Is the 95% AI failure statistic wrong?

It is unverifiable, which for a measurement firm is the same problem. The underlying argument that most AI deployments have not produced attributable profit and loss impact is well supported by other research. The specific number 95% traces to a report whose primary PDF is no longer served at its original address, and whose nearest supported figure refers to a narrower category than the way the number is usually quoted.

What should I use instead of the 95% figure?

PwC's 29th Annual Global CEO Survey, published 19 January 2026, surveyed 4,454 CEOs across 95 countries. It found that 33% report gains in either cost or revenue while 56% say they have seen no significant financial benefit to date. It makes the same argument with a disclosed methodology, a named population and a publisher that will stand behind it.

How do I check whether a statistic in a sales pitch is real?

Ask for the primary source rather than the article quoting it. Open it yourself. Look for the sample size, who was surveyed, when the fieldwork ran, and whether the number in the pitch matches the number the study actually supports. If the person selling to you cannot produce the source on request, treat the number as decoration.

We Published the 95% AI Failure Stat. Then We Checked It.

The number behind most AI sales pitches traces to 52 interviews and 153 conference survey responses, and its source PDF is no longer served at its own address. Here is what we found and what we use instead.

Author: Breaking Trail Technologies
Published: August 26, 2026
Reading time: 7 min read

We used to cite it. It was on our homepage, on our research pages, and in the title of one of our own posts. Ninety five percent of AI pilots fail to deliver measurable return.

It is a very good number to sell against. It is alarming, it is specific, and it sets up the obvious next sentence, which is that you need someone disciplined to keep you out of the ninety five. We are a firm that sells measurement. We were quoting it constantly.

Then we went looking for the source, the way we would for a client, and we stopped using it the same week.

Where the number comes from

The figure originates in The GenAI Divide: State of AI in Business 2025, published in July 2025 by MIT’s Project NANDA.

That much is real. The report exists, the authors are real researchers, and the finding was covered widely and in good faith by outlets that had no reason to doubt it. If you have seen the number in a deck, a newsletter or a YouTube video, this is where it came from, usually three or four citations removed.

What the study actually measured

The report describes its own method. Between January and June 2025 the researchers reviewed more than 300 publicly disclosed AI initiatives, conducted 52 structured interviews with people at organisations across industries, and collected 153 survey responses from senior leaders at four industry conferences.

Read that again with a buyer’s eye. Fifty two interviews. One hundred and fifty three survey responses gathered from people who were already at AI conferences, which is not a neutral sample of businesses. The report contains its own limitation note saying the percentages reflect the interview sample and may not represent broader market patterns.

That is a reasonable piece of qualitative research. It is not a measurement of what percentage of the world’s AI projects fail, and it does not claim to be. The claim got that big on the way down the citation chain.

The primary source is not at its own address

This is the part that decided it for us.

The report’s PDF was published at nanda.media.mit.edu/ai_report_2025.pdf. Request that address today and you do not get a PDF. You land on the NANDA group’s overview page at the MIT Media Lab.

That page describes NANDA’s current work on decentralised AI and the agentic web. It links two arXiv papers on index architecture and a perspective paper on decentralised AI. It does not mention The GenAI Divide, it does not mention State of AI in Business 2025, and it does not mention a 95% figure anywhere. Reports are now behind a request form.

The copies of the report still circulating are hosted by third parties who mirrored the file, not by the publisher. We could not find an Internet Archive snapshot of the original PDF address.

We are not asserting anything about why the file moved. We do not know, and guessing would be exactly the behaviour this post is about. What we can say is narrow and checkable: as of late August 2026, the number cannot be traced to a document served by the institution credited for it.

The 5% is narrower than the way it gets quoted

There is a second problem, and it is the one that shows up most often in the wild.

The figure is usually inverted for a more optimistic framing: only about 5% of businesses are getting real financial value from AI. We have heard it phrased that way in pitches and, recently, in a well produced video from a large channel.

That restatement is broader than what the report supports. The nearest supported 5% in the underlying research refers to integrated AI pilots, and specifically to custom enterprise tools, rather than to all businesses and all AI. Widening it from a category of deployments to the entire population of companies is a meaningful change, and it happens silently.

Kevin Werbach, Professor and Chair of Legal Studies and Business Ethics at Wharton, made the sharper version of this criticism in trade coverage in August 2025: there appears to be no further support for the 95% claim anywhere in the report’s documentation. The same coverage raised unlabelled chart axes, undefined terms and a conflation of non adoption with failure.

What we use instead

We did not stop making the argument. The argument is correct and it is why this firm exists. Most organisations spending on AI cannot show what it returned.

We changed the evidence.

PwC, 29th Annual Global CEO Survey, published 19 January 2026. 4,454 CEOs across 95 countries and territories, fieldwork 30 September to 10 November 2025. Only one in eight CEOs, 12%, say AI has delivered both cost and revenue benefits. 33% report gains in either cost or revenue. 56% say they have seen no significant financial benefit to date.

Same argument. Named publisher, disclosed sample, disclosed field window, and a document that is still where it says it is.

We also use the U.S. Census Bureau’s Household Trends and Outlook Pulse Survey from March 2026, published 11 August 2026, because it measures something the executive surveys cannot. Among workers who used AI at work in the week they were surveyed, 31% finished one to two hours faster and 15% saved more than four hours. But 10% saved no time at all, and a further 3% said AI made the work take longer.

That last pair is the whole problem in one dataset. The same category of tool returns half a day a week to one person and takes time away from another. Nothing about the tool tells you which one you bought. You have to measure it.

Why we are telling you this

Partly because we got it wrong and the correction belongs in public. We put an unverifiable number in front of prospects for months.

But mostly because of what it implies about the rest of the pitch.

If a firm is selling you measurement discipline while quoting a statistic it has not opened, that is a live demonstration of the thing you are being asked to buy, and it is failing in front of you. The number is easy to check. It took us an afternoon.

So use it as a test. When someone quotes a figure at you, ask for the primary source rather than the article quoting it. Open it. Find the sample size, who was surveyed, when the fieldwork ran, and whether the number in the pitch is the number the study supports.

Do it to us too. Every figure on this site carries its publisher, its date and its sample size, and we will send you the source on request. If we ever cannot produce one, that is a reason to doubt everything else we have told you.