About half of AI projects fail. Counted per project, between 46% and 54% did not reach production in five studies published from 2024 to 2026. The 80% and 95% figures that lead most articles are not project failure rates: one is an unnamed estimate that reached a RAND report through a magazine interview, and the other counts every organization studied, including the four in five that never piloted a task-specific tool.
The 46% to 54% range is an EltexSoft calculation, made in November 2026. The method: take every figure I could trace to a primary document and keep the ones that pass three tests. The figure is measured, not forecast. The primary document prints a share of projects or proofs of concept, not of companies. The share is about reaching production. Each is then restated as the share that did not reach production. Five qualify. Two of the five are floors, Gartner's "at least 50%" and IDC's more than 50%, so 46% to 54% is the span of the printed values and the top is not a ceiling. Two figures covered below are left out: the 88%, because its source prints no share and states no base, and Gartner's 30%, because it was a forecast. The five come from different studies with different questions, so the range is a reconciliation, not a single statistic.
EltexSoft builds AI systems for clients, and we have quoted the 95% ourselves. So I traced each number back to the page where it first appeared.
The Five Figures That Count Projects
52%, Gartner, late 2023. A Gartner survey of 644 respondents in the US, Germany and the UK found that "on average, only 48% of AI projects make it into production". The other 52% do not.
46%, S&P Global, late 2024. S&P Global's 451 Research surveyed 1,006 IT and business professionals and found "organizations on average reporting that 46% of projects are scrapped between proof of concept and broad adoption". The survey ran from October 21 to November 25, 2024. Its more quoted result, that the share of companies abandoning most of their AI initiatives rose from 17% to 42%, counts companies.
At least 50%, Gartner, end of 2025. In January 2026 Gartner wrote that "by the end of last year, at least 50% of generative AI projects were abandoned after proof of concept". The page says Gartner "analyzed hundreds of GenAI implementations" and gives no sample size or method. The figure covers generative AI only.
54%, IDC for Lenovo, late 2025. Lenovo's CIO Playbook 2026, an IDC survey of 3,120 IT and business decision makers, reports that "Nearly half (46%) of AI proof-of-concepts have already progressed into production". That leaves 54% that had not, some of them still in progress.
More than 50%, IDC, July 2026. A slide in an IDC conference presentation on AI agents states that "<50% of proofs of concept convert to production" and names its source as IDC's Future Enterprise Resiliency & Spending Survey, Wave 4, July 2026. More than 50% did not convert. The slide gives no sample size and no exact figure.
Four of the five come from surveys, two are floors, and a proof of concept that is not in production is not always dead. The five still agree on the size: about one project in two.
Where 80% Comes From
The usual citation is RAND's 2024 report on why AI projects fail: "By some estimates, more than 80 percent of AI projects fail." RAND did not measure that. Its own research was 65 interviews with data scientists and engineers about causes. The footnote on the sentence points to a 2022 Fortune article, an interview with the CEO of an AI software company. It says that in "a slew of recent surveys" business leaders "have put the failure rate of A.I. projects at between 83% and 92%". Fortune names none of the surveys and links none. The trail ends there.
RAND's next sentence says this is twice the failure rate of IT projects that do not involve AI. That footnote leads to a 2023 Harvard Business Review article, which says some estimates place the failure rate "as high as 80%", which it calls "almost double the rate of corporate IT project failures a decade ago". It gives no source for either half. Between the two documents "almost double" became "twice", and a comparison with a decade ago became a comparison with today. I found no study that measures AI projects and other IT projects on one definition in the same period.
What MIT's 95% Counts
The MIT NANDA report of July 2025 says "95% of organizations are getting zero return" from generative AI and, in the same paragraph, that "Just 5% of integrated AI pilots are extracting millions in value". Fortune's headline went with pilots: "95% of generative AI pilots at companies are failing". The report's own chart counts organizations.
The chart is a funnel. Of the organizations studied, 60% evaluated a custom or task-specific AI tool, 20% reached a pilot, and 5% reached production with a sustained effect on productivity or profit. Measured against the organizations that ran a pilot, 5 in 20 succeeded. That is a pilot failure rate of about 75%, not 95% (EltexSoft calculation from the report's own chart). The report prints the same kind of ratio for generic LLM chatbots: "high pilot-to-implementation rates (~83%)". The 95% also holds the 40% of organizations that never evaluated a tool and the 40% that evaluated one and never piloted it. Only 15 of the 95 points are pilots that stopped.
The report rests on interviews at 52 organizations, a survey of 153 leaders and a review of more than 300 public announcements, and it calls its figures "directionally accurate". The MIT address that hosted it now redirects to a page that does not mention the report. The link above is an archived copy.
Predictions and One Division
Three more numbers circulate as failure rates.
85%. A Gartner prediction from February 2018: "through 2022, 85 percent of AI projects will deliver erroneous outcomes due to bias in data, algorithms or the teams responsible for managing them". It forecast biased results, not failure, and its window closed four years ago.
88%. The IDC and Lenovo playbook for 2025 shows two numbers in a graphic, 33 AI proofs of concept and 4 production launches, and no percentage. CIO.com reported them as "88% of observed POCs" that do not reach wide deployment. That division is not in the range: the playbook prints no share and states no base for the 33. A year later the same sponsor and research firm report 46% in production, from a different survey.
30% and 40%. In July 2024 Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. Its own later figure, above, is at least 50%. Its other forecast, that more than 40% of agentic AI projects will be cancelled by the end of 2027, was repeated in October 2026 and is still a forecast.
Failing Outright Is the Smaller Part
"Fail" in these surveys mostly means stopped before production. When Gartner asked 782 infrastructure and operations leaders in late 2025, 28% of AI use cases were reported to "fully succeed and meet ROI expectations, while 20% fail outright". The remaining 52%, by subtraction, sit in between: running, and short of what was promised.
Company-level numbers are moving too. BCG reported in 2024 that 74% of companies had yet to show tangible value from AI. In September 2026 it says nearly half now generate value, and adds: "It is no longer accurate to say most companies get nothing from AI." Whether a company earns a return is a different question from whether a project ships, and we cover the return question separately.
What to Do With One in Two
At 95%, the sensible decision is not to start. At one in two, the decision is how to land in the right half, and that is an engineering question. Gartner's list of what kills generative AI projects after the proof of concept is "poor data quality, inadequate risk controls, escalating costs or unclear business value". None of the four is a model problem. A Gartner survey of 2,002 CIOs and technology executives in 2026 found that only 31% have guidelines for deciding when a pilot should not move to production. Most of the CIOs surveyed have no rule for when a pilot should stop.
When someone quotes you an AI failure rate, ask three things. Does it count projects or companies? Was it measured or predicted? What year was the survey in the field? We have written about why generative AI projects stall after the demo, and the short version is that a demo is not a system. If the project has to reach production, look at how we build generative AI systems.