The MIT NANDA initiative's report, The GenAI Divide: State of AI in Business 2025, produced the statistic that now opens every enterprise AI keynote: roughly 95% of generative AI pilots deliver no measurable profit and loss impact. It is the most quoted and most disputed number in enterprise AI; the methodology has been publicly challenged, and it is a research initiative's estimate, not an institutional MIT finding. S&P Global Market Intelligence supplies the corroborating trend from separate data; 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before.
Whatever the precise figure, the direction is not in doubt, and the way the number is being read is the real problem. In boardrooms it has become shorthand for "AI does not work yet", and that reading produces one of two responses: wait for the technology to mature, or buy a more capable model. Both responses miss what the research actually found.
What the report actually shows
Read past the headline and the MIT researchers are specific about where pilots died: overwhelmingly at the workflow boundary, not the model boundary. Generic tools produced fluent, plausible output that did not fit how decisions were actually made, and organisations quietly routed around them. The capability was present. The connection to a decision was not.
That diagnosis matters because it tells you what will not fix the problem. If the pilots were failing because the models were too weak, the next model generation would rescue them. But they failed because the model's output never reached a decision. A better model in the same dashboard changes nothing, so the 95% will survive every model release. The proof is in the timing. Abandonment climbed in the same year that model capability jumped.
What the two wrong responses cost
The misreading is not a semantic quibble, because each of its responses carries a bill.
Waiting treats the 95% as a maturity problem that time will solve. But the constraint the research points at, wiring outputs into owned decisions, does not improve with model releases; it improves with organisational work that takes quarters to do. An organisation that waits is not holding its position. It is donating those quarters to the minority already doing the wiring, whose systems are accumulating scored decision histories that a late starter cannot buy.
Upgrading treats the 95% as a capability shortfall, and it is the more expensive mistake because it looks like action. The new model is procured, the benchmark is better, the demo is impressive, and the output lands in the same dashboard, advisory, unowned and unscored, where its predecessor's output landed. Budget was spent making the unused thing more sophisticated.
The tell is in the timing: abandonment rates climbed in exactly the year model capability jumped. If capability were the constraint, those curves should have moved in opposite directions.
The reading that survives the evidence
The reading that fits the report's own evidence is simpler. The 5% did not have better models. They wired predictions into real decisions: a named owner, a threshold for acting, and a record of whether acting worked. How to do that wiring is its own subject, covered in the flagship essay. The point here comes first. None of that work can start until the statistic is read correctly, and in most boardrooms it is not.
Read the 95% as "the models failed" and you will wait or upgrade. Read it as "the decisions were never wired" and you have work you can start on Monday.
Wiring predictions to owned, scored decisions is the entire design brief of Prophesee. If you want to see what the 5% built, on your own data, start here.