Why AI projects fail: what Gartner, MIT, McKinsey and RAND actually found

Short answer: None of the four measured that “AI projects fail” the way it is quoted. Gartner’s numbers are mostly forecasts. MIT’s 95 percent is about generative AI with no measurable effect on the bottom line, based on a small sample. McKinsey measures how many respondents say AI affects earnings. And RAND did not measure a rate at all. But all four point to the same thing: it goes wrong between pilot and production, and when the problem is not clearly defined.

What the numbers really are

Source	Number	What it actually is
Gartner	30%	A 2024 forecast of generative AI projects abandoned after proof of concept
MIT	95%	Generative AI with no measurable bottom-line effect, from 52 organisations interviewed
McKinsey	37%	Share of respondents attributing some EBIT impact to AI, in 2026
RAND	80%	Not RAND’s number. Quoted from a 2022 Fortune article

These four numbers cannot be compared or combined. Articles claiming “70 to 85 percent of AI projects fail” do exactly that.

Gartner: forecasts, not measurements

In July 2024 Gartner predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025. The reasons given were poor data quality, inadequate risk controls, escalating costs and unclear business value. It was a forecast, with no published method.

In January 2026 Gartner wrote that at least 50 percent were in fact abandoned, based on analysing “hundreds of GenAI implementations”. How the number was calculated has not been published.

The third figure often quoted, that over 40 percent of agentic AI projects will be cancelled by 2027, is also a forecast. The survey attached to it asked 3,412 webinar attendees how much they had invested, not whether projects were cancelled.

Gartner’s pages block automated access, so we read the press releases through reprints of them.

MIT: 95 percent of what?

MIT NANDA’s report “The GenAI Divide”, July 2025, is based on a review of over 300 publicly disclosed AI initiatives, interviews with 52 organisations and responses from 153 senior leaders, collected from January to June 2025. It covers generative AI only.

The number comes from a funnel. Of the organisations that evaluated custom enterprise AI tools, 20 percent reached pilot and 5 percent reached production. General tools such as ChatGPT and Copilot were far more widely used, but according to the report they raised individual productivity, not the bottom line.

The authors are clear about the limits themselves:

The figures are based on interviews rather than official reporting, and the organisations willing to take part may differ from the rest.

Fortune spread the number under the headline “95% of generative AI pilots at companies are failing”, describing the sample as 150 interviews and a survey of 350 employees. The report says 52 organisations interviewed and 153 leaders surveyed.

The report’s own explanation is that most tools do not learn: they do not retain feedback or adapt to how the work is done.

McKinsey: self-reported impact

McKinsey’s annual survey asks leaders what AI does for earnings. In the 2026 edition, published on 25 August, 1,719 participants in 97 countries responded. 37 percent said AI has some effect on EBIT, about the same as the year before. Around 6 percent fell into the group McKinsey calls high performers.

What set them apart: nearly 3 in 4 high performers had redesigned workflows because of AI, against about 1 in 4 of the rest.

Note that these are respondents, not organisations, and the impact is self-reported. McKinsey’s pages would not load for us, so the 2026 figures come from The Register’s report on them.

RAND: causes, not rates

RAND interviewed 65 people with at least 5 years of experience in AI and machine learning, 50 from industry and 15 from academia, and published the report in August 2024. Projects that only used pretrained language models with prompts were excluded.

RAND measured no rate. The sentence people quote says that “by some estimates” more than 80 percent of AI projects fail, citing a 2022 Fortune article. That article in turn cites surveys it does not name.

What RAND actually found was 5 root causes:

Leaders misunderstand, or miscommunicate, what problem AI should solve.
The organisation lacks the data needed.
The focus is on new technology rather than on solving a real problem.
The infrastructure for data and deployment is missing.
The problem is too hard for AI.

The first two were raised unprompted by more than half of the interviewees.

Where they agree

The loss happens between pilot and production. Trying AI is not what fails. Getting it into ordinary use is.
An unclear problem or unclear value is a leading cause. Gartner, RAND and MIT each say so in their own way.
Changing the work matters more than adding a tool. McKinsey and MIT both point to redesigned workflows.

What it means for you

Decide what counts as success before the pilot starts, and measure it. Pick a problem that will still exist in a year. And stay open to the answer not always being AI, which we cover in the article on when the answer is not AI. How to measure whether an AI system actually works is in the article on evaluation.

Sources

All sources checked on 14 September 2026.