What is the AI project failure rate?

Different studies measure different things, which is why the failure rate is quoted as a range rather than a single figure. RAND’s 80% comes from interviews about AI projects that never reach reliable production. MIT’s 95% measures a narrower, harsher bar: enterprise generative-AI pilots that produce no measurable impact on the P&L.

They point the same direction. Whether you define failure as never shipped or shipped but moved no money, the large majority of AI initiatives don’t pay off. McKinsey’s finding that only 39% of adopters see any enterprise EBIT impact, despite 88% using AI, is the same story from the value side.

The consistent finding across RAND and the other studies is that failure is usually a governance and prioritisation problem, not a technology problem. The most common root causes are organisational:

An 80% to 95% failure rate doesn’t mean AI doesn’t work. It means most bets, judged one at a time, were never the ones worth funding. Teams that beat the average don’t have better luck. They compare every bet up front and demand evidence before capital.

That reframes the number. A high failure rate is exactly why a portfolio approach pays off: if most bets fail, the value is in cheaply identifying the few that won’t, before the build budget is committed.

Making evidence cheap changes the odds. Pretotyping produces real behavioural signal on a bet in days, for a fraction of a build. Score every bet, pretotype the close calls, and set fund, fix, or kill thresholds in advance.

That is what a Rapid Experimentation Operating Model installs: the discipline to stop funding the 80% to 95% that were never going to pay off, and back the few that will.