Why 95% of AI projects fail, what the MIT report behind that number measured, and why a small business should be watching capacity rather than cost savings.
Ninety-five percent of AI didn't fail. That number comes from one 2025 MIT report on enterprise AI pilots, which found 95% of organizations getting zero return. Those pilots only counted if AI replaced a cost. Most of what AI does in a small business never had a cost to replace.
The report is The GenAI Divide: State of AI in Business 2025, out of MIT's Project NANDA in July 2025. Three sources of data: more than 300 publicly disclosed AI initiatives, 52 structured interviews, and 153 senior leaders surveyed at four conferences. The authors call their own findings preliminary. Nobody audited anybody's books.
That's not a failure rate for AI. It's a failure rate for one kind of project, measured one way.
Where the money went gives the measuring stick away. Asked to split a hypothetical hundred dollars, executives put most of it into sales and marketing, around seventy in one section of the report and closer to fifty in others, so treat the exact split as soft and the direction as clear. The returns the report could actually find were somewhere else. Two to ten million a year from replacing outsourced customer service and document processing. Thirty percent off agency spend. The report's read is that the front office gets the money becuase the front office is easier to measure.
Every return in that report is something that stopped being bought.
I asked ChatGPT the same question this week. It cited two company blogs, a university page and a news story. The report itself was not in the list, so the most repeated AI statistic of the year is circulating through people summarizing each other, and all of them inherited the cost-cutting frame without noticing it was there.
Worth saying out loud: a statistic claiming almost everybody is doing this wrong is good for my business. I sell AI adoption work. That is exactly why I went and read the thing.
Replacing a cost is one of the things AI does. It's not the interesting one.
I have a should list. Things have sat on it for years, not because I didn't know they mattered, but because doing them meant hiring somebody and they were never worth a hire. Posting on LinkedIn consistently sat there a long time. No budget line to cut, no invoice to reduce, no hours to save. It just didn't happen. Now it happens, through digital employees instead of a person on payroll.
That's not money saved. It's money unspent.
The MIT measurement can't see that. Nothing was being spent, so there was never a number to compare it to. On a profit and loss statement, zero return and a thing that finally got done look the same.
What did show up for me is speed. I post weekly now, LinkedIn and my own blog, and the whole thing takes ten to twenty minutes.
Ten to twenty minutes is why it happens at all.
So the question isn't what you can cut. Write down your should list, not your to-do list. The things you've wanted done for two years and never assigned, because assigning them meant hiring. Then go through them and ask which ones a machine could carry now, set up properly, with you checking the work.
More of them than you'd guess.
If you want to go through that list with someone, I'm happy to have that conversation. A fair bit of what looks out of reach isn't, with the right approach.
Ninety-five percent of those pilots were asked to save money. That was never the best thing to ask for.