ISSUE 008
"95% of AI pilots fail" is not what the study found
If you have heard one statistic about AI in business, it is this one: ninety-five percent of pilots fail. It was everywhere last year and it is still being quoted. It comes from a real report, the report says something narrower than the headline, and the narrower version is considerably more useful to anyone running a small business.
An idea worth thinking about
The AI that works is the AI nobody approved
Buried under the famous number in the same report is a finding I think is the more important one. The research describes a shadow AI economy: employees at more than 90 percent of the firms studied use personal AI tools, including when the company's own official pilot is going nowhere.
So the picture is not an organisation failing to adopt AI. It is an organisation whose formal programme stalls while its people quietly route around it with whatever they can open in a browser. On paper the pilot fails. In practice work is getting done with tools nobody procured.
That inverts the usual advice. The question for most businesses is not how to launch an initiative, it's how to notice what people have already started doing and make it legitimate, supported and reviewable. The useful thing to go find out this week is which of your staff are already using something, for what, and whether it's any good — which costs nothing and is the one step the failing enterprises in this report skipped.
A tool I disappeared into
The enterprise AI failure studies, lined up with their caveats
Somebody assembled the major enterprise AI return studies into one comparison table with each one's limitation stated next to its headline figure, which is exactly the format this topic needs. MIT NANDA at roughly 95 percent showing no measurable profit-and-loss impact, flagged as preliminary and not peer-reviewed. S&P Global finding 42 percent of companies abandoned most AI initiatives and scrapped 46 percent of proofs of concept. Gartner forecasting over 40 percent of agentic AI projects cancelled by the end of 2027. RAND putting AI project failure above 80 percent.
Reading them together does something no single figure does. These are four different measurements of four different things — profit impact, abandonment, forecast cancellation, project failure — and they get quoted interchangeably as though they were corroborating each other. They aren't corroboration. They're a chorus.
The honest caveat on the source: it's published by a cloud cost-management company, so it has an interest in enterprises thinking carefully about AI spend. The table is still the clearest side-by-side I found, and every row links out.
Learn something
How a subsample became a headline
A technology analyst walked through the NANDA report looking for where the 95 percent actually comes from, and the answer is less solid than the coverage implies. The figure appears to be derived from a narrower group — organisations that piloted task-specific generative AI tools — rather than from all enterprises in the study. His objection is that treating the remainder as failures, including organisations that never ran a task-specific pilot at all, is an unsupported leap.
Then watch the compression happen in public. Within days the report was being reported as 95 percent of enterprise AI failing, 95 percent of AI pilots failing, and simply 95 percent of AI failing. Each restatement dropped a qualifier. The report's own language is about no measurable profit-and-loss impact inside its measurement window, which is not the same claim and is considerably easier to produce — partly, as critics note, because many of these companies never established a baseline to measure against.
This is a free lesson in how a statistic you will be quoted at gets built. Somebody measures something specific. A summary drops the specificity. A headline drops the hedging. By the third retelling it is a fact about the world. The defence is always the same: find the original, read the method, and notice what population the number actually describes.
What we’re seeing
Ninety-five percent showed no measurable profit impact. That is a different sentence from ninety-five percent failed.
MIT's Project NANDA published The GenAI Divide: State of AI in Business 2025, built on 150 interviews with leaders, a survey of 350 employees, and analysis of 300 public AI deployments. Its headline finding: roughly 95 percent of enterprise generative AI pilots produced no measurable profit-and-loss impact, against an estimated $30 to $40 billion in enterprise spending. About 5 percent achieved rapid value.
Read the sentence precisely, because the difference is the whole lesson. No measurable impact within a measurement window is not the same as failure. It can mean the pilot didn't work. It can equally mean nobody set a baseline, the window was short, or the benefit landed somewhere the finance team wasn't looking. The report is preliminary, has not been peer reviewed, and has been publicly criticised on exactly these grounds.
The two findings inside it that I'd actually act on have nothing to do with the 95. First, the money is in the wrong place: more than half of generative AI budgets went to sales and marketing tools, while the largest returns showed up in back-office automation — cutting outsourced processes, reducing agency costs, streamlining operations. Second, how you build matters more than what you buy: pilots combining internal people with outside expertise are reported at a 67 percent success rate against 22 percent for builds done entirely in-house by IT.
Now the part that applies to a business of six people, which is most of the people reading this. You do not have a pilot. You do not have a steering committee, a procurement cycle, or an AI budget that can be misallocated to a sales tool. The failure mode this report describes — an initiative that impresses in a demo and dies on contact with a legacy process — requires an organisation large enough to have both. Almost nothing in it describes you.
What does transfer is the back-office finding and the shadow AI one. The return is in the unglamorous recurring work, not the customer-facing showpiece. And whatever is already working is probably something somebody started doing without asking. Neither of those needs a 95 percent to justify it.
One question
Is anyone in your business already using AI without being asked to?
Reply with what they're using it for. The research says this is where the value actually is, and that most owners don't know it's happening.
More in this pillar as it's written.
— Shana
Sources link to the original publisher, never to a summary.