The AI Skills Gap Is a Specification Problem First

Where does the gap number come from?

Almost always from a survey asking managers or employees to rate a capability that the survey never defined. The respondent supplies their own definition, which drifts with how much AI coverage they read that month, and the aggregate is a mood rather than a measurement. That is not a reason to dismiss it. A widespread feeling that the organisation is behind is worth investigating, but it is where the work starts.

The tell is that the same question produces incompatible answers depending on who is counted. A Federal Reserve note comparing the three main US adoption surveys found they disagree by construction: about 18% of firms had adopted AI at the end of 2025, about 41% of the workforce reported using generative AI at work in November 2025, and an employment-weighted estimate put 78% of the labor force at firms that have adopted AI ¹. Those are three different units of analysis and none of them converts into another. The 78% describes whose employer has adopted AI, which is a different question from who uses it.

So the first question to ask about any gap figure is what it counted. A firm-level number and a worker-level number for the same period can differ by twenty points and both be correct, because most firms are small and most workers are not at small firms. The Census Bureau's biweekly survey put AI use among US businesses at 19.8% in May 2026, against 37% of firms with at least 250 employees ².

None of that answers whether your company has a gap. It establishes something more useful: a published gap figure is not evidence about your team, and there is no external number you can substitute for looking.

Why an average is the wrong shape for this

Because it collapses the distribution, and the distribution is the finding. Inside a company the same ambiguity that splits the published surveys operates on your own numbers: an internal survey and a manager's estimate of one team can differ by half the headcount without either party being careless. An average across that spread describes nobody on the team.

One study measured that spread directly. In the staggered rollout of an AI assistant to 5,179 customer support agents at one software firm, issues resolved per hour rose 14% on average, with a 34% improvement for novice and low-skilled agents and almost nothing for the experienced ones ³. One firm, one occupation, a pre-ChatGPT-era assistant on problems with correct answers: the numbers do not transfer to your team. The shape does. An average movement of 14% concealed two groups moving completely differently, and a programme designed against that average would have reached neither.

Aggregate effects are also smaller than the narrative suggests, which is worth knowing before the budget conversation. In a nationally representative US survey, people using generative AI at work reported mean time savings of 5.4% of their work hours, which works out to 1.4% across all workers including non-users ⁴. Self-reported, late 2024, and time saved is not value created unless the freed hours go somewhere. It is the honest counterweight to a deck implying the whole company is a step behind.

What survives every caveat here is a method. Count something specific, in your own company, against a definition you wrote down first.

Define the behavior for two or three roles first

Pick the roles where AI work is already happening and write, in a paragraph each, what good looks like in that job specifically. Not a competency framework and not a policy: four or five sentences naming decisions someone could be seen making. What goes to the assistant, what context it gets, what gets checked and against what, what stays with the person. Two weeks of calendar time, mostly other people's.

The drafting has to be done with the people doing the job, and it goes wrong the same way every time when it is not. A definition written in a leadership offsite describes work as leadership imagines it, so the resulting gap is between the team and an imagined role. The correction is cheap: two practitioners per role, an hour each, asked what they actually hand over and what they would never hand over. Their disagreements are the specification.

Then sample the work itself. Take five recent pieces of output per role and read them against the paragraph. Was anything traced to a source. Did anyone keep back part of the task and say why. This is slower than sending a form, and it is the only step in the exercise that produces evidence at all. Self-report and observed work are known to diverge in both directions.

Doing this without a survey nobody answers honestly is its own small problem, worked through in how to find out what the team can already do with AI. If the paragraphs are hard to start, four levels of AI skill written as observable behavior gives a ladder to borrow and edit.

What to do with whatever gap is left

Sort what you found into three piles. People already working the way the definition describes, who need naming and a role in teaching the rest. People a fortnight of supervised practice would move, which is a training item with a date. And capability nobody has at all, the only genuine hiring question. That third pile is usually the smallest, and it is the only one the gap figure claimed to measure.

The first pile is worth dwelling on, because it is the one specification reveals and surveys hide. Teams routinely have two or three people whose verification habits are exactly what the company is about to pay a vendor to install, and nobody had noticed because the behavior had no name and no line in a review. Naming it costs nothing and does most of the work: the practice spreads by imitation once it is visible and legible, and the internal teacher is more credible than the external course.

The second pile sets the size of any programme you buy, which is the number the whole exercise exists to produce. A programme sized to the second pile is a fraction of one sized to the whole company, and it is aimed at people who have a specific thing to learn rather than at everyone.

The third pile is a hiring decision with a real bar attached, and by this point the requirement writes itself out of the definition you already drafted. Whether to fill it by hiring or by training is the standing question, and whether to hire for AI skills or train the team you have takes it up. Which roles should have been in the exercise at all is answered in which roles actually need AI skills right now.

See the benchmarks >