AI Project Failure Rate: The Three Numbers Everyone Quotes

If you're following AI in any way, shape or form, somebody has quoted an AI project failure rate to you in the last month. It was probably 85 percent. It might have been 95 percent. If the person was being careful, it was 30 percent.

All three numbers are in circulation right now, often in the same conversation, usually from people who mean well. They are not three measurements of the same thing that happen to disagree. They are actually three different claims, from three different sources, related to three different questions — and one of them, as far as I can determine, has no original source at all.

Curious, I went looking for all three. Here is what I found.

What Is the Real AI Project Failure Rate?

There isn't one. There is no single measured AI project failure rate, because "AI project" and "failure" are not defined the same way twice. The three figures in circulation measure, respectively: nothing traceable, abandonment after a proof of concept, and enterprise pilots that never reached production with measurable profit-and-loss impact. Those are not the same question, and the honest answer to "what is the failure rate" is "of what, measured how, by whom, and when."

That sounds like a dodge. It is the opposite. Once you can ask those four questions out loud, the numbers stop being obtuse and start being useful.

The 85 Percent That Nobody Can Source

This is the one you probably hear most (or at least I do), and it is the one I would stop repeating today.

The claim is that Gartner found 85 percent of AI projects fail. It appears on consultancy blogs, in vendor decks, in LinkedIn posts, in articles from people I respect. It is almost always attributed to Gartner and almost never linked.

The earliest widely cited version I could find is a BMC blog post from December 2018, which says that "industry research firm Gartner made an audacious prediction: that 85 percent of AI projects won't deliver for CIOs." There is no link. No report title. No date. No named analyst. The piece does reference a real, dated Gartner artifact elsewhere — a survey of more than 3,000 CIOs — but it does not attribute the 85 percent figure to that survey either.

Neither I nor my AI tools could find a Gartner press release, research note or analyst quote that originates the number. What I found instead was a chain of articles citing "Gartner," each one linking to another article that also says "Gartner."

I want to be careful here, because absence of evidence is not proof the number was invented. Gartner publishes an enormous amount behind a paywall, and it is entirely possible that something said something like this in 2018 and the trail has simply gone cold. But a statistic you cannot check is not a statistic. It is a rumour with a decimal point.

The 30 Percent That IS Real, and it's Narrower Than You Think

This one holds up, and almost nobody quotes it.

Gartner, in a press release dated 29 July 2024, predicted that "at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025." Named analyst, Rita Sallam. Stated reasons: poor data quality, inadequate risk controls, escalating costs, unclear business value.

Read what it actually says. It is about projects that already got as far as a proof of concept, so it excludes everything that died before that. It is about abandonment, not about whether AI works. And it is a forecast, written in mid-2024 about what would happen by the end of 2025 — a prediction, not a measurement.

A correctly quoted 30 percent is more useful than a mis-quoted 85 percent, because it tells you where the mortality is. Things die AFTER the demo.

The 95 Percent That Measured Something Else

In July 2025, MIT's Project NANDA published The GenAI Divide: State of AI in Business 2025, and the headline travelled faster than anything else in this category — in the report's own words, "95% of organizations are getting zero return."

The report is real and the work is serious. What is missing from almost every retelling is the scale and the authors' own warning. The findings rest on a review of more than 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." And the authors wrote this, in their own report: "These figures are directionally accurate based on individual interviews rather than official company reporting. Sample sizes vary by category, and success definitions may differ across organizations."

Fifty-two companies is a respectable qualitative study. It is not a census, the authors never said it was, and they told anyone who read past the summary exactly that — directionally accurate, from interviews, with success defined differently by different companies.

What it measured also matters. The 95 percent refers to enterprise pilots reaching production with measurable impact on the books. That is a demanding bar, and failing it does not mean the technology did nothing.

And buried under the headline is the finding I would actually put on a wall. Strategic partnerships accounted for 66 percent of deployments. Internal development accounted for 33 percent. The report puts it plainly: "pilots built via strategic partnerships were 2x as likely to reach full deployment as those built internally." The same study that produced the scariest number in the category also found that the single biggest variable was not the technology. It was whether you tried to build it yourself. What a business without a software department is supposed to do about that is a longer conversation than this one. This post is about the numbers.

What the Three Numbers Together Actually Tell You

Put them side by side and the angst drains out.

One figure is untraceable. One is a forecast about post-demo abandonment. One is a careful study of fifty-two large companies, measuring a strict production bar, whose authors asked you not to generalize it.

None of them says AI does not work. Two of them say something much more useful and much less quotable: the failure is concentrated after the demo, in the part where somebody has to carry the thing into a real workflow, and doing that alone goes badly about twice as often as not doing it alone.

I have written before that every technology gold rush produces two bets — the loud one that everything changes overnight, and the quiet one that adoption happens one pain point at a time. These three numbers are what the loud bet looks like when it is measured. They are the wreckage of transformation projects, not evidence against the quiet bet. I still believe now is the time to apply new technologies to old pain.

How to Read the Next Statistic Someone Hands You

You do not need a data team for this. You need four questions, in order, and about ninety seconds.

1. Who produced it, and can I reach the original in one click? If the link goes to another article rather than to a report, you are looking at a citation, not a source. Stop there.

2. What exactly did it count? "Failure" can mean abandoned, unprofitable, never deployed, or merely disappointing. Those are four different numbers and they are not interchangeable.

3. How many, and who were they? Fifty-two large enterprises is a real finding about fifty-two large enterprises. Ask whether the sample looks anything like your business.

4. When was the fieldwork, not the article? A 2026 headline routinely sits on top of 2024 data. In this field that gap matters more than it used to.

Anything that survives those four questions is worth letting into your planning. Anything that fails them is somebody's marketing, and you are allowed to say so out loud in the meeting.

The Part That Should Actually Concern You

Here is what I take from the whole exercise, and it has nothing to do with AI.

We have reached a point where a number with no source can travel for eight years, through respectable publications, into board decks, and back out again, without anyone stopping to click. The 85 percent figure is not dangerous because it is wrong. We do not know that it is wrong. It is dangerous because thousands of people repeated it without checking, and that habit does not confine itself to one statistic.

The next number that lands on your desk will arrive with the same confidence. Give it the ninety seconds...and BTW, AI is a great tool for helping you with the search for that validation.

Jim Boudreau is the founder of Studio 1119, maker of CataSEO.ai & TruSync.ai.