Every company will say yes. The only measurement that means anything is where AI enters a decision — and whether anyone checks it when it disagrees with the boss.
Ask a leadership team how AI-driven their company is and you will get a number between seven and nine. Ask the same team to name the last decision an AI output actually changed, and the room goes quiet for a moment that tells you more than the number did.
That silence is not hypocrisy. It is the normal gap between what an organisation says, what it decides and what it does — three separate products, as the Swedish organisation theorist Nils Brunsson put it, that a company under pressure needs to keep apart. "We are AI-driven" belongs to the first product. We are interested in the other two.
Skip the scale from zero to ten. Ask these.
Where does AI enter? Take the last decision you were part of where AI played a role. Did it help frame the question — or did it write up the rationale for a choice that was already made? Framing is upstream of the decision. Writing the rationale is downstream of it. Same tool, opposite meaning. Our expectation, which we've published so it can be held against us, is that in most companies the honest answer sits downstream.
When it agrees with you, do you check it? Now: when it contradicts you? If those two answers differ, you have found the most important fact about AI in your company. Verifying disagreement harder than agreement is human — motivated reasoning was documented decades before language models — but a tool that is only audited when it's inconvenient is not a tool. It is a witness you call when you need one.
Who can say "I just see it this way"? In every company someone is allowed to decide without numbers and be followed. Founder only, any executive, recognised experts, almost nobody. That answer is a map of power, and it predicts what will happen to an AI output better than any adoption metric. If AI can be overruled by a feeling from the right chair, then AI is advisory exactly where it doesn't matter and decorative where it does.
Where does AI make the final call with nobody looking? Pricing, targeting, support routing, fraud — the answers are usually operational, frequent, reversible. Strategy, hiring, product bets — never. That split is not an accident. Companies delegate where being wrong is cheap and keep the ritual of human sign-off where being wrong is expensive. Which raises the uncomfortable follow-up: in the expensive cases, is the human judging the AI, or is the AI providing cover for the human?
Sociologists who study organisations have a name for structures adopted because they look like what a serious company has: rationalised myths. Total quality management was one. "Data-driven" was the next. The pattern is that the structure gets adopted for legitimacy, quietly decoupled from the real work, and everyone agrees not to inspect the seam.
AI is arriving at the moment the seam is cheapest to hide. A generative model produces the artefacts of reasoning — the analysis, the summary, the well-argued memo — for almost nothing. The performance of rationality gets richer. Whether the decision gets better is a separate question that the performance is designed not to raise.
So the measurement worth doing is not "how much AI" but where, checked by whom, overruled by whom, and blamed on what. "The model recommended it" is already appearing in post-mortems as a place where accountability goes to disappear.
The Institute's first study reconstructs real business decisions from both sides — the person who made the call and the person who produced the evidence — and codes exactly these four things: the stage where AI entered, whether verification depended on agreement, who was exempt from evidence, and what happened when the output contradicted the preference. We take any company willing to take part. We also seek out the ones that say "AI-driven" in public, because they are the natural test of the gap between saying and doing.
We don't have the numbers yet. That is the point of a study. What we have is a four-minute diagnostic that asks you the questions above about one real decision — and gives it back to you as a mirror, held up against the pattern of everyone else who answered.
Four minutes. One decision. No scale from zero to ten.
Four minutes. One real decision. Take the diagnostic →
Institute for Human Reasoning is a research and education organisation strengthening human reasoning for consequential decisions in a technology-shaped world.
Seventy years of research on evidence and decisions, in six findings that still embarrass every company calling itself data-driven.
ReadA founder's note on what the Institute for Human Reasoning is for, what it refuses to be, and what the word "institute" obliges us to do.
Read