If you've spent any time where business owners talk about AI, you've seen the number: 95% of AI projects fail. It shows up in sales decks, LinkedIn hot takes, and — with a straight face — in pitches from the same consultants whose services it supposedly indicts. It's become the industry's favorite scare statistic.
Almost nobody quoting it has read the source. Let's fix that, because what the study actually found is more interesting than the headline — and more useful to you.
The number comes from The GenAI Divide: State of AI in Business, a 2025 report from MIT's Project NANDA, built on 52 executive interviews, surveys of 153 leaders, and analysis of 300 public AI deployments. Three things about it get dropped every time it's quoted:
First — it's enterprise data.The population is large organizations running formal generative-AI pilots against billions in collective spend. It says nothing, statistically, about a twelve-person plumbing company or a forty-person logistics firm. Small businesses fail at AI too, but this study didn't measure them, and borrowing its number is somewhere between sloppy and dishonest.
Second — "fail" isn't what it says. The finding is that ~95% of pilots produced no measurable P&L impact. That phrase is doing enormous work. A pilot that quietly saves four hours a week but was never instrumented lands in the 95% — not because it produced nothing, but because nobody could prove it produced anything. The study's authors are explicit that the divide isn't model quality; it's the learning gap between tools and the organizations deploying them.
Third — the 5% is the actual story.A sliver of integrated deployments extracted millions in value. The difference wasn't better AI. It was that those teams picked one workflow, wired the tool into it deeply, and measured the result.
The 95% stat doesn't say AI doesn't work. It says almost nobody set up the measurement that would let them know either way.
Which reframes the whole thing. The scary version of the stat sells fear: AI is a casino, most people lose. The accurate version is closer to an indictment of process: companies bought technology before they built the yardstick. They ran pilots with no baseline metric, no owner, and no definition of "working" — then couldn't say whether the pilot worked. Of course they couldn't.
There's a supporting detail in the same report that deserves more attention than the headline: while only about 40% of the companies studied had bought an official LLM subscription, employees at roughly 90% of them were already using AI tools personally for work. The organizations were failing at AI while their own people were quietly succeeding at it, one unmeasured workflow at a time.
If you run a small business, the takeaway isn't "avoid AI" and it isn't "buy AI." It's: decide the number first. Pick the one figure that should move — calls answered, median time to first reply, hours of retyping — instrument it, and only then put software against it. Do that and you cannot end up in anyone's 95%, because whatever happens, you'll know.
That's the entire method we run deployments on, and it's also the free version: send us your website and we'll write up where your business is most likely leaking — no charge, no pitch.