AI pilots do not fail because the models are weak.
In McKinsey’s 2025 survey of nearly 2,000 respondents, 88 percent reported using AI regularly in at least one function. Only 39 percent reported any impact on enterprise-level profit. One 2025 MIT study of enterprise deployments reported that 95 percent of generative AI pilots produced no measurable return. That figure is contested, and it is worth reading alongside McKinsey’s rather than in place of it. Both point the same direction.
The model is not the problem. Adoption is nearly universal. Results are not.
You know the shape of it from your own rollout. You ran the pilot. Adoption climbed. The dashboards turned green. Your people move faster, and they will tell you so. And the thing you were trying to fix, the new initiative that takes months, the approval that takes weeks, takes as long as it always has.
That gap is not new. It arrives with every general-purpose technology, and the clearest case is the one most people have forgotten. Let’s go back in time.
When American factories began swapping steam engines for electric motors in the 1880s, the productivity gains took roughly forty years to arrive. Not because the technology was not disruptive. For decades, managers put the new motor exactly where the old steam engine had stood, spinning the same long drive shaft, with machines still crowded along it the way steam had demanded.
The motor was better. The factory was the same. The motor changed almost nothing.
The gains came only when a later generation redesigned the floor around what electricity made possible: a small motor at every station, machines arranged by the logic of the work instead of the logic of the drive shaft.
Now read that next to your copilot rollout. A model bolted onto an unchanged organization is the electric motor in the steam engine’s spot. Locally faster, sure, but structurally pointless, because the delay you were trying to remove never lived inside any one person’s work. It lived in the seams between them.
So here is the test, and you can run it this week on a single process. Take your slowest process that a customer actually feels. Ask one question of it: does the time go on what any one person is doing, or on the waiting between them? The drafting, the calculating, the writing, a copilot can reach. The queue, the “let me check with legal,” the handoff that sits in someone’s inbox overnight, it cannot. If most of your delay is in the waiting, you do not have a tooling problem. You have a wiring problem, and no copilot reaches the seams.
A copilot makes a person faster. It cannot make a handoff disappear, because it was never standing in the handoff.
You replaced the engine. You never rebuilt the factory.
There is a second trap waiting for the companies that do manage to speed up.
Next in the series: why running faster, when everyone can run faster, changes nothing at all.
Leave a Reply