95%, 80%, 40%: The Three AI Failure Numbers Everyone Quotes, and What Each One Actually Measures

Sources: MIT Project NANDA, The GenAI Divide: State of AI in Business 2025; RAND, The Root Causes of Failure for Artificial Intelligence Projects (2024); Gartner, press release, June 25, 2025.

Three numbers, three different things

If you've sat through an AI pitch or a board meeting this year, you've heard at least one of them. "95% of AI projects fail." "80% of AI projects fail." "40% of AI agent projects will be scrapped." They get quoted as if they were the same finding at different strengths. They aren't. One is a measurement from a modest sample, one is a borrowed estimate, and one is a forecast. If you're deciding whether to spend money on AI, the difference matters, because each one tells you something different about your own odds.

95%: MIT's measurement, from a modest sample

The 95% comes from MIT Project NANDA's report The GenAI Divide: State of AI in Business 2025. Its executive summary says that despite an estimated $30–40 billion in enterprise investment into generative AI, "95% of organizations are getting zero return," and that just 5% of integrated AI pilots are extracting millions in value while the rest show no measurable P&L impact.

What it actually measured: the report describes 52 structured interviews, a review of more than 300 public AI initiatives and surveys of 153 leaders. Success was defined as deployment beyond the pilot phase with measurable KPIs, with ROI measured six months after the pilot. The report's own limitations section warns that the sample may not represent every segment and that organisations willing to discuss AI may differ from those that declined.

So the honest version is: in MIT's sample, about 95% of organisations saw no measurable P&L return from integrated generative AI pilots within roughly six months. That is a real, useful finding. It is not proof that 95% of all AI everywhere fails. We go through the mechanisms in our breakdown of the MIT AI report.

80%: an estimate RAND quoted, not a finding RAND made

The 80% figure is usually credited to RAND. RAND's August 2024 report does say it: "By some estimates, more than 80 percent of AI projects fail—twice the rate of failure for information technology projects that do not involve AI." Look at the footnote and the source is a 2022 Fortune article, not RAND's own research.

RAND's own work was different and, for an owner, more useful. Researchers interviewed 65 experienced data scientists and engineers about why AI projects fail. They found five root causes: the problem was misunderstood or miscommunicated; the organisation lacked the right data; teams chased the latest technology rather than the user's problem; infrastructure was missing; and the problem was simply too hard for AI. Note the scope too: RAND focused on machine learning projects and excluded projects that only used pretrained language models through prompting, which is how most small firms use AI today.

Quote RAND for the five causes. Don't quote it for the 80%.

40%: Gartner's forecast about AI agents

The 40% is a prediction. On June 25, 2025, Gartner said "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." It's an analyst forecast about the future, published alongside a poll of 3,412 webinar attendees about how much they had invested in agentic AI. The poll measured investment posture, not cancellations.

The same release makes two points that matter more to a small firm than the headline. Gartner warned about "agent washing", the rebranding of existing assistants, automation bots and chatbots as agents, and estimated only about 130 of the thousands of agentic AI vendors are real. Gartner also said many use cases now pitched as agentic do not need an agentic implementation at all. If someone is selling you an agent, our agent vs automation test is a good place to start.

Side by side

• 95% (MIT): a measurement. Share of organisations in a 52-interview sample with no measurable P&L return from integrated GenAI pilots about six months in.
• 80% (quoted by RAND): an estimate attributed to "some estimates", traced to a 2022 press article. Not RAND's own finding.
• 40% (Gartner): a forecast. Agentic AI projects expected to be canceled by the end of 2027.

All three point the same way: most AI efforts don't pay off the way the pitch promised. That part is consistent. The precise percentages are not interchangeable, and none of them is your number.

What to do with this as an owner

These figures describe other organisations. Your odds depend on choices you control, and the three sources agree on most of them:

1. Start from a problem, not a tool. RAND's top root cause and MIT's pattern of tools that don't fit the workflow are the same lesson. Score your candidate tasks with the AI use case scorer.
2. Take a before number. MIT counted success as measurable KPIs. You can't show a return you never measured. The AI ROI calculator works it out per finished task.
3. Check the data and the plumbing. RAND's second and fourth causes. If the inputs are scattered, map the process first.
4. Be suspicious of agents you don't need. Gartner's own warning. Plain automation is cheaper and more predictable for fixed steps.
5. Set kill criteria. An end date and a decision owner turn a "failure" into a cheap lesson.

If you'd rather see where you stand before spending anything, take the 20-minute AI readiness assessment, or read AI hype vs reality for what tends to pay off in a small firm.

Frequently asked questions

Is it true that 95% of AI projects fail?

Not quite. MIT's GenAI Divide report found about 95% of organisations in its sample got no measurable P&L return from integrated generative AI pilots, measured six months after the pilot. That's a narrower claim than '95% of AI projects fail'.

Where does the 80% AI failure rate come from?

RAND's 2024 report says 'by some estimates, more than 80 percent of AI projects fail' and cites a 2022 Fortune article for it. RAND's own research was 65 interviews about why projects fail, not a count of failures.

Is Gartner's 40% agentic AI figure a measurement?

No. It's a forecast: Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing cost, unclear value and weak risk controls.

Which AI failure statistic should I use?

None of them can tell you whether your own project will pay. Measure cost per finished task before and after on one task in your business.