The sentence everyone quotes is not in the report
You have heard the number. Ninety-five percent of AI pilots fail. It has been in board packs, keynotes and comment threads since last summer. That sentence does not appear in the report it comes from.
The report actually says that 95% of organizations are getting zero return. Organizations are not pilots. A company running twenty projects counts once in that number. A company running one counts once. It is a headcount of companies, and it says nothing about whether any particular piece of work succeeded — which is the exact thing people quote it to prove.
The same report contains an eighty-three percent success rate
Every failure figure in that report is scoped to one narrow category: custom-built or vendor-sold enterprise AI systems. The same report measures the other category — ordinary AI assistants, the kind most people actually use — and puts pilot-to-implementation there at around 83%. An eighty-three percent success rate and a ninety-five percent failure rate, in one document, four pages apart, measuring different things.
The report is candid about what its numbers rest on. Not company reporting — conversations. Its own limitations note reads: these figures are directionally accurate based on individual interviews rather than official company reporting. Underneath the whole thing sit 153 senior leaders surveyed at four industry conferences and 52 interviews, and an appendix that adds the six-month observation window is short, and may be undercounting successes.
The number nobody forwards
A survey of 3,120 technology decision-makers a few months later found 46% of AI proof-of-concepts already in everyday use. That one was paid for by a company that sells AI equipment and its conclusion is that AI is working and you should buy more. Which is exactly as self-interested as the first one, in the opposite direction — the pessimistic report came out of a group building the very fix it recommends.
Nobody made anything up. A researcher wrote a precise sentence about organizations, somebody summarizing it reached for the punchier noun, and the punchier version travelled because it fit in a post. Every step was reasonable. What comes out the other end is a number that sounds like it is about your work and was never measuring anything of the kind.
A four-second habit that changes what you do next
When a number lands in front of you and it is about to change what you do, read the word straight after "of." Of organizations. Of pilots. Of the people we surveyed. Of the tools that were already working. Those are four different claims and they routinely wear the same percentage. If the word is not there at all, that is your answer too.
Then ask who paid, and whether the same source contains a number pointing the other way. It takes about twenty minutes to establish all of the above, and it is the difference between most AI fails, so let's wait and most companies have not got a return yet, which is a statement about companies. The distinction matters most when the statistic is telling you to do nothing — because that is the cheapest possible substitute for finding out.
The alternative to a statistic about other people is twenty minutes on where you actually stand: Find out where you actually stand →
Keep reading: The most common use is not the best one · The AI adoption incentives paradox
This is the thinking behind a Trusted AI Culture.
Sources: MIT NANDA (Project NANDA), "The GenAI Divide: State of AI in Business 2025," July 2025 — a preliminary, non-peer-reviewed report, read via a third-party mirror as it is not hosted on the Project NANDA site. Lenovo, "CIO Playbook 2026: The Race for Enterprise AI," January 2026, research by IDC, commissioned by Lenovo. Both are self-interested, pointing opposite ways, which is the point.
Written with AI in the loop: my idea, AI drafted and sharpened, my judgment on the way out. Every word is mine to stand behind. — Darren