80% of AI projects fail. Nobody ever measured it.
Jamie Watters
Operational resilience and AI delivery practitioner.
"80% of AI projects fail." You've seen it in pitch decks, board papers, keynote speeches. It's used to justify buying AI software, and to justify not buying it. I went looking for the study behind it. There isn't one.
What there is instead is a footnote, one line long, that nobody seems to have read. I traced the whole citation chain: RAND's 2024 report hedges carefully ("by some estimates"), and that hedge points to footnote 13, which points to a single Fortune magazine profile of an AI vendor's CEO, which cites unnamed survey respondents' opinions in a range of 83 to 92 percent. Not a measurement. Not a rate. Nobody's data. RAND's own interview study, the thing everyone actually cites, never produced a failure rate at all: 65 interviews, no denominator, no rate.
Then, while checking the literature, I found something else. Two separate AI models, asked independently to summarise AI failure research, invented the exact same fake study, with the same fabricated sample size. That study does not exist.
I traced 19 of these figures for a working paper. Seven weren't measurements of anything that had happened. Every source is named, every claim is checked, and the links are in the description.
If there's a number you repeat in your own work, a failure rate, a benchmark, an adoption figure, follow it back one hop. Just one. Then tell me in the comments whether it survived.
Written version, full citation chain: https://jamiewatters.work/journey/80-percent-ai-projects-fail
Say hello, I answer:
If this saved you from repeating a number, you can buy me a coffee: https://buymeacoffee.com/jamiewatters
0:00 The most precious thing in an AI world 0:36 "80% of AI projects fail" — the number everyone quotes 1:10 There's no study, just one footnote 1:28 The citation chain, and RAND's own careful hedge 2:02 Footnote 13, in full 2:29 It's a magazine profile, not a study 3:07 The bottom of the chain: opinions, not a rate 3:56 RAND's own data: 65 interviews, no denominator 4:40 The second number, also unsourced 5:23 Two AI models invented the same fake study 6:56 The real lesson: ask what a number counts 7:32 Four years, one unread footnote 8:37 19 numbers traced, seven were nothing 9:14 Check your own numbers before you repeat them