A widely quoted 95% AI failure figure rests on 52 interviews

What the report actually is

“The GenAI Divide: State of AI in Business 2025” is a July 2025 report from a group calling itself MIT NANDA, authored by Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari. The report itself, retrieved here through a third-party mirror since MIT's own project page no longer lists the file, describes its method plainly on its second page: a systematic review of “over 300 publicly disclosed AI initiatives,” structured interviews with representatives from 52 organisations, and survey responses from 153 senior leaders collected at four industry conferences, covering January to June 2025. NANDA is a research initiative at the MIT Media Lab, though its current project page, focused on later work on agentic-web architecture, no longer features this report.

The headline, and what produced it

The report's most quoted claim is that 95% of organisations are “getting zero return” from generative AI pilots, against an estimated $30 to 40 billion in enterprise investment, with just 5% of integrated pilots said to be extracting measurable value. It also reports that over 80% of organisations have tried general tools like ChatGPT or Copilot, mostly for individual productivity rather than profit-and-loss impact, while only 5% of custom enterprise AI tools it tracked reached production. The report itself labels its pilot-to-production figures as “directionally accurate based on individual interviews rather than official company reporting,” a caveat attached to the same data behind the 95% figure.

How the figure travelled

By the time the report reached wider coverage, its stated method had already shifted in the retelling. A Fortune article from 18 August 2025, used here as a secondary source, describes the same study as drawing on “150 interviews” and a “survey of 350 employees,” figures that do not match the report's own stated 52 organisations and 153 senior leaders. Neither number is necessarily wrong on its own terms, but the mismatch shows how a preliminary, interview-heavy study's headline percentage can circulate further than the sample size and definitions behind it, with each retelling smoothing over the report's own caveats about directional accuracy and anonymised, self-selected sources.

Questions to ask before you adopt it

• Does the “95%” figure describe your kind of pilot, custom-built enterprise tools, or does it get applied loosely to general tools like ChatGPT as well?
• Is the number you are quoting from the report itself or from a later summary that may have restated its method?
• What would “measurable return” mean for your own pilot, and is that the same definition the report used?

A report built on 52 interviews and a conference-floor survey of 153 leaders can be a genuinely useful signal about why pilots stall. It is not a measured, representative statistic about enterprise AI generally, and the gap between what the report says about its own method and what gets repeated about it is itself worth noting whenever the 95% figure appears unqualified.