The short answer
The AI failure numbers quoted everywhere measure different things and still agree. MIT NANDA: 95 percent of organizations see no P&L return from generative AI. IDC: 88 percent of proofs of concept never deploy. RAND: over 80 percent of AI projects fail, twice the IT norm. S&P Global: 42 percent of companies abandoned most AI initiatives in 2025. Gartner: 30 percent of GenAI projects dead after proof of concept, and over 40 percent of agentic projects canceled by 2027. Same direction, same causes: approach, not models.
Every pitch deck in the industry quotes one of these numbers, usually without saying what it measured. This page is the version we wanted to exist: every statistic, its exact claim, its source, and how they fit together.
The numbers, side by side
| Number | The exact claim | Who measured it | Population |
|---|---|---|---|
| 95% | Organizations investing in GenAI seeing no measurable P&L return | MIT NANDA, 2025 | 150 leader interviews, 350 employee surveys, 300 public deployments |
| 88% | AI proofs of concept that never reach widescale deployment (~4 of 33 graduate) | IDC with Lenovo, 2024 | Enterprise PoC portfolios |
| >80% | AI projects that fail, roughly twice the rate of non-AI IT projects | RAND, 2024 | Interviews with engineers and scientists on failed projects |
| 42% | Companies that abandoned most of their AI initiatives in 2025 (up from 17% in 2024) | S&P Global Market Intelligence, 2025 | 1,000+ enterprises, North America and Europe |
| 46% | Average share of AI proofs of concept scrapped before production, per organization | S&P Global Market Intelligence, 2025 | Same survey |
| 30% | GenAI projects predicted abandoned after proof of concept by end of 2025 | Gartner, July 2024 | Analyst projection |
| >40% | Agentic AI projects predicted canceled by end of 2027 | Gartner, June 2025 | Analyst projection |
What each study actually measured
MIT NANDA (95 percent) measured business return, not project completion: The GenAI Divide found the overwhelming majority of organizations getting nothing measurable from an estimated $30 to 40 billion of GenAI spending, and attributed the divide to approach rather than model quality. A system can ship and still land in this 95 percent if nobody's P&L moved.
IDC (88 percent) measured the pilot-to-production crossing: for every 33 proofs of concept started, about 4 deployed, with underfunded PoCs and missing business cases named as causes. This is the narrowest and most operational of the numbers, and the one our crossing playbook exists to beat.
RAND (over 80 percent) measured outright project failure and its root causes, by interviewing the people who lived the failures: misunderstood problems first among them. Its comparative finding is the sharpest: AI projects fail at roughly twice the rate of non-AI technology projects.
S&P Global (42 percent) measured abandonment at the company level, and its trend line is the alarming part: 42 percent of companies abandoned most AI initiatives in 2025, up from 17 percent the year before, with the average organization scrapping 46 percent of its proofs of concept. The failures are accelerating as the spending does.
Gartner (30 and 40 percent) are projections rather than measurements, but disciplined ones with causes attached: GenAI abandonment on data quality, risk controls, costs, and unclear value, and agentic cancellations on the same trio, alongside a memorable census: of the thousands of vendors claiming agentic products, Gartner counted roughly 130 genuine ones.
Why six different numbers agree
Stack them and the funnel appears. Companies start many proofs of concept; most never deploy (IDC's 88). Of what deploys or gets bought, much never enters the workflow, so returns never register (MIT's 95). Companies respond by abandoning initiatives in bulk (S&P's 42, doubling year over year), and the analysts project the same attrition forward (Gartner's 30 and 40). Different instruments, one funnel, and every study names versions of the same causes: problems misunderstood at the start, production never scoped, value never defined. The anatomy of those causes is its own article.
Different instruments, one funnel: designed far from the work, never scoped for production, abandoned at scale.
How to use these numbers
Not as doom, and not as decoration. The failure statistics are a checklist in disguise: every cause the studies name is a decision you control. Pick one workflow with an owner instead of a portfolio of demos. Define production and its evals before building. Put the system inside the workflow, with the people who run it. Own what gets built. Organizations that do those four things are, definitionally, not the population these studies describe. That is the entire argument for the forward-deployed way of buying AI, and it fits in one sentence: the failure rates measure distance from the work.