How to check an AI failure statistic before it justifies a purchase
Four failure statistics are quoted at buyers constantly. We traced each one to its source and recorded what we could read and what we could not. This is the method, so you can run it on whatever lands in your inbox.
Why the number gets into the room before the workflow does
Most AI proposals now open with a failure statistic. It works two ways. A vendor uses it to say that most attempts go wrong and theirs is the exception. An internal sponsor uses it to show the board that the risk is understood. Both uses move money, and the figure usually arrives without anyone having opened the document it came from.
We spent an evening tracing four of the numbers in widest circulation. This page is the method and what it produced.
The six questions
Ask them in this order.
- Who published it? Name the organization and the document. The blog post that quoted it does not count.
- What is its date? A 2022 estimate about machine learning projects answers a different question from a 2025 survey about generative AI.
- Is it a forecast, a survey or a measurement? A forecast is an analyst's view of the future. A survey records what respondents said about themselves. A measurement counts outcomes. All three are useful and they do not mean the same thing.
- What was the unit? Organizations, projects, pilots and proofs of concept are four different denominators. Most of the drift happens right here.
- What was the sample? Size, region, seniority, and how people were recruited.
- Can you open the original? A figure behind a login, or gone from the address where it first appeared, deserves less weight than a document you can read.
What the four numbers turned out to be
The Gartner forecast about agentic AI
Gartner published a press release on June 25, 2025 predicting that over 40 percent of agentic AI projects will be canceled by the end of 2027, because of rising costs, unclear business value and inadequate risk controls. The release is clear that this is a prediction, it attributes it to a named analyst, and it carries a date. It also reports a January 2025 Gartner poll of 3,412 webinar attendees about how much they had invested in agentic AI.
We could not open Gartner's own newsroom page, which answered our requests with an automated access check, so we read the release through two sites that republish it in full. Both carry the same date and wording.
Watch what happens in retelling. One trade article we read describes the same forecast as resting on a poll of more than 3,400 organizations actively investing in the technology. Webinar attendees and organizations actively investing are different groups. No dishonesty is required for that shift. It happens in ordinary summarizing.
The S&P Global figure about proofs of concept
S&P Global Market Intelligence published an analysis on March 12, 2025 reporting that the average organization scrapped 46 percent of its AI proofs of concept before they reached production, and that the share of companies abandoning most of their AI work had risen to 42 percent from 17 percent a year earlier. The underlying research is its Voice of the Enterprise survey on AI and machine learning, reported as 1,006 respondents in North America and Europe.
This is a number presented properly. The publisher, the survey, the sample size, the regions and the date are all stated, and it is described as a survey of what respondents said about their own organizations. Quote it and everyone knows what they are holding.
One limit on our own check. S&P's page refused our requests, so we took the figures and the sample from its published summary and from trade reports of the release. We have not read the full note.
The figure attributed to MIT
This document exists and is readable. It is a report from MIT NANDA dated July 2025, by four named authors, whose executive summary says that 95 percent of organizations are getting zero return. Its front matter calls it preliminary findings, gives a research period of January to June 2025, and describes a method of reviewing over 300 publicly disclosed AI initiatives, interviewing representatives of 52 organizations, and collecting 153 survey responses at four industry conferences.
Two things follow. The first is the unit. The report's own sentence counts organizations getting zero return from their investment. The version that reaches executives is usually about pilots failing, which is a different denominator counted a different way. The second is that the report is franker about its limits than its retellings are. It names selection bias among organizations willing to discuss AI, it notes that success definitions varied, and it says its six-month observation window may be too short for complex systems and could understate success rates. The limitations page describes a more careful document than the headline suggests.
What we could not do is read it where it was first published. The address it was originally posted at now redirects to a research group overview page, so we read a copy hosted elsewhere. We cannot say from MIT why that is, and we are not going to guess.
The figure attributed to RAND
The RAND document people cite is a 2024 research report on the root causes of AI project failure, and it measures no failure rate at all. In that 2024 RAND report the figure is an aside in the introduction, which says that by some estimates more than 80 percent of AI projects fail, with an endnote crediting a magazine column.
We opened the column. Published in Fortune on July 26, 2022, it says business leaders have put the failure rate of AI projects somewhere between 83 and 92 percent, and credits a group of recent surveys without naming any of them. That is where the trail stops.
None of this makes the estimate wrong. It means nobody in the chain we could follow measured it, and that it describes projects from before the current generation of tools. What RAND did in that report is interview 65 experienced practitioners between August and December 2023 about why projects fail, defining failure as a project the organization itself perceived as failed. That is careful work about causes, quoted for a rate it never claimed to establish.
The worksheet
Take this to the next proposal that quotes a number.
| Question | Where to look | What a weak answer looks like |
|---|---|---|
| Publisher and document title | The slide's footnote, then the publisher's own site | A link to a vendor blog that links to another vendor blog |
| Date of publication | The document itself; the page quoting it does not count | No date, or only the date of the article quoting it |
| Forecast, survey or measurement | The document's own description of its method | The word "research" with no method named |
| Unit counted | The sentence the figure appears in | The unit changes between the headline and the body |
| Sample and recruitment | A methodology or limitations section | A sample size with no description of who was in it |
| Can you open it | Try the original address yourself | Login wall, paywall, or a dead address |
If four of those six come back weak, treat the figure as decoration on somebody's argument. Ask the firm for a number they can source, or go ahead without one.
Where we sit in this
We sell AI integration work. A buyer who believes almost every AI project fails is a buyer who feels urgency, and urgency is good for anyone selling a first engagement. We benefit from these statistics being repeated. That is exactly why we keep them off our own pages and out of our proposals, and why the only numbers on this site come with a named source and a date attached.
Apply the six questions to us too. If we ever quote a figure at you and cannot answer them, do not accept the number.
What to use instead of a rate
None of these figures tells you whether your project will work. Your own baseline does. Measure the workflow before anything is built, write down what a passing result looks like, and name the person who decides. Evidence about your own company carries more weight in a board meeting than any industry figure.
When you compare firms, our guide on what to ask an AI consulting firm includes the question about sourcing claims.
If you want to talk about a specific workflow, tell us which one, who does it today, and what a passing result would look like.