AI maturity is three instruments, not one number

Every leadership team wants a single number for where they are on their AI journey, something tidy enough to put on a slide. When they give you that number, notice who’s producing it. It’s nearly always the person who’s best at AI, because they’re the one trusted with the question, and they answer it from inside their own fluency. So it comes back higher than the company has earned. There’s no dishonesty in that. The person fluent enough to be asked is the same one whose own skill makes the whole place look further along than it is.

A single score is also an average, and an average hides the spread it’s drawn from. Some of that spread runs between departments. Engineering might be running a real practice while accounting has barely opened the box, and a company-wide figure splits the difference and matches neither. More of it hides inside a single team, where one person works at a high level and everyone around them is still asking AI to clean up their emails. Average that and you get a middling number that describes nobody, reading as steady progress while almost all the real capability sits with one person. Ask what’s left if they take a month’s leave or move on, and you find out quickly.

For anyone with an oversight role, that’s where the number turns dangerous. Capability concentrated in one or two people is a continuity risk long before it’s an advantage, and the score does nothing to flag it.

The three instruments

A single figure can’t hold the thing that matters, which is the distance between how good your best people are and how good your organisation is. You need three readings.

The first is the organisational ladder we published earlier this year, five rungs from exploring up to directing, measuring how well a company absorbs and compounds what AI makes possible rather than what it’s bought. You run it per department rather than once for the whole company, since they rarely move at the same speed. Even then it reads the aggregate on purpose, and that’s exactly why it can’t stand alone, because two or three strong people are easily mistaken for a practice. That piece called the trap the lone wolf problem, and it closed on three questions I still reach for, about who you’d lose if your two strongest people left, whether a new hire onboards into the practice or only the tools, and whether anyone’s measuring if it’s working. Those three are the second instrument, the one you can run in a single conversation.

The third, the one that’s been missing, reads where each person actually sits. It’s the reason a single number fails, so it’s worth doing properly.

The instrument that was missing

“Good with AI” is almost useless as a description of a person, and the individual ladder is there to fix that. It runs across five rungs. At task assistance you ask for a draft or a summary and take what comes back. At workflow support AI is folded into the work you repeat, though you’re still checking it more than trusting it. The rung that matters is agentic, where you hand over a defined piece of work and review the result instead of producing it yourself. That’s where the payoff really starts. Above it, systems orchestration is running several agents at once and building the checks that keep them honest. At the top sits champion, held back on purpose, because by that rung the measure has moved off your own output and onto what the people around you can do.

Nobody sits on one rung across the board. The same person will be orchestrating agents in one part of the job and asking for a tidied-up email in the next, so a single label just buries the picture. We score seven dimensions of the work on their own:

whether AI turns up where it actually helps
how much they trust what it gives back, and how well they check it
how they handle tool and data safety
whether their skill travels or dies with one tool
how well they manage the context they feed it
when they drive it step by step and when they let it run
how they hold a whole workflow together

Fluent, to us, means holding agentic across all seven at once, and almost nobody does that evenly. The gaps between the dimensions are usually the useful part.

Read them apart

Keeping the two ladders separate is the whole point. One reads the aggregate, the other reads people one at a time, and the moment worth watching is when they disagree. You can have an individual ladder full of people working at a high level while the organisation sits near the bottom. That looks like a contradiction until you see what it is, the lone wolf problem showing up as a measurement. You only catch it because you kept the readings apart. Roll them into a single number and the disagreement vanishes, and that was the finding.

Stewardship closes the gap

Reading the instruments tells you where you are. It doesn’t move you. The move that matters for most companies is the one from owning the tools to running a practice, and in our experience the thing that produces it is stewardship.

Stewardship is coaching done on someone’s real work, while they do it. Someone brings a live, unfinished piece of their own, and a steward sits alongside, watches how AI does or doesn’t show up in the actual flow, and coaches as they go, asking why they did it that way and what they expected rather than taking over. What makes it lift the organisation and not just the person is where the learning ends up. A workshop lifts the person in the room for an afternoon. A stewardship session turns what they worked out into something the company keeps after they’ve moved on, the fix from a real bug becoming a line in an instruction file the AI reads on every future task, one developer’s hand-off becoming the way everyone hands off. Do that enough and the practice holds its shape when any single person is away, which is the actual distance between the second rung and the third.

A few things we’ve learned running them. They’re worth far more mid-implementation than mid-planning, when there’s live work to watch instead of a plan to discuss. People need to help shape the approach and then be left to run it, not handed a finished plan to nod at. And an observer with no stated role reads as judgement rather than support, which quietly wrecks the session. None of it is glamorous, it’s the dull, repeatable ground the compounding is built on.

For the finance side of the table, that same concentration belongs on a risk register, even though it never shows up on the maturity slide. The rest doesn’t need a number at all. Read your best people and your organisation as two separate things, watch where they disagree, and treat the gap as the risk it is. Then do the slow work of closing it, which is the part no number was ever going to do for you.

Daniel van der Merwe is the Technical Director of Rokkit200, an AI transformation agency that works with engineering and product organisations to build compounding AI practices.