The AI project failure rate, reconciled to its sources
AI project failure estimates range from more than 80% of projects (RAND, 2024) to 95% of organisations getting zero return from generative AI (MIT, 2025), while BCG found only 5% of companies achieving value at scale. These figures are not in conflict and they are not comparable: they count different things, over different populations, and two of the most quoted are forecasts rather than measurements.
By Sunny Patel published figures last checked
The number, and what it is a number of
Asked plainly, the best-supported answer is RAND's: by some estimates, more than 80% of AI projects fail, roughly twice the rate of IT projects that do not involve AI. Notice how much hedging survives in that sentence. RAND is not reporting its own count. It is characterising other people's estimates, in a report whose actual method was 65 semi-structured interviews.
That hedge does not survive contact with the internet. The figure you have almost certainly seen is 80.3%, complete with a decimal place, often accompanied by a tidy four-way breakdown of outcomes that sums to exactly 100.0%. Neither appears anywhere in RAND's report. We looked.
Why the estimates disagree
They do not, really. They are answers to different questions, and they are routinely stacked against each other as though they were rival measurements of the same thing. Here is every headline figure in this field, next to what it actually counts.
| Source | The figure, in its own words | Unit counted | What it actually measures | Sample |
|---|---|---|---|---|
| RAND 2024 | By some estimates, more than 80% of AI projects fail, about twice the rate of non-AI IT projects | projects | Projects that fail. RAND is quoting other estimates, hence "by some estimates". | 65 interviews (50 industry, 15 academic) |
| MIT Project NANDA 2025 | 95% of organisations are getting zero return from generative AI | organisations | Organisations getting zero return from generative AI. Not projects, and not "pilots failing". | 52 organisations interviewed, 153 survey leaders, 300+ public initiatives reviewed |
| Gartner 2024 | At least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs or unclear business value | projects (forecast) | A prediction, issued in 2024, about what would be true by the end of 2025. | not published |
| Gartner 2025 | Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls | projects (forecast) | A prediction, issued in 2025, about what will be true by the end of 2027. | not published |
| BCG 2025 | 5% of companies are achieving AI value at scale, and 60% are not achieving material value at all (stagnating 14%, emerging 46%, scaling 35%, future-built 5%) | companies | Companies achieving AI value at scale. A maturity classification, not a project outcome. | 1,250 executives across 68 countries and 25+ sectors |
| IBM 2025 | CEOs report that only 25% of AI initiatives have delivered expected ROI over the last few years, and only 16% have scaled enterprise wide | initiatives | Initiatives judged against the ROI their own CEOs expected of them. Self-reported. | 2,000 CEOs across 33 countries and 24 industries |
| McKinsey 2025 | 88% report regular AI use in at least one business function, 72% use generative AI, only 7% have fully scaled it, and just 39% report enterprise-level EBIT impact | organisations | Organisations that have fully scaled AI, and those reporting enterprise EBIT impact. | 1,993 respondents across 105 countries |
| Deloitte 2025 | 74% of organisations are hoping to grow revenue through their AI initiatives, compared with just 20% that are already doing so | organisations | Organisations already growing revenue through AI, against those hoping to. | not published |
Compiled by implementai.today from the sources named. We ran no survey. Machine-readable version: /data/ai-failure-rates.json.
Three distinctions do most of the work. Projects are not organisations. RAND counts projects; MIT and BCG count companies. A company can run five AI projects, fail four, and still appear in the successful 5%. Failure is not the absence of return. A project can ship, work, and still show nothing on the profit and loss statement, which is what MIT's 95% describes. And a forecast is not a measurement. Both Gartner figures are predictions, written before the years they describe.
The figures that were never published
This is the part nobody else does. Of the 15 sources this site cites, 9 circulate in a form the original does not support. Some of the drift is trivial. Some of it is invention.
The mechanism is worth understanding, because it will keep happening. Primary sources block
automated readers: rand.org, gartner.com and hbr.org all
return an error to anything that is not a browser. So the AI assistants and content mills that
quote them cannot actually read them. They quote each other instead, and precision accretes with
every hop. A hedge becomes a number. A number acquires a decimal. An opinion survey of
65 people becomes an audit of an industry.
