Most UK SMEs Are Buying AI. Few Are Profiting From It.

Only 31% of UK businesses investing in AI are reporting a positive return. That figure comes from the UK Government's own AI Adoption Research — not a sceptic's think piece or a vendor survey with a narrative to push.

For a 5–20-person consultancy being pitched AI tools every other week, it deserves sitting with. Seven in ten businesses putting money into AI are seeing no measurable financial return. That isn't an indictment of AI — the same tools are producing results for the 31%. It's an indictment of how most of the money is being spent.

The 71% problem

The sharpest number in the UK Government's own AI Adoption Research isn't the adoption rate. It's the 71%.

Seventy-one per cent of UK businesses said they hadn't identified a clear use for AI in their organisation — UK Government AI Adoption Research, via Spicy Advisory. The skills gap gets more coverage (60% of businesses cite it as the primary blocker). The use-case gap is the more interesting problem.

Skills you can hire for. A clear use case requires you to think carefully about your own operations — where the manual work is, what inputs it runs on, what output it's supposed to produce. That's not a skills problem. It's a diagnosis problem.

Most SMEs shop for AI before they've diagnosed the workflow. They see a demo — AI that qualifies leads, answers calls, summarises documents — and they buy in, hoping it'll find a home. Sometimes it does. Mostly it doesn't.

What's working

The SMEs seeing returns are doing something structurally different. They're starting with a specific, bounded workflow — not a category of capability.

Lead triage is the clearest example. A five-person professional services firm gets 20 inbound enquiries a week, some qualified, most not. The AI reads each form submission, scores it against a defined set of criteria, routes the top five to a human, puts the rest into a nurture sequence. No platform purchase. One workflow, structured inputs, a measurable output, conversion lift visible within 60 days.

The same shape holds across adjacent builds. Voice AI for out-of-hours calls: defined questions, structured transcript, booking or escalation as the output. Invoice processing: identical. The specific details change; the underlying structure doesn't.

The five ops processes most likely to follow this shape — lead intake, onboarding, invoicing, reporting, content scheduling — share three features: clear inputs, a defined output, and a human review step for edge cases. That shape automates reliably. Workflows with fuzzy inputs and shifting definitions don't — and no amount of tooling changes that.

The deployment gap

Sharp surveyed 2,500 SME leaders across 10 European markets earlier this year. Sixty-four per cent said AI is fully embedded in their business. Fifty-five per cent said they're concerned they're not using it as effectively as they could be.

Hold both numbers together. A meaningful share of that 64% falls into both camps simultaneously — deployed, but not delivering.

The reasons are predictable. Thirty-five per cent of leaders say their staff worry about lacking the technical skills to use it well. Thirty-four per cent report a lack of trust in AI-generated outputs — which means the automation runs but gets overridden or quietly ignored anyway. The licence is live, the integration is done, the team broadly knows it exists. But no one is accountable for what it produces, and there's no benchmark to compare against.

That's where the budget disappears in year one. Not through a dramatic failure. Through quiet non-use.

The question before the next purchase

Before the next AI investment, one question is worth asking: what specifically will this automate, and how will I know if it's working?

Not "can AI help with this?" — almost certainly yes, at some level. The narrower question is whether the workflow has structured inputs, a defined output, and a measurable baseline to compare against.

If yes to all three: build it. The cost structure for a scoped SME automation is simpler than vendors make it sound — a single focused build typically pays back within 60–90 days when the use case is correctly scoped.

If no to any of them: don't buy yet. Do the diagnosis first. It takes an afternoon, costs nothing, and is the only thing separating the 31% who see returns from the 69% who don't.