Your AI status report is not a measurement

Companies rate their own AI progress higher than a scoring does, and the error runs one way: always optimistic. If your answer to "how is AI going here" is a sentence rather than a number, you have a status report, not a measurement — and the two disagree in a known direction.

What the benchmark found

IDC's 2026 AI MaturityScape Benchmark (summarised on IDC's blog on 11 August 2026 by Andrea Siviero and Xiao Liu) scored 1,900 organisations across 20 markets against a five-stage maturity model, from an ad-hoc "AI scramble" through opportunistic, managed and advanced to optimised.

• 3.1% reached the optimised stage.
• 61.3% were still in the two least-mature stages.
• 12.8% sat in the two most advanced.
• The mean score moved from 2.39 to 2.43 in a year.

And the sentence that does the real work, in IDC's own framing: of the organisations that identified as thrivers, only about one in six reached the managed or optimised stages when scored against IDC's methodology.

Two caveats worth stating before anyone repeats the figure. This is IDC's blog summary of a paywalled benchmark — the underlying data tables are not public, so "one in six" is IDC's framing rather than something independently re-derived. And no field dates or market split are disclosed, so it must not be presented as European data. On the other hand, IDC is an analyst house rather than a vendor selling the remedy, which makes it a less conflicted source than most numbers in this space.

The flat mean is the second half of the story

A move from 2.39 to 2.43 is four hundredths of a point across a year in which AI budgets went up almost everywhere. Whatever that spend bought, it did not buy movement along a maturity curve. That is a harder question for a CFO than the headline gap, and a more useful one.

How do you know if your AI adoption is actually working?

You know when you can produce, on demand, four numbers: how many people use the tool weekly on real work, what it costs per active user, how many hours it gives back on a named workflow, and how many licences have had zero activity in 90 days. If those take a week to assemble, they are not being managed — they are being described.

• Weekly active users as a percentage of paid seats. The denominator is seats you pay for, not seats you hoped for.
• Cost per active user. The honest ROI denominator, and the one most finance teams cannot produce on request.
• Hours returned on one named workflow. Not "productivity." A specific process, before and after, in hours.
• Licences with zero activity in 90 days. The waste line, and the fastest euro in the building.

Why the gap runs one way

Nobody in a reporting line is rewarded for saying "our AI programme is at stage two of five." The information passing upward is filtered by people whose standing depends on the answer, at every hop. That is not dishonesty; it is the ordinary physics of self-assessment, and it is exactly why a score and a status update diverge.

Which is also why the fix is structural rather than motivational. You do not get a truer answer by asking more sincerely. You get it by replacing the question with an instrument.

What to do with it

Measure before you buy the next tier. The reflex when AI is not paying back is to buy more of it — a higher licence tier, another tool, a bigger pilot. If the current tier is sitting at a few percent weekly usage, the next one will sit there too, at a higher price. Establish the four numbers above on what you already own. Most organisations discover the gap is not capability. It is that nobody was ever shown how the thing fits their Tuesday.

A company's account of how its AI is going is systematically better than its score — and always in the same direction.