Ninety-five percent. That's the share of corporate generative AI pilots that are failing, according to a new MIT report highlighted by Fortune. For every twenty initiatives a company launches, nineteen aren't delivering what their sponsors promised.

Talk to almost any CIO right now and you'll hear some version of the same story. It usually starts in a boardroom, with a mandate: we need an AI strategy. A pilot gets scoped, a vendor gets signed, a chatbot or copilot gets bolted onto an existing workflow. Then, months later, nobody can point to a number that moved.

MIT's finding puts a hard figure on that quiet frustration. And the number matters because it lands at a moment when AI budgets are still expanding — which means most of that new spend is flowing into projects that, by the report's own measure, don't work.

What does this change for enterprise AI programs?

In practice, it shifts the burden of proof. If 95% of pilots fail, the question stops being "are we doing AI?" and becomes "can this specific deployment show a measurable return?" For CIOs, that reframes vendor conversations, too: a demo is no longer evidence, and neither is a pilot that never graduates into production.

The failure rate also sets a realistic planning baseline. If you greenlight ten generative AI initiatives, MIT's data suggests you should expect roughly nine of them to stall. Portfolio thinking — spreading bets, killing weak projects fast, reserving real investment for the one or two that show traction — beats betting the transformation budget on a single flagship pilot.

There's a reputational dimension as well. Executives who stood up AI programs and promised productivity gains now face boards asking harder questions. A 95% failure rate doesn't mean AI is worthless; it means the deployment model most companies are using isn't working yet.

The pattern behind the number

The report's headline figure echoes what practitioners have been reporting anecdotally for the past two years: pilots that demo well but never integrate into core processes, tools that employees try once and abandon, and business cases built on assumptions that production workloads don't survive.

For technology leaders, the takeaway is less about abandoning generative AI and more about how these projects get scoped. The gap between a working demonstration and a deployed, measurable system is where most of that 95% lives.

The bottom line

MIT just put a number on what many CIOs already suspected — and now every AI budget line has to survive it.