The AI Disappointment Is Real. It's Also a Head Fake.

There's a mood shift happening around AI, and if you run a business, you've felt it. Two years ago the promise was total. Now the room is quieter, more skeptical, a little tired. The pilots didn't pay off. The magic looked like a demo trick. And a growing number of leaders have quietly concluded the whole thing was overhyped.

The disappointment is real, and the data backs it up. But the conclusion people are drawing from it is exactly wrong, and I think it's about to become one of the more expensive misreadings in recent business history.

Let me show you the number everyone is reacting to.

That number went off like a bomb, and it wasn't alone. Roughly 42% of companies abandoned most of their AI initiatives in 2025, more than double the year before.² Among the businesses that stuck with it, the reported gains turned out to be modest: about 74% of leaders say AI improved productivity, but for most of them the improvement sits under 25%, and Goldman Sachs found no meaningful link yet between AI adoption and productivity across the economy as a whole.³ Real gains, but local ones. Nothing like the revolution on the slides.

So the skeptics have their evidence, and I'm not going to argue with it. I'm going to argue with what they think it means.

What the failures actually measured

Read the MIT study past the headline and it says something very different from “AI doesn't work.” It says the failure was almost never the model. The tools performed. What broke was everything around them: bad data, no workflow integration, and pilots aimed at problems that were never defined clearly enough to have a measurable answer.¹

In other words, the 95% didn't fail a test of the technology. They failed a test of implementation, and they failed it in the most predictable way possible, by treating a demo as a deployment and expecting a general-purpose tool to reshape a business it was never connected to. The disappointment isn't evidence that AI is overhyped. It's evidence that using it well is much harder than buying it, which is a completely different problem with a completely different fix.

Confuse those two, and you draw the losing conclusion. You look at a wave of botched implementations, decide the technology is the issue, and step back to wait it out. Which brings us to the timing.

The worst possible moment to quit

Here is what makes the retreat so costly. The companies pulling back are doing it at the exact moment the technology crosses from novelty into infrastructure.

Until recently, most business AI was a chat window: you asked, it answered, and nothing happened unless a human carried the output somewhere. That's the version that disappointed everyone, because a chat window doesn't run anything. What's arriving now is different in kind. Gartner projects that 40% of enterprise software applications will include task-specific AI agents by the end of 2026, up from less than 5%.⁴ Agents don't just answer. They do the work: they reconcile, they route, they follow up, they act inside your systems without a person in the middle.

So the leaders concluding “AI didn't deliver” are judging the whole technology by its weakest, earliest form, and shelving it the season before the capable version shows up. It's the equivalent of test-driving a car in the driveway, deciding cars are overhyped because it didn't take you anywhere, and selling it the week before you learned to leave the garage.

What the 5% did differently

The gap isn't luck, and it isn't budget. MIT's own data shows the split. The initiatives that worked were the ones built for a specific workflow, integrated deeply into real systems and data, and owned by someone accountable for the result. Externally built, workflow-integrated tools succeeded roughly twice as often as generic or internal ones.⁵ The winners didn't have better AI. Everyone has the same AI. They did the unglamorous work underneath it that the 95% skipped.

Put those three numbers next to each other and the picture is almost absurd. A large share of companies are quitting a technology whose useful form is arriving in months, on the basis of experiments that failed for reasons that had nothing to do with the technology. The retreat and the breakthrough are happening in the same quarter.

What to do with this

Don't let a bad first attempt talk you out of the whole thing, and don't let the backlash convince you the window closed. Both reactions are the head fake.

The right move is narrower and duller than the hype and more useful than the retreat: pick one workflow that actually matters, connect the tool to your real data, put someone accountable on the outcome, and measure it. That is the entire difference between the 5% and the 95%, and it is available to a business of any size. The technology was never the hard part. Doing this part was, and still is.

The companies that win the next few years won't be the ones that felt the most optimism in 2024 or the most cynicism in 2026. They'll be the ones who quietly kept building through both.