Why AI Projects Fail. 88% Use AI. Only 6% Profit.

Ask ten owner-led companies whether they use AI and nine will tell you yes. Someone in ops is running reports through it. Someone in sales is rewriting a proposal with it at 9 PM. Someone is summarizing the meeting nobody wanted to take notes in. Adoption stopped being an interesting question: 88% of organizations report regular AI use in at least one business function.

Almost none of them are getting paid for it. Around 6% pull 5% or more of EBIT out of what they’re doing with AI. Any measurable bottom-line impact at all, of any size, shows up in 39%, and for most of that group it’s scraps. Two-thirds or so never leave the pilot stage. Those figures are McKinsey’s, from its State of AI survey, and they match what we find walking into a company for the first time. Everybody’s using it. Almost nobody’s making money on it.

Why AI projects fail: the tool was never the problem

The failure looks the same nearly every time. A company buys licenses, hands out logins, books a webinar, and then asks people to use AI on work that is still shaped exactly the way it was before. Approvals still take three days. The handoff between sales and ops still happens in a spreadsheet somebody emails around. Nothing about how the work moves changed, so nothing about the result changes either, and eight months later someone asks what the subscriptions are actually for.

The companies that get value did the harder thing. They took one workflow that mattered, pulled it apart, and rebuilt it around what the tool is good at. That’s the whole difference. Not a better model, not a bigger budget, not an earlier start. McKinsey’s read on its high performers lands in the same place: what separates them is workflow redesign, not tooling. MIT’s State of AI in Business arrives from the other direction, putting 95% of generative AI pilots in the no-measurable-return column, and the ones that die are the ones where a generic tool got dropped into unchanged work and stalled the moment the job needed context or a second attempt.

Adoption everywhere, transformation nowhere. That’s the pattern, and it’s an organizational failure, not a technical one.

The proof is already inside your building

Here’s the part that makes the gap strange rather than just disappointing. Walk into most of these companies and the evidence that AI works is already there. It’s just not on anyone’s dashboard. Somebody in accounts payable worked out how to reconcile a report in eight minutes instead of forty, on a personal account, and never mentioned it because they weren’t sure they were allowed to.

That isn’t an anecdote, it’s the normal case. Workers at more than 90% of the companies MIT surveyed were already using personal AI tools for real work, while only around 40% of companies had bought an official subscription to anything.

Read those two facts next to each other and the 88/6 gap changes shape. It isn’t that 82% of companies bought something that doesn’t work. It’s that 82% of companies have people quietly proving it works, in disconnected pockets nobody can see, govern, or repeat, while the organization has no structure to notice what its own staff already figured out.

Which makes it a governance problem before it’s a technology problem. Every one of those private workarounds is simultaneously the best evidence you have and a live exposure: company information moving through accounts nobody approved, with no record of what went where.

What closing that gap looks like up close

When we run a training cohort, the work isn’t teaching ten people to use AI. Most of them already have, on their own laptops. The work is taking what someone already worked out privately and rebuilding the actual workflow around it, with governance in place so IT signs off instead of looking away.

Across the cohorts we’ve run, that shows up as 15 to 20% efficiency gains by the third week, and each person walks away having automated two or three specific tasks. Not hypothetical ones. The ones they were already half-solving alone before anyone gave the effort a structure. Multiply that across a ten-person cohort and you’re recovering somewhere between one and two full-time employees’ worth of capacity, out of work that already exists rather than headcount you have to add.

None of it required a better model than the one already sitting on those laptops. It required treating the individual workaround as a signal instead of a compliance problem, and then building the rest of the company around it.

The 6% didn’t win by adopting AI first

Nearly everyone adopted AI first. The companies getting value won by noticing what their own people had already proven, and by having the discipline to rebuild a real workflow around it before the mess talked them out of it. That’s harder to buy than a subscription, and it’s why the distance between 88% and 6% is still this wide with everyone holding the same tools.