Skip to content
Almanet
Field Notes
Reading the sources14 September 2026 · 5 min read

Eighty percent of AI projects fail. Nobody measured that.

RAND is cited for “more than 80 percent of AI projects fail.” RAND did not measure that. It quoted it, with a hedge, from a magazine — and then warned that its own findings may be skewed. Both parts get cut.

If you have read anything about AI projects failing, you have read a number. More than 80 percent fail. Or 95 percent see no return. The numbers are real in the sense that they appear in real documents. Almost every time they are repeated, the part that would let you judge them has been removed.

The study everyone cites

The 80 percent figure is attributed to a RAND Corporation report — The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, by James Ryseff, Brandon De Bruhl and Sydne J. Newberry, report RR-A2680-1, completed April 2024.[1] Here is the sentence it is taken from, in full:

By some estimates, more than 80 percent of AI projects fail. This is twice the already-high rate of failure in corporate information technology (IT) projects that do not involve AI.

“By some estimates.” The sentence carries two footnotes. The first points to a Fortune article by Jeremy Kahn from July 2022, which says business leaders have put the failure rate “between 83% and 92%” on the strength of “recent surveys” it does not name.[2] The second, for “twice the rate,” points to a Harvard Business Review piece from late 2023.[3]

RAND did not measure a failure rate. It measured causes. Between August and December 2023 the researchers interviewed 65 data scientists and engineers, each with at least five years of experience building AI and machine learning models, selected across company sizes and industries. From those interviews they identified five leading root causes, in this order: the problem to be solved is misunderstood or miscommunicated; the organisation lacks the data to train an effective model; the organisation focuses on the latest technology rather than on solving real problems for its users; the infrastructure to manage data and deploy models is not there; and the problem is too difficult for AI to solve at all.

That is the finding. The number in front of it is somebody else’s estimate, passed along with a hedge.

The sentence that gets cut

RAND also wrote down a limitation of its own method:

Because the majority of our interviewees were nonmanagerial engineers instead of business executives, the results may disproportionately reflect the perspective of individuals who do not hold leadership positions. Thus, the results may be skewed toward identifying leadership failures.

That is the authors, about their own headline finding, in the report itself.

We read eight articles that cite the report, published between August 2024 and August 2026.[4] None of the eight includes that limitation. Four of the eight present the 80 percent as RAND’s own result — one gives it as “80.3 percent,” a decimal that exists nowhere in the report. The other four keep RAND’s “by some estimates,” and two of them say outright that RAND was repeating a number it did not measure. Even the careful ones drop the caveat.

What survives the retelling is the shape that travels well: a big number, an institution’s name behind it, and a villain. What gets dropped is everything a reader would need in order to judge whether the finding applies to them.

The same happens to the other numbers. The 95 percent figure — that 95 percent of organisations see no return from generative AI — comes from a July 2025 report by MIT’s Project NANDA whose cover reads “Preliminary Findings.”[5] It circulates as settled. The claim that AI failure “doubled” in a year traces to a survey by S&P Global Market Intelligence of more than a thousand companies in North America and Europe, in which 42 percent said they had abandoned most of their AI initiatives in 2025, up from 17 percent the year before.[6] That is a real and striking result. It is not a failure rate, and 17 to 42 is not a doubling.

Why this matters if you are the one deciding

A number without its context cannot tell you what to do. “Eighty percent fail” is frightening and useless. “Sixty-five engineers, asked what went wrong on projects they lived through, mostly named things that happened before any model was trained — and the authors warn that engineers were doing the naming” is less quotable and far more actionable.

Our reading of the five causes is that most of them are decided before the software exists: what problem you are solving, whether the data for it exists, whether the machinery to run it exists. That is our interpretation, not RAND’s wording. But it tells you where to look — at the part of your business that comes before the software.

What we see

The pattern we run into is consistent enough to describe without numbers, because we have not measured it.

A company wants to automate something. The intent is real and the budget exists. Then the work starts and the process turns out to be undocumented — it lives in the head of the person who does it, and that person does it slightly differently depending on the day. The data that would feed the automation is spread across a spreadsheet, an inbox, and a paper log, with no agreement on which is correct when they disagree.

Nothing here is an AI problem. It is the problem of a business that has grown without ever writing down how it works. The automation cannot be built on top of it, because there is no “it” — there is a set of habits that happen to produce the right answer most of the time.

The sequence that works is unglamorous. Find out what the process actually is, as opposed to what people say it is. Write it down. Decide where the data lives and what makes one copy authoritative. Then automate, on a foundation that exists.

The rule we apply to ourselves

Every number on this site carries a source, and where the source has a limitation we state it. Not because it is a nice gesture — because a number you cannot check is not evidence, it is decoration.

If we ever quote a study here and leave out what the study says about itself, hold us to the same standard.

Sources

  1. 1
    RAND Corporation, RR-A2680-1, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, James Ryseff, Brandon De Bruhl, Sydne J. Newberry, 2024. Quotations are from the PDF, retrieved 2026-09-14. rand.org/t/RRA2680-1
  2. 2
    Jeremy Kahn, “Want Your Company’s A.I. Project to Succeed? Don’t Hand It to the Data Scientists, Says This CEO,” Fortune, 26 July 2022. RAND’s footnote 13. fortune.com
  3. 3
    Iavor Bojinov, “Keep Your AI Projects on Track,” Harvard Business Review, November–December 2023. RAND’s footnote 14.
  4. 4
    Eight articles citing the RAND report, each read 2026-09-14 and checked for how it attributes the 80 percent and whether it carries RAND’s limitation on its interview sample. Two further articles could not be retrieved and are not counted.
  5. 5
    MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, July 2025, marked “Preliminary Findings.” We did not find an official address for the report on MIT’s sites; we quote from a copy on a third-party mirror (v0.1), and that is a limitation of the quotation.
  6. 6
    Lindsey Wilkinson, “AI project failure rates are on the rise: report,” CIO Dive, 14 March 2025, reporting an S&P Global Market Intelligence survey of more than 1,000 respondents in North America and Europe. The underlying S&P report is not publicly linked. ciodive.com
Almanet

AI building and automation for small and mid-sized businesses. Building since 2022.

team@almanet.ai

© 2026 Almanet. All rights reserved.Privacy

Live figures are research counts, read from the Wayseering public API.