the 95 percent myth

The most quoted AI statistic of the moment, that 95 percent of AI pilots deliver no measurable result, comes from an MIT report whose methodology is described inconsistently even in its own coverage. Nearly every other failure figure making the rounds has similar flaws: a small sample, data kept private, or a vendor footing the bill. Yet the solid sources all point the same way: a small minority of companies demonstrably makes money with AI. The pattern is real, the precise percentages are shaky.

In the production gap I work out the thesis: the demo is 20 percent of the work, and most projects die between pilot and production. This essay is the fact-check that belongs with it. Where do the failure numbers come from, how sturdy are they, and what is left standing when you get strict?

is it true that 95 percent of ai projects fail?

Nobody knows, and the MIT report the number comes from certainly does not prove it. In August 2025 one headline went around the world: 95 percent of generative AI pilots deliver no measurable effect on the profit and loss statement, despite 30 to 40 billion dollars of investment (Fortune, 2025). The source was “The GenAI Divide: State of AI in Business 2025” by MIT Project NANDA. The number reached boardrooms, market commentary and countless LinkedIn posts, and nearly a year later it still turns up in almost every presentation on AI strategy.

Anyone who wants to read the report first has to fill in a request form. The raw data was never released. And the methodology is described inconsistently: the report itself speaks of 52 structured interviews, 153 surveys collected at four conferences and a review of more than 300 public AI initiatives, while Fortune’s own coverage mentions 150 interviews and 350 surveys (Futuriom, 2025). Those numbers contradict each other within the coverage of one and the same report. Analyst site Futuriom also laid the chart the figure leans on next to the conclusion and disputes whether the 95 percent can be derived from it at all. Wharton professor Kevin Werbach publicly voiced doubt about the report, in careful terms. The harder demand, that MIT release the full data or withdraw the report, comes from Futuriom itself. Those two must be kept apart, and most retellings do not.

Maybe the real percentage is lower, maybe higher. The honest answer is that the most quoted AI figure of the moment is also the least verifiable one. A report about sloppy AI implementation that itself looks methodologically sloppy: that irony deserves to be named out loud.

where do the other failure numbers come from?

From research that is almost always smaller, softer or more interested than the headlines suggest. Walk down the best-known ones and you keep seeing the same thing: the percentage in the headline and the research underneath rarely fit together.

The second most popular number comes from RAND Corporation: more than 80 percent of AI projects supposedly fail, twice as often as IT projects without AI (RAND, 2025). Read the publication itself and it literally says “by some estimates”. RAND is citing an existing estimate by others. Its own research consisted of interviews with 65 experienced data scientists and ML engineers and produced a qualitative analysis of five root causes. That is valuable work, and it does not substantiate the percentage itself. Nearly every retelling glues those two things together as if they were one measurement.

Then the doubling that made the rounds in 2025: 42 percent of companies abandoned the majority of their AI initiatives that year, against 17 percent in 2024, and on average 46 percent of proofs of concept died before production. Those figures come from the “Voice of the Enterprise” series by S&P Global Market Intelligence, a survey of more than a thousand IT and business decision makers in North America and Europe. The series is real and multiple independent sources cite exactly the same numbers, but I have not managed to access the primary publication either. So I pass them on as I found them: confirmed secondhand (WorkOS, 2025).

In June 2025, Gartner predicted that more than 40 percent of agentic AI projects will be canceled before the end of 2027, driven by rising costs, unclear business value and inadequate risk controls (Gartner, 2025). That is a prediction, so by definition it cannot be checked against outcomes yet. I find a side catch from the same press release more interesting: of the thousands of vendors calling themselves agentic AI, only about 130 actually deliver agentic capabilities, according to Gartner. The rest sell polished chatbots and RPA. Gartner calls that “agent washing”, and that word alone is worth remembering at every vendor demo.

And then the Dutch number: 21 percent AI project success, the lowest of six European countries studied, combined with the highest internal resistance (38 percent). It comes from “State of Integration & AI 2026”, research commissioned by integration platform Frends and carried out by Sapio Research (Frends, 2026). Two caveats belong with it. An integration platform benefits from the conclusion that an integration-first approach works. And the sample covers organizations from 201 employees up, so large companies and the upper end of the midmarket. For a company of fifteen people this number says little.

what survives once you strike the shaky numbers?

A pattern confirmed again and again by methodologically independent sources: nearly everyone has AI in use, and the measurable gains stay with a minority. Fortunately that pattern rests on sturdier material than the headline percentages above.

