The slide said ninety-two percent adoption. Someone had put it in green. A room full of senior people nodded at it the way you nod at good weather, and the chief operating officer asked the only question that mattered, which was what all of it had actually produced. The green number stayed on the screen. Nobody had an answer that survived a follow-up.

I have sat in a version of that meeting more than once over the past year, and the shape of it barely changes. A company buys the tools, the licenses light up, usage climbs, and then the return everyone expected fails to show up anywhere a finance team can find it. When reviewers at MIT went looking for that return across more than three hundred corporate initiatives, the number that came back was brutal in its plainness. Ninety-five percent of organizations were getting no measurable return on their generative AI spending, with only one in twenty turning a pilot into real profit-and-loss impact.1

The reflex is to blame the technology

In most executive rooms the low number reads as a technology problem. The model is not good enough yet, or the data is too messy, and another quarter with another integration will surely close the gap. What the MIT reviewers found underneath the failures was not a weakness in the models. The tools were being dropped into workflows nobody had redesigned, with no mechanism to learn from real use.1 The outcome each tool was bought to deliver had no owner.

Adoption is not the same as integration

There is a gap between a tool being used and a tool changing how the work gets done, and most dashboards cannot see it. Harvard Business Review looked at this early in 2026 and found that while eighty-eight percent of companies reported regular AI use, employees were mostly experimenting at the edges rather than folding the technology into the actual mechanics of their jobs.2 People opened the tool, ran a few queries, and closed it again. The usage counter climbed. The way decisions got made stayed exactly where it was.

That distinction is where the money leaks out. A leader watching the adoption curve rise feels like progress is happening, because something is measurably increasing. The thing being measured is activity. The thing that was promised was outcome, and the two have stopped tracking each other.

Usage tells you a tool was opened. It tells you nothing about whether anyone is answerable for what the tool was supposed to accomplish.

The accountability that never got assigned

Ask who owns an AI result and watch the room. Not who approved the budget, not who ran the rollout, but who is answerable if the thing produces nothing, or produces something wrong. A survey by the compensation firm Pearl Meyer caught the seam cleanly, with boards largely assuming the C-suite owned AI strategy while the executives themselves did not agree that they did.3 Ownership was presumed at one altitude and dodged at another, and in the space between them the accountability simply evaporated.

This gets more serious as AI moves from drafting emails to touching real decisions. When a model shapes who gets screened for a role, or how a case gets prioritized, or what a customer is quoted, the question of who answers for the result stops being a governance nicety. If the answer is a vendor, or a committee, or a transformation office that reports to everyone and therefore to no one, then a technology now sitting inside consequential calls has no human clearly standing behind it.

What the five percent do first

The leaders who get something back do a boring thing before they touch a license. They name the outcome the tool is supposed to move and the person answerable for it. Then they fix the process the tool should improve, rather than laying it on top of one that was already broken. The technology decision comes last, because the technology was never the hard part.

None of that shows up on an adoption slide. It is slower, it is less impressive in a board update, and it is the only version of this that produces a number a finance team can actually find. The uncomfortable truth for a lot of senior teams is that the ninety-five percent are not being let down by their models. They are being let down by a decision they never quite made about who owns the result.

So the question worth carrying into your next AI review is not how many people are using the thing. If this tool delivers nothing next quarter, whose name is on that? If you cannot say it out loud, you have a pilot, and you do not yet have a change.