Nornic Register · Living document
“95% of enterprise AI pilots fail.”
A preprint that describes itself as preliminary, drawn from a self-selected sample, published by a group that sells the remedy it recommends.
The source
- Study
- The GenAI Divide: State of AI in Business 2025
- Publisher
- MIT Media Lab — Project NANDA
- Publication type
- Preprint. Self-described as "Preliminary Findings"
- Sample
- ~52 structured interviews, ~150 survey responses, ~300 public deployments; self-selected
- Fieldwork
- Six-month observation window, 2025
- Funding & conflicts
- Project NANDA commercialises agentic-AI infrastructure of the kind the report recommends
- Verified on
- 2026-07-26
The arithmetic
The underlying observation is almost certainly directionally true: most pilots do not reach production. Several better-sampled studies agree.
The precision is what fails. "95%" implies a measured rate across a defined population. The sample is self-selected and the document is explicitly preliminary.
The conflict of interest is material and rarely disclosed alongside the figure.
A more defensible number for the same phenomenon: 42% of companies abandoned most AI initiatives in 2025, up from 17% the year before — S&P Global Market Intelligence, n > 1,000.
The finding in the same report that almost nobody quotes: deployment succeeded in 67% of cases via external partnerships versus 33% for internal builds. That is the more actionable result.
Nornic does not use the 95% figure as a claim. It is listed here because correcting it is useful and repeating it is not.
What would change this verdict
A peer-reviewed or large-sample study reporting a comparable failure rate on a defined population, published by a party without a commercial interest in the conclusion.
Sources