In 2025 researchers at MIT's Project NANDA examined hundreds of corporate generative Ai pilots. 95% produced no measurable return. Not a small effect. Not a rounding error. Nineteen in every twenty.
That number gets quoted a lot, usually to argue that Ai does not work. Read the study and it says something closer to the opposite. The tools performed as advertised. The projects around them did not.
The number
The failed pilots had things in common. Nobody had named the problem the tool was meant to solve. Nobody owned the outcome. Nobody measured anything before switching it on, so nobody could tell whether it helped.
The 5% that returned money were duller projects. Invoice handling. Enquiry triage. Document drafting. One narrow job, one owner, a number written down before and after.
The study is preliminary and has critics. The pattern it describes is older than the study.
The boring answer
This is why we map before we build. A tool pointed at a named problem, with a number attached, tends to pay for itself. A tool bought because the board read about Ai tends to join the 95%.
Boring wins. It usually does.