability. A well-chosen, low-drama first task is the cheapest way to remove it.

Pick something you can measure

If you can't say what success looks like in a number before you build, you won't be able to prove it afterwards — and an unprovable win is no win at all. Define the metric first: hours saved per week, response time cut from days to hours, error rate down by a fixed percentage, throughput per person up. Write down today's baseline, then commit to checking the same number in a few weeks. The discipline is simple and it changes everything: it forces you to choose a use case where the value is visible, not vague.

Avoid the glamour projects

The temptation is always the showpiece: the customer-facing chatbot, the autonomous decision-maker, the thing you could demo to the board. These make terrible first projects. They are customer-facing, so mistakes are public. They are high-stakes, so the cost of error is real. And they are ambiguous, so "good enough" is hard to define and harder to defend. You end up with a science project — long, expensive, and impossible to call a clear success. Save the ambitious work for project three, once the organisation already believes.

Score your candidates

When you have a few options on the table, stop arguing about them in the abstract and plot them. Two axes do the job: business value up the side, effort to deliver along the bottom. The top-left quadrant — high value, low effort — is where your first use case lives. Everything else can wait.

The exact placements don't matter as much as the conversation they force. The moment you have to position the glamorous chatbot bottom-right — high effort, uncertain value — against a quiet "draft the replies" task top-left, the right first move usually becomes obvious to everyone in the room.

A shortlist most SMEs can start with

You rarely have to invent the first use case from scratch. Across the small and mid-sized businesses we work with, the same handful keep proving themselves — repetitive, measurable, and low-risk by nature:

• Drafting routine replies — first-draft responses to common emails and enquiries, with a person reviewing before send.
• Summarising documents — turning long reports, contracts or call notes into a tight, readable brief.
• Moving data between systems — the copy-paste re-keying that quietly eats hours and breeds errors.
• First-line triage — sorting and routing incoming requests so the right person sees the right thing faster.
• Reporting — pulling scattered numbers into a regular snapshot someone will actually read.

Any one of these can be stood up quickly, measured honestly, and either kept or quietly retired within weeks. That speed is the whole point. The first use case isn't there to transform the business — it's there to prove the business can do this. Pick the one you can win, win it visibly, and the harder projects get a lot easier to start.

Sources: S&P Global Market Intelligence, Voice of the Enterprise (2025); McKinsey, The State of AI (2025); UK Government DSIT AI Adoption Research, conducted by IFF Research and Technopolis Group (2025). Figures vary by how "AI use" is defined.