A CIO forwarded me a board deck last month. Slide 4 had one number on it, in 90 point type, with no footnote. His board had already voted on it.
The research is real. The number is real. What people do with it is not.
The number never measured what it is used to prove.
The source is The GenAI Divide: State of AI in Business 2025, published July 2025 by MIT's Project NANDA.
Page two of the report labels it Preliminary Findings. The methodology is 52 structured interviews, 153 survey responses collected at four industry conferences, and a review of 300 publicly disclosed AI initiatives.
The 95% measures the inverse of a single funnel in section 3.2, for custom and task specific GenAI tools only. Sixty percent investigated. Twenty percent piloted. Five percent reached production.
General purpose tools were not in that funnel. The report says elsewhere those convert at roughly 83%.
Directly beneath the exhibit, the authors write that the figures are “directionally accurate based on individual interviews rather than official company reporting.”
They also define success as a deployment that users or leaders “remarked as” causing sustained productivity or P&L impact. Remarked. Not audited. Not measured.
The observation window was six months. In the appendix, the authors note this “may be insufficient” and could be “understating success rates.”
None of that survives the trip to a slide.
The report is preliminary, self-reviewed, and published by a party with a stake in the answer.
The reviewer is one of the authors. Page two lists a single reviewer, Pradyumna Chari, Project NANDA. Page one lists Pradyumna Chari as a co author. This is not peer review. It was never presented as peer review.
The report contradicts itself on its own second most quoted figure. Section 3.4 puts sales and marketing at “approximately 70 percent” of GenAI budget. Section 6.3 puts it at 50%. Both appear in the same document.
The publisher has a position. The appendix states that Project NANDA “builds on Anthropic's Model Context Protocol and the Google/Linux Foundation A2A to create infrastructure for distributed agent intelligence at scale.” The report's conclusion is that the fix is agentic systems with persistent memory. That happens to be the category NANDA builds.
That does not make the work dishonest. It makes it interested. Interested research can still be correct. It just should not be quoted as a neutral scoreboard.
Strip the headline and the findings underneath are worth more than the number that made it famous.
Ninety percent of surveyed employees use personal AI tools for work. Forty percent of their companies bought a subscription. That gap is the real finding and almost nobody quotes it.
Externally sourced tools reached deployment about 67% of the time. Internal builds, about 33%. Twice the success rate for buying over building.
Mid market firms went from pilot to production in about 90 days. Enterprises took nine months or more.
Half to seventy percent of budget went to sales and marketing, while the documented savings sat in the back office. Two to ten million a year from eliminating outsourced customer service and document processing.
Read that way, the report is not a verdict on AI. It is a verdict on how organizations buy it.
We are telling you in advance what evidence would move this score.
Release of the underlying data. A replication with a defined, audited success metric and a window longer than six months. Either would move this score.
Five components, twenty points each. Higher means more hype risk.
The score above is the council's. Five models from five laboratories read the fetched text of the sources listed below, score the claim against the same published rubric independently, then review each other anonymised. Each edition is scored three times, because one run is a sample of size one, and the five component scores are the median of those runs.
| Model | Score | Verdict | In one line |
|---|---|---|---|
| GPT-5.6 | 57 | Half True | The claim reflects a real but imperfectly measured finding about zero-return organizations or integrated pilots, overstated as a universal failure rate for enterprise AI pilots. |
| Claude Opus 4.8 | 59 | Half True | A single preliminary MIT NANDA report saying 95% of organizations saw zero GenAI return has been reworded into '95% of pilots fail' and circulated sourceless in vendor decks. |
| Grok 4.5 | 63 | Overstated | A real but narrowly scoped NANDA finding has been stripped of its organizational and 'zero-return' framing and turned into a free-floating failure statistic. |
| Sonar Pro | 29 | Holds Up | The claim is broadly supported by one MIT report but is being repeated more broadly than the evidence cleanly justifies, especially when treated as a universal failure rate for all enterprise AI pilots. |
| DeepSeek | 45 | Half True | The claim is based on a reputable but non-peer-reviewed study with a reasonable methodology, but it is often cited without its original context and qualifiers. |
The claim reflects a real but imperfectly measured finding about zero-return organizations or integrated pilots, overstated as a universal failure rate for enterprise AI pilots.
A stronger case for a lower score is that the report itself explicitly states the 95% figure and is based on a named, multi-method study, so the popular shorthand may be only a mild simplification rather than hype.
No peer reviewed replication will put enterprise AI ROI failure at 95%. When better studies land, the honest range will fall between 60 and 80 percent. Most of that spread will come from how each study defines the word success.
July 26, 2026. Re-scored against rubric v1.0 when it was published. Source quality 14 to 13, Sample and method 15 to 14, Independence 16 to 12, Replication 15 to 12, Drift 14 to 16. Total moved from 74 to 67. Publication date corrected from July 26 to July 21, the Tuesday it went out. The source line now states that the report PDF is a third-party mirror rather than a publisher-hosted copy.
July 26, 2026. Challenge C-004 upheld. Sample and method moved from 14 to 16. Fifty-two interviews is a small sample for a claim about enterprise AI generally, which is the 15 to 17 band, and scoring it as adequate size contradicted this edition's own headline statistic. Total moved from 67 to 69. The verdict is unchanged at Overstated.
July 27, 2026. Rescored. This edition was originally scored 69 out of 100, Overstated, by the editor. Scoring moved to a council of five models on 2026-07-27, and the council put the same claim, read from the same sources, at 53, Half True. The council score is now the published one. Total moved -16 points. The original score is kept on the record and both readings will be graded against this edition’s call when it resolves.
The Hype Index, July 21, 2026. “95% of enterprise AI pilots fail.” scored 53/100, Half True. Primary source: The GenAI Divide: State of AI in Business 2025, MIT Project NANDA, July 2025. Scored by a council of five models, median of 3 runs, range 6; published by Mark Lynd. https://thehypeindex.com/edition/95-percent-of-enterprise-ai-pilots-fail/
| Claim | Verdict | Index | Call resolves |
|---|---|---|---|
| Security keys stop 100% of phishing. August 6, 2026 |
Half True | 43● | January 2027 |
| There are 4.8 million unfilled cybersecurity jobs worldwide. August 4, 2026 |
Overstated | 76● | January 2027 |
| There is a plastic spoon's worth of microplastic in your brain, and it is causing dementia. July 30, 2026 |
Half True | 53 | January 2027 |
| Superhero movies are dead. July 28, 2026 |
Half True | 50 | January 2027 |
| 95% of enterprise AI pilots fail. July 21, 2026 |
Half True | 53● | July 2027 |
Tuesday and Thursday at 1:00 PM Central. Every call is dated and graded in public, including the ones we get wrong.
One email, one scored claim, no pitch. We never sell or share your address.