The 95% AI Failure Stat: What the Study Actually Says

If you have read anything about AI disappointment in the last year you have met this number. Ninety-five percent of enterprise AI pilots fail. It is in board decks, on cable news, in the AI bubble arguments, and in roughly every LinkedIn post arguing that the emperor has no clothes.

It comes from one MIT working paper called The GenAI Divide. I went and read it, because a number doing that much work deserves twenty minutes.

The number is real. It also does not mean what almost anybody using it thinks it means.

What the study actually measured

The paper defines success as deployment beyond the pilot phase, with measurable KPIs, and ROI impact measured six months after the pilot.

Read that again, because everything turns on it. A project that saves a department real time, and that everyone in it would fight you to keep, counts as a failure under this definition unless somebody attached a KPI to it and measured profit and loss impact within six months.

The report is not even consistent about it. Its own research note counts a tool as successfully implemented when users or executives reported a marked and sustained productivity or P&L impact, which is broader. The narrow six-month KPI definition in the methodology section is the one the headline rests on.

That is a defensible research choice. Direct P&L attribution is the hardest bar and the one executives claim to care about. It is not the same thing as "did not work," and the gap between those two readings is where the entire public use of this number lives.

Paul Roetzer of the Marketing AI Institute made the point sharply: the definition ignores efficiency gains, cost reduction, churn reduction, lead conversion and pipeline velocity. Those are not soft benefits. Those are most of why firms buy this stuff.

The funnel, and the arithmetic nobody does

Here is the part that changed how I read the headline.

The 5 percent figure comes out of a funnel for task-specific GenAI tools, and the base is organizations, not projects. Sixty percent of organizations evaluated such tools. Twenty percent reached the pilot stage. Five percent reached production.

So the 5 percent is 5 percent of every organization that looked at these tools, including the ones that never got past evaluation. Measured against the organizations that actually ran a pilot, about a quarter cleared a deliberately high bar.

One in four is a different story from one in twenty. It is not a scandal, and it does not support the sentence it is usually attached to.

The finding that got buried

This is the one that should be in every article citing the study and is in almost none of them.

The same report found that general-purpose AI tools were doing well. Pilot-to-implementation rates above 80 percent. More than 80 percent of organizations had explored or piloted general-purpose LLM tools, with close to 40 percent reporting deployment.

Sit with that. The paper that produced the most-cited evidence that AI does not work in the enterprise also found that when firms deploy AI as a general-purpose tool rather than a bespoke embedded system, it works most of the time.

The 5 percent is not a verdict on custom work against bought software, either. The report says enterprise-grade systems, custom or vendor-sold, are both being quietly rejected. Its sharper finding is about who builds: customized tools bought through external partnerships reached deployment about 67 percent of the time, against about 33 percent for tools built in-house. General-purpose tools, meanwhile, mostly raised individual productivity rather than P&L.

The study's actual thesis is not "AI fails." It is closer to "partner on customized tools rather than building alone, and expect general tools to help individuals before they show up in P&L." That is genuinely useful. It is also the opposite of how the number gets used.

How solid is the underlying research

Worth being even-handed here, because the study is being over-attacked as well as over-cited.

The headline rests on about 52 structured interviews, supported by a systematic review of 300-plus publicly disclosed initiatives and 153 survey responses gathered across four industry conferences. The authors themselves describe the figures as directionally accurate and based on individual interviews rather than official company reporting. The report does not spell out the method for the 300-plus review.

That is a preliminary working paper doing what preliminary working papers do. The authors were careful. The people citing them have not been.

Conference-recruited respondents are also a specific population. People who attend industry AI conferences are not a random sample of firms.

What I think you can honestly say with this number

Fair uses:

• Task-specific GenAI tools, custom or vendor-sold, have a low rate of reaching production with measured P&L impact inside six months.
• Tools built in-house reached deployment about half as often as customized tools bought through a partner.
• Measuring AI ROI within six months is hard enough that most firms do not manage it.

Unfair uses:

• "95 percent of AI projects fail." The study did not measure whether projects worked. It measured whether they cleared one specific bar.
• "AI does not deliver value in the enterprise." The same study found general-purpose tools widely piloted and deployed.
• Citing it as settled evidence. It is a working paper built substantially on 52 interviews.

Why this matters beyond one statistic

The 95 percent number is load-bearing in the AI bubble argument. It is the empirical leg. If enterprise adoption is genuinely producing nothing, the capital expenditure case collapses and the valuations follow.

So the strength of that leg matters, and it is weaker and more specific than its usage suggests. As of December 2025, Goldman Sachs had revised its estimate of 2026 AI capital expenditure up from 465 billion to 527 billion dollars, so the spending side of the argument is real. The adoption side leans on this number, and it is being carried further than it can go.

If you are going to use a number in a board deck, read the definition of success it uses. That is the whole lesson, and it generalizes well past this one paper.

The practical read for a firm

The study's real finding is good advice. Internal builds reached deployment about half as often as customized tools bought through a partner, and general tools showed up first as individual productivity.

Start with the general tool and prove the work changes. When you need something customized, partner rather than build alone, and set the KPI before the pilot starts so the result can be measured.

That advice would be correct even if the number were 50 percent, or 20.