A manufacturing company brought me in a couple of weeks ago to scope out some training. Before I got to my first question, one of their leaders told me exactly where they stood. Microsoft Copilot was on every laptop in the building and had been for a while, the licenses were paid, and his people were using it like a search box. He had a name for it. Google mode. Type the question, skim the first answer, move on. Months into a real investment, that was the honest state of things, and he volunteered it before anyone asked him to.

I have sat in a lot of rooms like that one. What stood out was not the Google mode, which is ordinary. What stood out was that he knew.

The gap between the two floors

WalkMe put a number on that distance this spring. The company sells software for driving exactly this kind of adoption, which is worth saying out loud before leaning on its research, and the survey itself is substantial: 3,750 people across fourteen countries, seventeen hundred senior leaders and just over two thousand office and hybrid workers at companies of a thousand employees or more.1 Sixty-one percent of the executives said they trust AI to handle complex, business-critical decisions. Among the workers, nine percent said the same.1

Fifty-two points is not a difference of opinion about software. It is two groups appraising the same tool from positions with almost nothing in common.

Here is what I think sits underneath it. Confidence in a tool rises with distance from the consequence of being wrong about it. The executive who would hand an agent a business-critical call is rarely the person keying the inputs, and almost never the person standing in front of a customer when the schedule that agent produced turns out to be impossible. The nine percent belongs to people who would be doing both.

The same survey found eighty-eight percent of executives confident their people had the tools they needed to work well. Twenty-one percent of the workers agreed.1 Fifty-four percent of them said they had gone around an AI tool and finished the job by hand at least once in the previous thirty days.1

The people who would have to live with the decision are the least willing to hand it over, and they are also the only ones who have watched it fail.

What the workaround is telling you

There is a reading of that behavior available from any conference room, which is that people are dodging the tool and need to be coaxed onto it. I would offer a different one. Somebody who performs a task forty times a week and chooses to do the forty-first without the tool has run more trials on it than the person who approved the purchase. That choice is a finding. It is very likely the most current information the company owns about where the thing holds and where it comes apart, and it is being filed as an attitude to be corrected.

A much-quoted number points the same way, and it deserves more care than it usually gets. MIT's Project NANDA reported in 2025 that ninety-five percent of generative AI pilots produced no measurable impact on profit and loss.2 That figure has travelled a long way, often cited as though it closed the matter. It rests on fifty-two organizational interviews, a survey of a hundred and fifty-three executives, and a review of three hundred public deployments, which is a modest base for a claim carried that far.2 I would treat it as directional rather than settled. Pointed in that direction, it happens to agree with the nine percent.

None of this is an argument against the technology. I use these tools every working day and they have changed how I do the job. The question is narrower and considerably harder. When the people closest to the work and the people furthest from it disagree about whether a tool is ready, whose reading governs? Most organizations answer that question by seniority and never notice they answered it.

The company that already knew

Which brings me back to that room. The leader who described Google mode before I asked had no better instrumentation than anybody else. He had asked his people, and then he had believed what they told him, and that put him further along than most of the companies I walk into. His adoption problem is a solvable one now, for the plain reason that it is described correctly. Nobody is going to spend the next two quarters fixing an imaginary version of it.

I have written before about rollouts that skipped the people expected to be changed by them, and I am not going to relitigate that here. This is the step after. Suppose the rollout did go out as an email, as most of them do. The information you need is still available, sitting in the behavior of everyone who has been working around the tool since April, and it is free.

So the question I would put to anyone sponsoring an agent this quarter. Do you know, specifically and recently, what your people actually do with the tool when nobody is watching them do it? And if the answer turns out to be that they go around it, are you willing to read that as information rather than as something of theirs to fix?