Why most AI projects fail, and what the “95%” really says.

The number everyone quotes

In 2025 an MIT project reported that 95% of organisations saw no measurable return from generative AI 1. It spread everywhere.

Read closely, it’s narrower than the headline. It wasn’t peer reviewed, its data isn’t published, success meant a measurable profit effect within six months, and many of the pilots it counted had no baseline to measure against 2. A pilot with no baseline can’t show a return even when it has one.

Other figures agree in direction. A survey of over 1,000 firms found 42% had abandoned most of their AI initiatives in 2025, up from 17% a year earlier, and the average firm scrapped 46% of its proofs of concept 3. RAND’s “over 80% fail” is a figure it repeats from others; its own evidence is 65 interviews 4.

What actually goes wrong

Nobody acts on it. Only about a quarter of employees use business intelligence tools, a share that has barely moved in 25 years 5. A list nobody works is worth nothing.

One visible mistake ends trust. After people see an algorithm make an error, they choose a worse human forecaster instead 6. A reason that says “enquired twice” when it was one duplicated enquiry can undo a month of good work.

Nobody measured it fairly. When eBay switched off brand search ads in some markets, 99.5% of that traffic arrived anyway, credit the ads had been taking for years 7. Without a comparison group, every result becomes an argument.

AI used beyond its ability. In a preregistered study of 758 consultants, AI made them faster and better on tasks inside its ability, and 19 percentage points less likely to be right on a task outside it 8.

What this means for a business

• Choose the problem before the technology, and check the data before promising anything.
• Put the output where people already work, and count adoption as the first result.
• Agree a baseline and a comparison group before building. Report ranges, and say what you can’t attribute.
• Check every fact an AI states against the record, in code, before a person sees it.

What we don’t know

• Almost none of these studies looks at small businesses specifically.
• Failure rates for simple, narrow tools (one list, one report) are probably lower than for broad programmes. We haven’t found good evidence either way.

Sources

1. Claim MIT NANDA, The GenAI Divide: State of AI in Business 2025 (coverage) · Not peer reviewed; data not published
2. Claim Critique of the 95% figure (method, baseline, six-month window)
3. Fact S&P Global Market Intelligence 2025, via CIO Dive · Survey, self-reported, 1,000+ firms
4. Claim RAND, The Root Causes of Failure for AI Projects, 2024 · The 80% is RAND repeating other estimates
5. Fact BARC and Eckerson, via TechTarget
6. Fact Dietvorst, Simmons and Massey, “Algorithm Aversion”, 2015 · Peer-reviewed experiments
7. Fact Blake, Nosko and Tadelis, Econometrica 2015, via Chicago Booth Review
8. Fact Dell’Acqua et al., “Navigating the Jagged Technological Frontier”, HBS and BCG, 2023 · Preregistered field experiment

Related problems