Do 95% of AI pilots fail? Read the fine print

Ever since Fortune published it on 18 August, not a week has gone by without someone sending me the same headline: “95% of generative AI pilots are failing”. It comes from the report The GenAI Divide: State of AI in Business 2025, by MIT Media Lab’s NANDA project.

I have read it in full. And I think it says some very useful things, though not exactly the ones the headline repeats.

First, the method

The document itself is presented as “preliminary findings”. It is based on an analysis of more than 300 public initiatives, interviews with 52 organisations and a survey of 153 senior leaders at four industry conferences. The authors acknowledge that their figures are “directionally accurate” based on interviews rather than official company reporting, and that definitions of success may vary.

It is not a peer-reviewed study. That does not invalidate it, but it does mean reading it for what it is: an interesting snapshot, not a universal law.

What it really says

The central claim is that, despite $30–40 billion in enterprise investment, 95% of organisations are getting no measurable return. The most interesting part is the why. According to the report, the problem is not the quality of the models, but:

• tools that neither learn nor adapt to real workflows;
• projects that start with the technology rather than a specific process;
• initiatives that remain stuck as pilots because nobody integrates them into day-to-day work.

Anyone who has worked on implementation will recognise this. The demo works; the following Monday, the team goes back to its Excel spreadsheet.

What the 5% do

My reading is that the report describes the difference between buying AI and implementing it quite precisely:

1. Start with a process that has an owner, not with a tool looking for a use.
2. Measure from day one: hours, errors, turnaround times. Without a baseline there is no return to demonstrate.
3. Integrate into the tools the team already uses, instead of adding yet another screen.
4. Iterate with the people who use it, because the first version is never the right one.

A note on the August noise

This month also saw the arrival of GPT-5, on 7 August, and on 2 August the AI Act obligations for general-purpose AI models began to apply. Lots of movement at the top. But the MIT report is a reminder of something no launch changes: the value of AI is decided at ground level, in how it is integrated into each person’s work.