Skip to content
← All articles

On This Day in AI · 18 August 2026

95% of AI Projects Fail. Here Is What That Number Counted.

On this day, in 2025, a report came out saying that 95% of company AI projects were producing nothing at all, and within two days it was being blamed for a slide in tech stocks.

August 18, 2025: a report from MIT saying 95% of company AI projects deliver no measurable return goes round the world

August 20, 2025: tech stocks slide, with the report and Sam Altman’s warning about overexcited AI investors both named in the coverage

August 2026: a year on, the 95% is still the number people quote, and there is now a small industry selling you a plan to be in the other five

On this day, in 2025, a report came out saying that 95% of company AI projects were producing nothing at all, and within two days it was being blamed for a slide in tech stocks.

Back then it felt like the moment somebody finally said it out loud. If you had been quietly wondering whether all of this was an expensive fashion, here was MIT with a number. If you had already spent money on AI, it read like a diagnosis. Either way it was the first AI statistic in years that ordinary business owners repeated to each other, and it is still doing the rounds now, usually with nobody knowing where it came from.

Here is what it actually counted. The report was called The GenAI Divide, and it looked at large organisations building their own custom AI systems, not at people using ChatGPT. To pass, a project had to get out of trial, go into real use, and show up in the profit and loss within six months. That is a hard bar for any new thing. A firm could have saved its team eight hours a week and still been counted as a failure, because eight hours a week does not reach the accounts by Christmas. It is a bit like testing whether every new hire has paid for themselves by month six, then announcing that 95% of hiring fails.

Then watch what happened to the number. People who read past the headline found that the 95% rested on 52 interviews, and that the report itself said those figures “may not represent broader market patterns”. By April this year a full teardown had gone further, pointing out that around 80% of the companies surveyed had never built a custom AI tool in the first place, so they were never in a position to fail at one. The interesting findings were in there all along, and none of them made the headline: buying something ready made worked about twice as often as building your own, most of the budget went into sales and marketing while the better returns were sitting in the back office, and staff across those same companies were already using AI at work on their own, without asking anyone. A year on, none of that is what gets quoted. The 95% is, and there is now a small trade in selling you a plan to be in the other five.

So the lesson is not that AI works, or that it does not. It is that a frightening number is always counting something specific, and it is worth thirty seconds finding out what, because usually it is not counting you.

If you want to read more:

Yesterday’s breakthrough is today’s normal.

The newsletter

Get On This Day in AI every week.

One story from AI's recent past, what happened next, and what it means for running a small business. Five minutes, no hype.

Free 30 minute call

Tell us what's eating your week.

You leave with two or three ways to save time, whether we work together or not.