Closing the AI skills gap: a practical guide for mid-market leaders

The AI skills gap is the distance between the tools a business has bought and what its people can actually do with them. It is tempting to picture this as a technical problem, a shortage of data scientists and engineers. In most mid-market firms it is not that. The gap is operational, and it is about judgement: knowing when to use AI in a real piece of work, how to check what it produces, and when to overrule it.

That reframing changes who you train and how. If you treat the gap as technical, you hire specialists and wait. If you recognise it as an everyday capability, you build it into work most of your people already do.

It is a real constraint, not a soft one

In aibl’s State of UK AI Adoption Survey 2026 (755 UK mid-market leaders), a quarter of leaders named skills gaps as their single biggest source of friction with AI, second only to security and compliance. So this is not a nice-to-have. It is one of the two things most likely to stall the whole effort.

The payoff for closing it is large. Where people are well equipped to use AI and the company is governed, measurable ROI reaches 81 per cent in the survey. Where neither is true, it is 15 per cent. Capability is not the whole story, governance sits right alongside it, but the two together move the return further than any single tool does.

Which capabilities actually matter

Four capabilities do most of the work, and none of them need a technical background.

The first is using AI in a real task rather than a demo: taking an actual report, email, analysis or customer query and getting AI to do a useful first pass. The second is oversight, checking that output against what you already know before it goes anywhere. The third is prompting, which is mostly just asking for what you want clearly and giving the model enough context to answer. The fourth, and the one people skip, is knowing when AI is wrong: spotting the confident answer that does not hold up, the invented figure, the summary that missed the point.

That fourth skill is where business judgement earns its place. A finance manager who knows the numbers will catch an AI error a data scientist never would. This is exactly why the gap is not mainly technical, and why the people who close it are not always the ones you would expect.

Build the capability or buy it in

In the survey, relying mainly on outside consultants came last of every capability approach, at 14 per cent measurable ROI. Bought-in expertise that never transfers to your people tends to leave when the invoice is paid.

A separate study points the same way. The 2026 AI Enablement Services Buyer Survey from 10x Humans and AI Enablement Insider, which asked 100 senior buyers, found a training-led approach returned 73 per cent measurable ROI against 14 per cent for a consultant-only model. Every buyer in that survey agreed that tools alone are not enough without enablement. (These are two different datasets; the two 14 per cent figures are a coincidence, not the same number reported twice.)

None of this means never use consultants. Use them to transfer skills to your people, not to hold the capability on your behalf. When they leave, the capability should stay behind.

Training that actually sticks

The training that works looks different from the training most companies run. A one-off course, everyone in a room for a day, teaches AI in the abstract and fades within a few weeks.

What sticks is role-based. A marketer learns AI on marketing tasks, a finance analyst on finance tasks, both using the tools and data they touch every day. And it happens in the flow of work, applied to a live task this week rather than saved for a training day next quarter. People keep what they use.

How to measure capability

Measuring capability is awkward, because the easy metric, how many people attended training, tells you nothing about whether they can do the job.

Better signals are behavioural. What share of a team uses AI in real work each week. Whether they can point to a task it improved. Whether they catch its mistakes. That last one is the real test. Someone who never questions the output has not learned to use AI, they have learned to trust it, and that is the more dangerous state.

Who to train first

Pick the teams where AI touches the most valuable work and where a mistake costs the most. Usually that is a mix of a customer-facing team and a decision-heavy one like finance or operations. Train a first group properly, let them show results, and use them to pull the rest of the organisation along. Training everyone at once spreads the effort too thin to land anywhere.

A practical first step: choose one team, name the AI tasks that matter in their week, and train them on those specifically over the next month. Then watch whether they use it and whether they catch its mistakes. That tells you more than any company-wide rollout plan.

If you want outside help, the aibl AI Enablement Directory lists vetted UK AI training providers for business, each reviewed by the aibl team.

Frequently asked questions

How much should we spend on AI training?

There is no single right number, but budget it against the teams doing your highest-value work rather than a flat figure per head across everyone. The 10x Humans buyer survey found a training-led approach returned 73 per cent measurable ROI against 14 per cent for buying in consultants alone, so the money tends to go further when it builds your own people’s capability than when it rents someone else’s.

Should we train everyone or specific teams?

Specific teams first. Training everyone at once sounds fair but spreads the effort thin and lands nowhere. Start with the teams where AI touches valuable or risky work, prove it there, then widen out.

How long before we see results?

Weeks for basic usage, longer for judgement. People can be doing useful first-pass work within a month of role-based training. The harder skill, knowing when AI is wrong, builds over a few months of real use. A one-day course will not deliver either on its own.