PROVENANCE REGISTER
Where the AI statistics actually come from.
The numbers that get quoted in board meetings mostly travel without their sources. We opened the primaries. Several of the most repeated figures in the industry rest on nothing — a panel remark, a forecast that expired unscored, a footnote pointing at a magazine interview.
One of them was ours. We quoted it for months before we opened the source. It is in the register, marked as such.
The register
"87% of data science projects never reach production"
Traces to nothing
A sponsored VentureBeat article (July 2019) reporting a panel remark. On stage, IBM's Deborah Leff said "I think CIO Dive Magazine says that only 13%…" — a verbal recollection of a trade-magazine claim, inside sponsored content. No study, no dataset, no method. It underpins much of the MLOps vendor category's marketing.
Open the primary →"85% of AI projects fail"
A forecast, not a measurement
A Gartner forecast from 2018: that through 2022, 85% of AI projects would deliver erroneous outcomes from bias in data, algorithms or teams. "Erroneous outcomes" is not "failure". The window expired in 2022 and was never scored.
"RAND: more than 80% of AI projects fail"
Traces to nothing
RAND report RR-A2680-1 (2024). The sentence carries footnote 13, which leads to a Fortune interview with a vendor CEO — not to data. RAND's own hedge, "by some estimates", is dropped every time it is quoted. RAND studied why projects fail; it never measured a rate.
Open the primary →"MIT: 95% of GenAI pilots fail"
Preliminary, and misreported
MIT NANDA, The GenAI Divide, July 2025: 52 interviews and 153 leaders recruited across four industry conferences. Self-labelled "Preliminary Findings", version 0.1, with no bibliography of any kind. "Successfully implemented" is defined as something users or executives "remarked as" causing impact — an impression, not a measurement. Its disclaimer states the views are not those of any affiliated employer, so it is not an MIT institutional finding.
"BCG: 10-20-70 — 70% of AI value is people and process"
Traces to nothing
A heuristic from a 2019 TED@BCG talk, later restated in a press release as though it were a survey result. It was never empirically derived, and it is not an output of the 1,803-executive survey it appears alongside. The "MIT Sloan found 70%…" corroboration that circulates with it appears to have been invented by SEO content.
"42% of companies show zero ROI; the profitable 58% set KPIs first"
Traces to nothing
A vendor blog post. No survey, no method, no sample disclosed. We flag this one specifically because it is the claim we would most like to be true — it argues for exactly what we sell — and it is the weakest item in the register.
"By 2028, one in four enterprise software purchases will be made by AI agents"
Untraceable
Attributed to Gartner wherever it appears. We could not trace it to any primary Gartner release.
"Gartner: 41% of prototypes reach production"
Untraceable
Appears only inside a client-gated Gartner document. Nobody quoting it can have verified it, and neither could we.
"82% of banks do not measure ROI on any technology investment"
Untraceable
Attributed to a 2025 Bank Director survey of 141 respondents, but reachable only through a secondary blog. We could not open the primary.
The one we got wrong
"40% of AI agent projects get abandoned (Gartner)"
A forecast, not a measurement
Gartner's press release of 25 June 2025 says that over 40% of agentic AI projects will be cancelled by the end of 2027. Future tense, stated horizon. Its only underlying data point is a January 2025 poll of 3,412 webinar attendees — self-selected, no sampling frame, no response rate.
We carried it in our own offer document as a present-tense abandonment rate, and made it support a causal claim Gartner never makes: its stated causes are cost, unclear value and risk controls. We corrected it on 24 August 2026. The figure can be cited — but only as a prediction, and never without its horizon.
Open the primary →What does hold up
Debunking is the easy half. Across the surveys that do disclose their method — Gartner (n=644 and n=782), S&P Global (n=1,006), IBM (n=2,000), Deloitte (n=3,235) — the picture is consistent and much less dramatic than the headlines:
Roughly 20 to 30% of enterprise AI initiatives reach production and meet their ROI expectations. Outright failure runs at about 20%. Between them sits a large partial-success middle that no headline number describes.
The best-methodology work we have found: the University of Melbourne and KPMG global study on AI use and trust (n=48,340, 47 countries, nationally representative), and Google Cloud's DORA report on AI-assisted software development (~5,000 professionals, Bayesian models with credible intervals). Neither produces a scary headline, which is probably why neither circulates.
And the finding nobody quotes from the MIT report itself: general-purpose LLM tools reach 40% successful implementation. The 5% applies only to custom, task-specific builds.
How to check one yourself
- Find the primary. Not the article quoting it — the release, the PDF, the report. If three clicks do not reach a document, that is already the answer.
- Read the tense. "Will be" is a forecast. A forecast with an expired window that nobody scored is not evidence of anything.
- Find n, and find who was asked. Webinar attendees and conference recruits are self-selected. A probability sample is a different object entirely.
- Read the definition behind the number. "Failure" and "success" are defined in the method section, and the definition is often far weaker than the headline implies.
- Follow the footnote. More than once here, it led to a magazine interview with someone selling the remedy.
This register is maintained from our internal evidence file. If you find an error in it, we want to know — that is the whole point of publishing it.
This is the same discipline behind Counsel's public verdicts: every claim quoted from a real source. Or read the long version.