Everyone Says 95% of Enterprise GenAI Pilots Fail. That Isn't Quite What the Study Measured.

What the Study Actually Measured

The source is State of AI in Business 2025, a July 2025 preliminary report from MIT Project NANDA (https://airnd.center/v0.1_State_of_AI_in_Business_2025_Report.pdf). It combined interviews across 52 organisations, surveys of 153 leaders, and analysis of more than 300 publicly disclosed initiatives.

Start with scope, because the slogan has already widened it. The report is about generative AI, and the funnel everyone quotes is specifically about task-specific GenAI tools. Everyone takes this as "AI."

That funnel reports that 60% of organisations investigated a solution, 20% reached pilot stage, and 5% reached what the researchers called successful implementation.

That is not the same as showing that 95% of pilots failed. If the stages use the same organisational denominator, the chart implies that roughly one quarter of the organisations reaching pilot stage subsequently reached successful implementation. The report nevertheless also describes a "95% failure rate," illustrating how easily organisations, initiatives, pilots, deployment and financial impact can become interchangeable in transmission.

Success is also defined more than once. In the body, "successfully implemented" means a tool that users or executives described as producing a marked and sustained productivity or P&L impact. The appendix defines it somewhat differently, as deployment beyond the pilot with measurable KPIs, and says ROI was assessed six months after the pilot. The two overlap, but they are not the same test. That is worth noticing rather than holding against the researchers: if the work itself needed more than one formulation of success, a single percentage was never going to carry it intact through a news cycle.

The authors caution that the figures come from individual interviews rather than official company reporting, and that definitions and sample sizes varied. They describe the findings as preliminary and name their own limitations: selection bias, inconsistent success metrics, possible confounders, and a six-month window that may understate longer-term results. The relevant pages are 2, 6 to 7 and 24.

Three Findings, One Percentage

The report contains three distinct claims or measurements that the headline blends.

Organisation-level return

Ninety-five percent of organisations are described as getting zero return from generative AI. That is a statement about aggregate business performance, not about any individual system.

Pilot-to-production conversion

The funnel's own stages, 20% piloting and 5% implementing, describe how many organisations moved from one stage to the next. Read on a common denominator, that is roughly 25% conversion, not 5%.

A definition of success

"Successfully implemented" is a test rather than a count: sustained productivity or P&L impact in the body, deployment beyond the pilot with measurable KPIs in the appendix, with ROI assessed at six months. A deployment that runs well but has not yet produced a detectable signal by then does not clear it.

A listener hearing "95% fail" pictures the second. The number being quoted is the first. The test doing the real work is the third.

Four Things to Formalize Before the Next One

The practical response to a contested measurement is a better one of your own, and it is cheaper than it sounds. Before the next pilot starts, write down four things and get them agreed by whoever will judge it.

What uncertainty this pilot resolves. Feasibility, adoption, quality, or economics. These need different designs and different durations, and most pilots are quietly asked to resolve all four and then judged on the last.

What result would justify continuing. A threshold, stated in advance, by the person controlling the budget. If nobody will commit to a number beforehand, the pilot's real purpose is political and should be planned accordingly.

What result would justify stopping. Almost nobody writes this one, which is why so few pilots ever formally end.

When the economics should first be visible, and why you believe that. If the answer is six months, say why six. If the honest answer is eighteen, then a six-month review is a checkpoint rather than a verdict, and everyone should know that going in.

Four sentences. They take a few structured sessions and they remove most of the ambiguity that later gets reported as failure.

That fourth sentence is the one the report's own caveat points at, and it deserves a name.

The Measurement-Horizon Trap

The measurement-horizon trap is judging a change in how an organisation works against a window that was chosen for administrative convenience rather than for economic realism.

The pilot window exists for good reasons. It bounds spend, forces a decision and keeps a portfolio moving. But it is a governance period, not an economic one. When an organisation measures financial impact at the end of that window and finds little, there are two readings available: the intervention does not work, or the instrument has a shorter reach than the phenomenon. The slogan encourages the first reading without leaving much room for the second. The NANDA authors, to their credit, flag exactly this about their own six-month window.

This is not an argument that everything pays back eventually and patience is always warranted. That reasoning has funded a great deal of waste. It is an argument that the horizon should be chosen deliberately, defended explicitly, and stated before the measurement rather than after it.

Underneath all of it sits a definition worth correcting. A pilot is not intended to fail. It is intended to resolve uncertainty cheaply. A pilot that runs for eight weeks, produces a clear answer of "no," and costs a fraction of a full deployment has performed exactly as designed. Recording that as a failure is a category error that then propagates into a statistic. The real failure is the pilot that resolves no uncertainty at all, regardless of what it shipped.

Why This Is Not Pedantry

Because the statistic is doing real work in real organisations right now.

It can become an argument for pausing programmes, for not building internal capability, for waiting until the technology matures. Some of those may be correct decisions. None of them should rest on a number whose scope the person quoting it cannot describe.

A widely repeated figure that has become detached from its methodology is not worthless. It is evidence with uncertain applicability and uncertain fidelity, which is a different thing from being wrong, and a different thing again from being sufficient to change a capital allocation.

The durable habit is the general one. When a convenient figure arrives, ask what was measured, over what period, in which population, against what definition. If the person citing it cannot answer, that is information about the citation rather than about the underlying reality.

The Question

Think about the last pilot your organisation stopped.

What would it have had to prove for you to have called it a success, and did anybody write that down before it started?

I suspect the second answer is usually "no", and that the first answer moved during the pilot without anyone announcing it. If your experience is different, I would like to hear how you did it, because the discipline is rarer than it should be.

Originally published on LinkedIn on September 15, 2026.