The headline numbers, and what each one actually measured
AI failure statistics travel further than their methodologies. Boards hear "95% of AI projects fail" in one meeting and "30%" in the next, and both speakers are quoting real research. The numbers differ because they measure different things over different windows. Here is the evidence base as it actually stands, with the caveats the headlines drop.
RAND Corporation (August 2024) reported, drawing on prior estimates and 65 structured interviews with experienced AI practitioners, that more than 80% of AI projects fail, roughly twice the rate of IT projects that do not involve AI. RAND's contribution is qualitative depth rather than a new survey: the figure is literature context, the interviews are the evidence.
S&P Global Market Intelligence (fieldwork late 2024, published March 2025; 1,006 respondents across North America and Europe) found the share of companies abandoning most of their AI initiatives jumped from 17% to 42% in a year, and that on average 46% of proofs of concept were scrapped before reaching production.
Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs and unclear business value. Gartner's subsequent analysis suggests the realised rate exceeded half. Its separate 2025 prediction: more than 40% of agentic AI projects will be cancelled by the end of 2027.
MIT's Project NANDA (August 2025) produced the most quoted and least caveated number: 95% of generative AI pilots showing no measurable profit-and-loss impact. The caveats matter. Failure was defined as no measurable P&L impact within six months, a short horizon for enterprise change; the survey sample was small and drawn from conference attendees; and the authors describe the figures as directional. Quote it, but quote it whole.
BCG (September 2025; 1,250+ senior executives) found about 5% of companies achieving AI value at scale and roughly 60% achieving no material value. McKinsey's State of AI (November 2025; 1,993 respondents) found 88% of organisations using AI somewhere, but only 39% reporting any enterprise-level EBIT impact. Adoption is nearly universal. Value is not.
The honest synthesis: most AI projects do not return their investment, the abandonment rate roughly doubled between 2024 and 2025, and the gap between adopting AI and getting value from it is the defining pattern. No serious source disputes the direction.
Six failure modes the evidence keeps finding
The same research that produces the failure rates also explains them, and the explanations converge.
First, the problem was never defined: business leaders and technical teams described success in incompatible terms, so the system solved a problem nobody had. RAND ranks this misalignment first among root causes. Second, the data was not there: the volume, quality or permissions the model needed did not exist, and nobody checked before procurement did. Third, technology went looking for a problem: the tool was chosen before the use case, and fit was presumed rather than established. Fourth, the demo never survived production: systems that worked in isolation broke against real workflows, legacy systems, approval chains and exception handling. Fifth, the problem was harder than the tools: the use case demanded reliability that current models cannot supply, and no confidence threshold was set to catch the shortfall.
The sixth failure mode runs underneath the other five: a governance vacuum. No named owner. No success metric agreed before the spend. No kill criteria. No evidence trail that would let anyone say, six months in, whether the project was working. The National Audit Office's review of AI in UK government found exactly this shape: enthusiasm for deployment, thin post-implementation evaluation. A project without a defined finish line cannot fail loudly, so it fails slowly, which costs more.
The board questions that prevent a repeat
Failure rates are not an argument against AI. They are an argument for governing it like any other material investment, which most of the failed projects were not. Six questions, asked before money moves, close off most of the failure modes above.
What decision or workflow does this change, and how would we measure that in twelve weeks? Who is the named senior owner, and what evidence will they bring to each board meeting? What data does this need, and have we verified it exists, at quality, with a lawful basis? What are the kill criteria, agreed now, that would stop the project? What happens when the system is wrong, and who catches it? And if a regulator, auditor or affected person asked us to explain an output, what record would we produce?
A project that cannot answer these questions has not earned its budget. A project that can answer them has, by most published measures, already separated itself from the majority that fail. If the project has already failed, the same six questions become the diagnosis: the answer that was missing is usually the failure mode that killed it.