Why 95% of AI projects return nothing (and what the 5% do)

Both are wrong, and the report it comes from is more interesting than the argument it started. Here is what it really says, what it does not say, and what to do with it when you run a company.

Where the figure comes from

It comes from a report titled The GenAI Divide: State of AI in Business 2025, published in July 2025 by MIT's NANDA initiative, signed by Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari.

Its method, as the authors describe it: a review of more than three hundred publicly disclosed AI initiatives, structured interviews with fifty-two organisations, and a survey of one hundred and fifty-three executives, all between January and June 2025.

The report's exact claim is this: despite thirty to forty billion dollars invested by companies in generative AI, 95% of organisations get no return.

Three qualifications are needed before going further, and the authors give them themselves. The report is labelled “preliminary findings”. It rests on individual interviews rather than published financial statements. And the authors acknowledge that the definition of success varies from one organisation to the next.

In other words: it is a strong, documented signal, not an accounting measurement. Anyone quoting it to you as a law of nature has not read it.

What the figure does not say

It does not say the technology does not work. That is the most widespread confusion, and it leads to the wrong decision.

Several solid academic studies measure real gains. Work by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, covering more than five thousand agents in a customer support centre, measured a productivity increase of about 14% on average, and 34% among the least experienced agents. An experiment published in Science by Shakked Noy and Whitney Zhang, run with 453 graduate professionals, measured a 40% reduction in writing time and an 18% improvement in quality.

Both studies point the same way, and it is valuable management information: AI helps most those who master least. It spreads the knowledge of the best to everyone else. It is not a tool for genius, it is a levelling tool.

You should also know the opposite result. In July 2025, the organisation METR ran a controlled trial with sixteen experienced developers working on their own projects, across 246 real tasks. With AI, they were 19% slower. The most troubling part is not that: they believed they had been 20% faster. METR notes that this result concerns experts, on code bases they knew by heart, with early-2025 tools, and that it does not generalise.

What to remember above all is the gap between perception and measurement. That is exactly why a management team can be convinced its AI project is working without a single indicator confirming it.

What the 5% do

This is the useful part of the report, and oddly the part nobody quotes.

They buy more often than they build. The finding is clear: internal builds fail twice as often as external partnerships. The reason is simple and familiar to any executive who has run an IT project: building is the easy part, maintaining and evolving is the part that costs.

They invest where the value is, not where it shows. The report notes that about half of generative AI budgets go to marketing and sales, while administrative and back-office functions offer a better return. That is human: we fund what can be shown to the board.

They require the tool to fit the process, not the reverse. The organisations that get a result ask for customisation by business process and judge the tool on operating results, not on demos or technical scores.

They know what they are measuring before they start. That is the condition most often missing. A project with no indicator defined in advance cannot succeed, since nobody will be able to say whether it did.

One last point deserves executive attention: according to the same report, around 90% of employees use personal AI tools at work, while only 40% of companies have official subscriptions. Your teams already use AI. The question is not whether you will get there, but whether it happens inside a framework or outside one.

An operations director quoted in the report sums up the gap between the talk and the ground: on LinkedIn everything has supposedly changed; in his operations, nothing fundamental has moved.

The four questions to ask before starting a project

If you take one thing from this report, let it be this list.

1. 1Which precise process does this tool change? Not “customer service”, but “handling order-tracking requests received outside working hours”. If the answer stays vague, the project will join the 95%.
2. 2What am I measuring, from what baseline, and over what period? Settle it before signing. A project assessed after the fact always assesses well.
3. 3What happens when the tool gets it wrong? An AI that gives a false answer to a customer, a candidate or a supplier speaks in your name. Ask how the error is prevented, detected, and who corrects it.
4. 4What is left when the supplier disappears? Is your data exportable, does your process survive, can your team work without it?

What an executive should conclude

Neither “AI does not work” nor “we have to go all in”. The right conclusion is duller and more profitable: the gap is not about technology, it is about execution. Pick a precise process, define what you measure, prefer a tool maintained by someone whose job that is, and insist on knowing what happens when the machine is wrong.

These are the same rules as for any investment. What is new is how quickly they can be forgotten.