Why enterprise AI pilots stall

If you have sat through a board discussion about artificial intelligence in the last year, you have heard the number. Ninety five percent of AI pilots fail. It arrives without a source, it settles the room, and it is repeated by people arguing for more investment and by people arguing for less.

It is not a reliable number. It is worth understanding exactly how it is unreliable, because the error points at the thing that actually goes wrong between a pilot and a rollout.

Is the 95 percent AI failure statistic real?

The figure comes from a working paper called "The GenAI Divide: State of AI in Business 2025", published by four authors associated with Project NANDA at the MIT Media Lab. The document is cover dated July 2025, began circulating in August 2025, is labelled "Preliminary Findings", and carries the file version v0.1.

Two things are worth saying plainly before anything else. The paper has not been retracted, and nothing here alleges bad faith. It is a preliminary working paper that was read as though it were a finished study, and most of the damage was done in transit rather than at the source.

That distinction is the whole problem, and the paper contains the arithmetic that shows it.

What the paper actually counted

The adoption exhibit reports a funnel for task specific generative AI tools. Sixty percent of organizations evaluated such tools. Twenty percent reached pilot stage. Five percent reached successful implementation.

Read those three numbers in order. The five percent is a share of all organizations surveyed, not a share of organizations that ran a pilot. Among the organizations that actually piloted something, five in twenty reached production. That is twenty five percent, not five percent.

Roughly four fifths of the famous ninety five percent is made up of organizations that never attempted a pilot at all. Counting a company that never started as a company that failed is not an interpretation you can argue with. It is a division error, and anyone repeating the headline is publishing it.

What the paper counted as success

The second problem is the definition. The exhibit note states it directly: successfully implemented tools are ones that "users or executives have remarked as causing a marked and sustained productivity and/or P&L impact".

Success is a remark in an interview. The same page describes the figures as "directionally accurate based on individual interviews rather than official company reporting". The methodology section states that impact was measured six months after the pilot, and lists as an explicit limitation that six months "may be insufficient", "potentially understating success rates".

The paper states its basis as "a systematic review of over 300 publicly disclosed AI initiatives, structured interviews with representatives from 52 organizations, and survey responses from 153 senior leaders collected across four major industry conferences". The 300 are public announcements rather than companies studied, and the 153 are people recruited at conferences rather than a random sample of enterprises.

How the number got worse on the way to you

Most people did not meet this statistic in the paper. They met it in Fortune, on 18 August 2025, which described the study as "150 interviews with leaders, a survey of 350 employees, and an analysis of 300 public AI deployments".

The interview base was overstated roughly threefold and the survey roughly twofold, and "employees" became the wrong word for a group of senior leaders recruited at industry conferences. The numbers 150 and 350 appear nowhere in the paper. Every article that repeats them is repeating Fortune, not the research.

The paper was also hard to read at the moment it went viral. On the 80,000 Hours podcast episode "The story behind the bad AI stat that moved markets and misled millions", published 28 April 2026, Rob Wiblin noted that "the article wasn't even available to journalists around the world trying to cover the story".

The methodological objection has been made by named academics. Wharton professor Kevin Werbach, quoted in LeadDev in November 2025, observed that "the 95% figure is presented in one sentence, but the authors offer no detail on how they came up with it".

The paper carries no conflict of interest statement beyond a line stating that the views are the authors' own. It also names NANDA, the authors' own project, among the infrastructure it describes as the way forward. That is worth knowing when a paper concludes that the cure is the category its authors work in.

Where can I find a figure I can actually use?

There is a properly sampled measurement of the same gap, and it is not hiding. McKinsey's "The state of AI in 2025" was published on 5 November 2025, surveyed between 25 June and 29 July 2025, and drew 1,993 responses across 105 nations.

A documented field window, a published sample size, a named institution that stands behind the work. The distance between those two numbers is the pilot to production gap, measured properly. Adoption is close to universal. Running the business on it is rare.

McKinsey's word is "scaled", not "deployed". The two are often swapped in secondary coverage, and they are not the same claim.

So what actually goes wrong between the pilot and the rollout?

A pilot runs on a narrow slice of the business, with somebody in the room who knows the slice. That person chose the example. They know which of the four systems holds the true version of the record. They know which project the invoice belongs to, which approval it needs, what state the job is in, and who owns it. If the answer had been wrong, they would have caught it.

None of that was written down. It was supplied, live, by a human being who understood the organization.

At ten users the supply is invisible. At a thousand it is the constraint. This is why the failure looks like a technology failure and is not one: the model performed the same in both cases, and what changed was whether anyone was still filling the gap.

What is missing is a model of the organization, not a better model

The systems already hold the data. The ERP holds the transactions, the document system holds the files, and the spreadsheets hold everything that never found a home. What none of them holds is the structure: what the central operating object is, and what the organization has agreed is true about it - what it cost, who approved it, what governs it, and where it stands today.

Every department sees that object from its own angle and calls it something slightly different. Finance sees a cost centre, operations sees a job, and legal sees an engagement. The object is the same. The disagreement is not about data quality. It is about the absence of an agreed structure for the thing the company earns from.

That structure is what the model is. It is also why a better language model does not close the gap: the missing information was never in the text.

What to ask before funding the next pilot

1. 01What is the central operating object, and does every system agree on its identity? If four systems hold four identifiers for the same job, that is the first problem, and it is not a model problem.
2. 02Who supplies the context in this pilot that nobody will supply at scale? Name the person. If you cannot, the pilot has not started yet.
3. 03What is the definition of success, written down before the pilot starts? A remark from a satisfied executive is not a measurement.
4. 04Over what window is it measured, and who agreed to that window in advance? Six months is a choice, not a fact, and it is short for anything that touches an operating routine.
5. 05When the system produces an answer, what does it change in the systems you already run? If the answer is nothing, you have bought a report.
6. 06If it acts, what is the record of what it did? An action with no receipt is not something you can run a business on.

None of these questions require a vendor to answer. They are answerable by the people who already run the work, and they are the questions that separate a pilot that can scale from one that cannot.