| What circulates | What the source actually says | Source |
|---|---|---|
| "80.3% of AI projects fail", and a four-way outcome breakdown crediting RAND with 33.8% abandoned, 28.4% no value, 18.1% partial, 19.7% succeed | RAND publishes no decimal figure. Its words are "by some estimates, more than 80 percent", a hedge on someone else's estimate. The four-way breakdown appears nowhere in the report, and could not: RAND's method was 65 semi-structured interviews, which cannot produce an outcome taxonomy. The four numbers sum to exactly 100.0%, which is what invented data tends to do. | RAND 2024 n=65 interviews (50 industry, 15 academic) |
| "84% of AI project failures are caused by leadership, not technology" | The 84% is real, and it is narrower than that. It is the share of RAND's 65 interviewees who named a leadership-driven cause as the primary reason AI projects fail. It is expert opinion from 65 people, not an audit of how failures were caused. Confusingly, a second and unrelated 84% appears in the same report, taken from the Cisco AI Readiness Index, describing business leaders who believe AI will significantly affect their business. That coincidence is probably where the muddle began. | RAND 2024 n=65 interviews (50 industry, 15 academic) |
| "95% of AI pilots fail", and "general-purpose tools succeed 83% of the time against 5% for custom tools" | The 95% describes organisations getting zero return, not pilots failing. The 83% and the 5% are not comparable: 83% is a pilot-to-implementation rate for general-purpose chatbots, while the 5% is the share of all organisations that evaluated a custom tool and reached production. On like-for-like pilot-to-implementation, the report's own chart puts general-purpose tools near 80% and embedded custom tools near 25%. The sample is also inflated in retelling: it is 52 organisations interviewed, 153 survey leaders and 300+ public initiatives, not "150 interviews and a 350-employee survey". | MIT Project NANDA 2025 n=52 organisations interviewed, 153 survey leaders, 300+ public initiatives reviewed |
| "Gartner found that 30% of generative AI projects are abandoned" | This is a forecast, published in July 2024, about what would happen by the end of 2025. Gartner measured nothing. A prediction quoted as an observation is the most common error in this field. | Gartner 2024 no sample published |
| "40% of agentic AI projects have been cancelled" | A forecast issued in June 2025 about the end of 2027. Gartner also notes that only a small fraction of vendors claiming agentic capability actually offer it, a practice it calls agent washing. | Gartner 2025 no sample published |
| "88% of companies use AI", cited as evidence that AI adoption is solved | The same survey reports that only 7% have fully scaled AI and nearly two thirds have not begun scaling at all. Regular use in one function is a low bar. Adoption and value are different measurements. | McKinsey 2025 n=1,993 respondents across 105 countries |
| "Only 4% generate substantial value" and "26% have capabilities to move beyond experimentation" | The 4% is BCG's 2024 figure; the 2025 report puts future-built firms at 5%. No 26% figure appears anywhere in the 2025 report. It is not a BCG category. | BCG 2025 n=1,250 executives across 68 countries and 25+ sectors |
| "46% of small firms cite lack of knowledge as the barrier", credited to DSIT | That figure comes from a Federation of Small Businesses survey, not this study. DSIT's own barrier figure is 60% of businesses citing limited skills or expertise. | DSIT 2025 n=3,500 businesses with 5+ employees, plus 100 qualitative interviews |
| "UK AI adoption rose from 9% to 25%" | ONS states a rise of 15 percentage points to 25%, which puts the baseline at 10%, not 9%. A small error, and it is in almost every retelling. | ONS 2025 n=the BICS business panel |
The 84% that means something narrower than you think
Worth dwelling on, because it is the most instructive case. RAND's report contains the sentence: "Eighty-four percent of our interviewees cited one or more of these root causes as the primary reason that AI projects would fail." Those root causes are the leadership-driven ones. So the honest gloss is that 84% of 65 interviewed experts pointed at leadership. That is a survey of opinion, and a small one. It is not a finding that 84% of real failures were caused by leadership, which is how it now travels.
Then it gets worse. The same report contains a second 84%, drawn from the Cisco AI Readiness Index, describing business leaders who believe AI will significantly affect their business. Two unrelated 84% figures, one document. You can watch the confusion form.
What to do with this
If you are building a business case, quote RAND for the comparative claim (AI projects fail about twice as often as other IT projects) and quote nothing to two significant figures. If someone hands you a deck with 80.3% on it, you now know they did not read the source, and you can reasonably ask what else in the deck came from a content mill.
And if you want to know whether your own project carries the conditions these studies keep finding, the free diagnostic walks the eight factors, showing which study each one comes from. It cannot predict anything. Nothing can. It can tell you which of the documented risks you are currently carrying.
Questions, answered from the sources
What percentage of AI projects fail?
RAND wrote in 2024 that, by some estimates, more than 80% of AI projects fail, about twice the rate of non-AI IT projects. That is the most defensible answer to the question as asked. The widely repeated figure of 80.3% appears in no study, and the four-way breakdown often attributed to RAND was never published by RAND.
Why do the estimates differ so much, from 80% to 95% to 5%?
Because they measure different things. RAND counts projects that fail. MIT counts organisations getting zero return from generative AI. BCG classifies companies by maturity, finding 5% achieve value at scale. IBM asks CEOs whether initiatives met the ROI expected of them. Comparing these numbers to one another is a category error.
Which figure is most reliable?
For the question "do AI projects fail more than other IT projects", RAND, because it is the only source addressing that comparison directly, and it says roughly twice as often. For anything about generative AI specifically, treat every figure as contested: the most cited one is not peer reviewed and is not hosted on an MIT domain despite being universally called an MIT report.
Is the 30% Gartner figure a measurement?
No. It is a forecast published in July 2024 predicting that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. Gartner measured nothing. Its 40% agentic AI figure is likewise a forecast, about 2027.
Do AI projects fail more often than other IT projects?
RAND says so: more than 80% of AI projects fail, which it describes as twice the failure rate of information technology projects that do not involve AI. This is the single clearest comparative claim in the literature, and it is a hedged estimate rather than a measurement RAND performed.
Related: why AI projects fail, the causes rather than the count. And AI adoption statistics, where the UK numbers disagree with each other by a factor of two.