Ninety-five per cent of what? Reading the most quoted AI statistic

In August 2025 a report from MIT NANDA called The GenAI Divide produced a line that travelled fast: 95% of generative AI projects deliver no measurable business return, despite $30 billion to $40 billion of enterprise investment.¹

It has been repeated widely since, usually as shorthand for the state of AI investment. It is a perfectly good number to quote. It becomes risky when it travels without its definition.

What the study did

According to trade press coverage of the report, the authors reviewed more than 300 publicly disclosed AI initiatives, carried out 52 structured interviews and collected 153 survey responses from senior leaders at four conferences, between January and June 2025.¹

Their yardstick was measurable impact on profit and loss. On that test, only 5% of integrated AI pilots were extracting millions of dollars in value, and only 5% of custom enterprise AI tools had reached production.¹

What it can support

A narrow statement holds up well: in this sample, over this window, very few projects showed a measurable effect on profit and loss. The funnel the study describes also holds up. Many organisations evaluate tools, fewer pilot them and fewer still reach production.

The authors name learning capability as the central barrier, meaning tools that do not keep feedback, adapt to context or get better with use.¹ That observation is worth testing in your own pilots, and it is cheap to test.

What it leaves open

Public initiatives and conference surveys are a particular slice of the projects that exist. Private pilots may be under-represented, and so may projects whose value is real but hard to pin to one line of the P&L. Six months of fieldwork can also miss projects that pay back in their second year.

Other studies ask other questions, so their answers sit in a different range. McKinsey’s 2026 survey found 37% of respondents attributing at least some EBIT impact to AI.² IBM’s 2025 CEO study reported that 25% of AI initiatives had delivered the expected ROI.³ You cannot average these with MIT’s figure, since each one counts something different.

A failure rate means little until you know how the study defined success.

Before you quote it

Three details should go wherever the number goes. The measure of success, because revenue, cost, EBIT, adoption and expected ROI each give a different answer. The period, since a twelve-month window and a thirty-six-month window describe different projects. And the sample, because public cases, surveyed executives and a random sample of firms each lean in their own direction.

A board paper that quotes 95% from MIT and 37% from McKinsey on the same page is fine, as long as both carry those three details. Then nobody mistakes them for a contradiction.