License count is not adoption.

Somewhere in your last board deck there was a slide about AI adoption, and the number on it was almost certainly a count: seats purchased, licenses assigned, percentage of employees "enabled." Two thousand employees on the enterprise AI plan. Ninety-one percent coverage.

That number measures exactly one thing: what you spent. It says nothing about what you got. Between "has a license" and "works differently because of it" sits the entire return on the investment — and license count is silent about all of it.

The flat middle

Pull the actual usage data behind any enterprise AI rollout and the same distribution appears. A small tail of power users — usually single-digit percent — who have genuinely restructured how they work and account for a wildly disproportionate share of total usage. A large middle who use the tools occasionally and shallowly: the meeting summary, the email polish, the thing they saw a colleague do once. And a silent bottom — routinely a fifth to a third of licensed seats — who signed in during onboarding week and never came back.

Averages hide this completely. "Weekly active usage is up" can be true while ninety percent of the value accrues to five percent of the people. And here's the strategic point most reporting misses: the power users didn't need your program, and the silent bottom isn't where the money is either. The return on the entire investment lives in the middle — the majority who use the tools at a fraction of their depth. Moving that group from shallow, occasional use to confident delegation is worth more than anything you can do for either tail. It's also the group about which login metrics tell you precisely nothing, because they are logging in. They're just not getting much out of it.

Measure capability, not logins

Why is the middle stuck? Talk to them and it's rarely access, and rarely attitude. It's specific missing skills: not knowing which parts of their role can be delegated to a model, not trusting themselves to verify the output, not knowing what's safe to paste. One bad early experience — a confident wrong answer they almost forwarded — and they quietly retreat to the meeting-summary use case forever.

None of that appears in telemetry. Usage data can show you that someone stopped at shallow use; it cannot show you why, or what they'd need to go deeper. Those are questions about judgment and skill, and the only way to see them is to assess them — role by role, because the capability that matters for an account executive (drafting, verification against the customer record, tone risk) is different from what matters for an engineer or a financial analyst.

A real capability baseline asks, per role: can this person identify delegable work, construct a decent request, verify the output against ground truth, and recognize what must not go into the tool at all? Score it on a handful of axes, and suddenly the flat middle isn't a blob — it's a map. This team is capable but doesn't trust the output. That team pastes things it shouldn't. This function never got past summarization because nobody showed them the workflow that matters for their job.

Enablement follows measurement

With a map, enablement spend stops being generic and starts compounding:

• Training lands where the gap is. The team that can't verify output gets verification technique for their domain, not another generic prompt-writing webinar — the current default, which power users don't need and the silent bottom won't attend.
• Power users become multipliers. You know exactly who they are and — with role-level data — exactly which stuck team each one's workflows would transfer to. Structured pairing beats another all-hands demo by an order of magnitude.
• Verification becomes an explicit skill. The single biggest unlock for the cautious middle is confidence in checking AI output. Treat it like a competency: taught, assessed, expected. The organizations that do this get the productivity and fewer of the incidents that come from misplaced trust.
• Re-measurement closes the loop. Baseline, intervene, re-assess. Now the board slide shows capability moving quarter over quarter — a delta, not a seat count — and the enablement budget can defend itself with evidence.

There's a governance dividend, too. The AI-literacy expectations now arriving in regulation — the EU AI Act expects organizations to ensure staff operating AI systems have adequate competence — are trivially satisfied by a program that already measures and develops capability by role, and awkwardly satisfied by a procurement receipt.

The seat-count slide felt good in the year everyone was racing to have AI. That year is over. The question boards are learning to ask next is the one that separates spend from return: not how many people have the tools — how many people are actually better at their jobs because of them, and how do you know. It pays to have an answer before the question arrives.