The maturity score that flattered the room

Meta description: AI maturity is revealed in real workflows, repeatable results, error correction, and sustained human oversight — not license counts.

Miklos Roth

The maturity assessments I encounter in board packs are confident documents. They score the company on dimensions, place it on a curve against peers, and conclude — in the cases that interest me — that the company is roughly where it should be. Then I watch the same company discover, usually by accident, that a model has been quietly making a class of decisions for months and nobody in leadership can say which decisions those are. The score was high. The reading was wrong. Both overrating and underrating, I have found, come from the same error.

A company may own several sophisticated AI subscriptions and still struggle to produce one reliably usable document. Another may use a smaller set of tools and achieve repeatable improvements in daily work. Counting licenses will not reliably distinguish the two. Neither will a polished demonstration or a survey asking managers how confident they feel about AI.

What do executive teams count when they say they are mature?

The countable things. Licenses deployed, people trained, pilots launched, vendors contracted. These numbers are real and they are also easy, and I have noticed that assessment instruments gravitate toward whatever the company can count without leaving the conference room. What rarely appears in the count is capability: what the organization can actually do differently on Monday because AI exists. Access and capability share a budget line, so they get scored as one thing.

Where do overrating and underrating come from the same error?

For me, the more revealing evidence is operational. I want to see what happens in an actual workflow, how the work is checked, what happens when an output fails, and whether the team becomes more capable as a result. Maturity is not a claim to perfection. It is the ability to keep improving without surrendering human responsibility.

The opposite mistake is also possible. A small team might appear immature because it has few formal AI positions or limited enterprise infrastructure. Yet if its staff can use the tools productively, inspect the results, correct defects, and repeat the process, it may possess more practical capability than the larger organization. Both errors arise from treating access as a substitute for demonstrated performance.

What I can say is that the two failures look like opposites and behave like twins. One company believes it is advanced because it bought early and broadly. Another believes it is behind because it bought cautiously. Both are measuring the shelf. I have watched the cautious company outperform the eager one on the decisions that mattered, and I have watched neither of them find that fact in their assessments.

Why do peer benchmarks stabilize the error?

The false explanation I hear is benchmarking: compare against peers, close the gap, stay competitive. Benchmarking against peers who are all counting the same wrong object does not correct the error, it certifies it. I have seen entire industries arrive at a shared, flattering, incorrect picture of their own maturity, each company’s assessment citing the others as evidence. The curve everyone is placed on was drawn from the same miscount.

What pattern separates access from capability?

Access is visible in procurement. Capability is visible in decisions. The pattern I look for — and this is as far as I go here — is whether anyone can point to a decision that is made differently, by a different route, at a different speed, because the model exists. In companies that are mature in the way that matters, that question has answers. In companies with excellent maturity scores and no capability, the question produces a pause, then a story about a pilot.

I would rather see a complete example of work than a high-level score. Show the input, the relevant data, the AI’s contribution, the human corrections, and the final business result. Show what changed after an error. Show whether people can repeat the result when the task is longer, more complex, or less predictable than a demonstration.

At our agency, AI has increased our content-production capacity across several areas. On one occasion, we sent a client a document of roughly sixty pages. We had reviewed the first thirty pages, which looked strong. We did not examine the remaining pages with the same care. The second half was poor enough that the client rejected the material. The issue was not that AI could never produce the document. The issue was that our own quality-control process had stopped halfway through. We changed the workflow, reviewed the whole document, and adjusted the instructions and iteration process. The revised result met the client’s expectations.

What question replaces the maturity score?

The maturity score survives because it is reportable. It fits a slide, it moves year over year, it gives the board a number to watch. The question that would replace it is less comfortable, because its answer changes the room’s mood rather than its metrics. I have watched leadership teams avoid it for exactly that reason.

The question beneath an AI maturity score is concrete: Can this organization repeatedly produce better work with AI while retaining the human expertise needed to recognize when the result is wrong?

PR excerpt (80–120 words, Miklos’s own words): At our agency, we once sent a client a document of roughly sixty pages. We had reviewed the first thirty, which looked strong; we did not examine the rest with the same care. The second half was poor enough that the client rejected it. The issue was not that AI could never produce the document — it was that our quality-control process had stopped halfway through. Success at the beginning of an output cannot validate the rest. Maturity is the ability to keep improving without surrendering human responsibility.

By Miklos Roth — AI marketing consultant, Budapest. More: his field notes.