One AI framework, twenty local realities: adapt without losing comparability

The request almost always arrives in the same order. First: we want to know where we stand on AI, everywhere. Then, a few weeks later, when the first assessments come back: this is not comparable, our subsidiaries do not do the same job. Both statements are true at the same time, and that is the whole problem.

Group management needs a number that reads across entities. A local manager needs a diagnostic that speaks about their own ground. Answering one by sacrificing the other is easy, and the two ways of getting it wrong are symmetrical.

The first is to impose one framework, written at headquarters, identical everywhere. It produces scores that are perfectly comparable and perfectly useless: the subsidiary with no dedicated legal team answers no to a governance question that does not apply to it, and gets penalized for an organization that is the right one at its size. After two cycles, nobody fills it in seriously.

The second is to let every entity adapt the framework. Each then gets an accurate diagnostic, and the group gets twenty diagnostics it cannot add up. The dashboard lines up scores side by side and implies they compare, which is more dangerous than showing nothing at all.

It is worth naming precisely what varies, because the list is shorter than people expect. Regulation first: an entity established in the Union applies the European AI regulation, an entity outside it only does so when placing systems on the European market. Starting point next: a team of three running experiments is not in the same place as an IT department of two hundred. Use cases last: spotting a defect on a production line and drafting customer replies do not raise the same control questions.

A large part of the framework, on the other hand, does not vary. Knowing which AI systems run in the organization, who answers for them, what data they rely on, how their quality is measured, and what happens when they get it wrong: those questions land identically in Hamburg, Lyon and Porto. They are the core.

Hence the only construction that holds both ends. A common layer, decided once, identical for every entity, carrying comparability. A local layer, added by each entity, carrying relevance. Core scores compare, consolidate, and feed group steering. Local layer scores do not compare, and nobody asks them to.

One rule keeps the whole thing honest: the local layer adds topics, it never removes any from the core. An entity that finds a core question irrelevant does not delete it. It answers, and documents why it sits where it sits. That is often where the most useful information for the group turns out to be.

A second rule, less obvious: the core must be short. The bigger it grows, the stronger the temptation to work around it, and the harder entities push to get their own specifics into it, which empties it of meaning. A core you can hold in thirty questions beats an exhaustive one nobody fills in twice.

The instinct is to assume comparison exists to rank. That is the least interesting reading, and the one that makes teams defensive. Comparison exists first to spot where one entity has already solved what another is facing, then to move the answer across. A subsidiary that has built a register of its AI systems has produced a transferable asset. The score is how you find it, not how you reward it.

And the most useful comparison is not between entities, it is against yourself. A score set against last year's, against the target you had set, and against the action plan you had written says something no ranking says: whether the money committed produced the effect expected. That is the only measure an investment committee cares about.

The entry point depends on which question is pressing. If the deadline is regulatory, the European AI framework sets the scope and the calendar, and the diagnostic tells you what is already in place. If the question is organizational, an AI management system framework structures governance over time. If the question is more open, an AI maturity diagnostic simply says where the organization stands before committing to anything.

In all three cases the move is the same: measure what exists, name the target, price the gap, then work through it. The framework is not the goal, it is the measuring instrument. What counts is what you do next, and being able to do it again next year under the same rule.