Morgan Stanley tracks each quarter how many S&P 500 companies can name a measurable AI benefit: 10 percent at the end of 2024, 15 percent in the third quarter of 2025, 21 percent at the end of 2025 (Morgan Stanley, 2026). The trend is rising, and it remains a minority, at the largest and best-funded companies in the world no less. IBM asked two thousand CEOs about their AI investments in early 2025: a quarter of the initiatives delivered the expected ROI, and 16 percent have been scaled company-wide (IBM via Fortune, 2025). McKinsey’s own State of AI survey, consultancy research that is, shows the same picture: 88 percent of organizations use AI in at least one business function, 39 percent see measurable impact on the bottom line and roughly 38 percent are past the pilot phase (McKinsey via CX Today, 2026). BCG, also a consultancy, finds among 1,800 executives a quarter reporting significant value (BCG, 2025).

For small business, the OECD survey of more than two thousand SMBs in twelve countries is the most relevant: AI adoption among SMBs grew from 7.1 percent in 2023 to 17.4 percent in 2025, while the gap with large companies actually widened, from 23.4 to 34.6 percentage points (OECD, 2026). Eurostat measures 55 percent AI use at large companies in the EU against 17 percent at small ones (Eurostat via Ipsos, 2026). I quote those numbers EU-wide on purpose: I have not found a reliable figure for Dutch SMBs, and the numbers circulating on Dutch marketing blogs contradict each other so hard that I use none of them.

Add it up. Equity analysts, a CEO survey, two consultancies and official statistics, all with different methods, arrive at the same picture. You do not need to believe MIT’s 95 percent at all to take the production gap seriously. In fact, whoever builds their story on that one contested percentage is doing exactly what the report accuses the companies of. Impressive demo, thin substantiation.

what do the companies that do capture value do differently?

Three patterns keep returning in the research, each with its own source and its own caveat.

Buying over building. The contested MIT report contains one finding that received far less attention than the headline: purchased AI solutions and partnerships succeeded in roughly 67 percent of cases, internal builds in roughly 33 percent (Fortune, 2025). Same report, so same caveat. It does match what I see in practice: those who build in-house systematically underestimate management, maintenance and the road to production.

Starting small and focusing. The quarter of companies that captures significant value according to BCG concentrates on a small number of initiatives, scales them fast and adapts the underlying core processes (BCG, 2025). Consultancy research, with an advisory practice as its interest, and it lines up with the IBM picture that scaling is the bottleneck.

Data and integration in order before the pilot. Network vendor Cisco calls 13 percent of companies “Pacesetters” in its AI Readiness Index; that group converted four times as many pilots into production (Cisco, 2025). Vendor research, just like the Frends report that finds the same pattern in Europe. More independent is the study of agentic AI at industrial companies that Forbes cited this month, with a term that describes exactly what I see at clients: the “capability-deployment verification gap”. A pilot runs neatly in the test environment and loses all confidence the moment it runs against live business data (Forbes, 2026).

how to handle failure numbers and pilots yourself

For the small-business reader who finds an AI proposal on their desk tomorrow, this is the order I follow myself:

1. Distrust every round percentage. Ask about the sample, who commissioned it and whether the raw data is public. A number from vendor research is marketing until proven otherwise.
2. Pick one process where the euros or hours are already measurable: quotes, customer questions, planning, invoicing.
3. Agree before the pilot what success means, in money or hours per week, and who measures it.
4. Buy or rent a proven solution before you have anything built. The best-substantiated success patterns point that way.
5. First check whether the data the system needs exists, is correct and is accessible. As small as a tidy FAQ, as large as your ERP integration.
6. Plan the road to production before the pilot starts: who will manage it later, what it costs per month, what happens when it fails.
7. Stop in time. A three-week pilot you call off costs a fraction of a year-long zombie project.

What you can safely skip: training your own model, setting up a company-wide AI transformation program, and signing with every party that has put the word agent on its homepage. Not a single number in this essay gives any reason for that, and Gartner’s agent-washing estimate gives every reason for the opposite.

where I stand

I do not believe the 95 percent, and I do not believe the reverse hype either. My position: pumping failure numbers around is itself demo culture, an impressive headline on a thin measurement, while the underlying pattern needs no headline. Independent sources keep showing the same thing: measurable AI value stays with a minority, and that minority buys in, starts small and has its data in order before the pilot starts. That is duller than a percentage of 95, and it is something you can build on.