The email arrives with a subject line that sounds like progress. "We've deployed AI." There is a slide deck somewhere behind it, a short FAQ, a working group with a name, a link you are encouraged to explore. You click the link. The tool opens, clean and confident, and you ask it to do the one thing that would actually save you an afternoon — open the real file, the messy quarterly workbook with the tabs nobody documents, and pull the number you keep re-deriving by hand. It cannot. It can chat. It can summarize the paragraph you paste into it. It will, if you are feeling bleak, write you a short poem about reconciliation. It cannot reach the file.
So you close the tab. And in the closing of that tab there is a small, private verdict, and the verdict is about you. *Maybe I'm not using it right. Maybe everyone else figured out the trick and I'm the one who didn't.*
I want to be the person who tells you, plainly, before anything else: you are not using it wrong. You got the default.
For a long time that was the most I could offer — recognition, the warm hand on the shoulder, "me too." What's changed is that the experience now has numbers attached to it, and the numbers are not small or hedged or buried in a footnote. They are the headline finding of the most-cited reports of the year. The thing you felt at your desk turned out to be the thing the whole market was quietly measuring.
The numbers are in
Start with the one that made the rounds. A review out of MIT's NANDA initiative last year looked across the enterprise AI pilots — the deployments, the platforms, the transformations — and found that roughly ninety-five percent of them returned nothing. Not "less than hoped." Nothing measurable on the books. Tens of billions of dollars of spend, and the overwhelming majority of it produced the glossy intranet page and not much else. (The five percent that worked did something specific, which I'll come back to in a moment, because it matters more than the ninety-five.)
Then the one that names your exact tool. Forrester went looking at the Copilot rollout — the one most likely to be the AI your company "deployed" — and found that something like seventy percent of the Fortune 500 are licensed for it, and only twenty to thirty percent of those paid seats get used in a given week. Read that again slowly. The company bought the seats. The seats are sitting there, lit up, unused, because the thing behind them cannot do the job people have. That is not a seat-count problem. That is the gap between "we paid for it" and "you can use it."
The tool isn't weak. It's fenced.
This is the part I need you to hold onto, because it is the difference between a story about your competence and a story about a configuration screen. The tool in front of you is not weak. It is "fenced." A locked-down corporate tenant is, by design, a tool with the useful settings turned off — the connectors to your files disabled, the access to the database withheld, the permissions that would let it touch anything real left in the box. The analysts have a tidy way of putting it: Copilot is a governance project wearing the costume of an AI project. Which is a polite way of saying the thing you were handed was decided by people thinking about risk, not about your Tuesday.
There's a line I keep coming back to, because it's the cleanest description of the feeling I've ever found. They bought us a car, then removed the engine and the steering wheel. The radio still works, though. That's it. That's the rollout. You got the radio. You were told it was the car, and when it wouldn't drive you assumed you'd forgotten how to operate a car.
And here is the quiet cruelty of it, the part that keeps it personal long after it stopped being personal: the people who decided this mostly think it's going great. There's a survey from BCG and Columbia — fourteen hundred people — where seventy-six percent of executives believed their employees were enthusiastic about the AI rollout, and only thirty-one percent of the actual employees were. That's not a rounding error. That's two different companies occupying the same building. The all-hands slide and your actual desk are more than two times apart, and measured, and the distance is not your imagination.
So no — you did not get bad luck. You got the default. The default is a press release about distributed capability shipped by an organization that has not yet figured out how to live with distributed capability, which is a harder thing to admit than a missing toggle, so nobody admits it. The disabled tool on your screen is the shape of that unfinished thought.
The radio still works
I'm not going to pretend that diagnosis is a solution, and I'm not going to hand you the whole method in a blog post and call it generosity. (It isn't generosity; it's a chapter, and it deserves the room.) But I'll tell you the thing that turns this from a complaint into a door, because it's the thing the working five percent understood and the ninety-five missed.
The radio still works.
Not metaphorically — actually. The fenced tool can still read, still draft, still summarize, still take a messy paragraph and hand it back as a structured one. That is not nothing. That is a wedge. The whole difference between the person who closes the tab and the person who quietly gets faster is not access to a better tool — most days it's the same locked-down radio. The difference is what you decide the radio is "for", and whether you direct it like you mean it or visit it like a tourist and wait to be impressed. One of those people is going to look up in six months noticeably ahead, working inside the exact same constraints as everyone who gave up. I would like it to be you. The move that gets you there is small, and learnable, and it has almost nothing to do with the tool — but it's the whole back half of the argument, and it's the book, not the blog.
For now, the only thing I want you to take to your desk is the correction. The number is in. The tab you closed was not a verdict on you. It was the default, and the default is fixable, and the first repair is to stop reading the locked door as a mirror.
Still reading every rollout for what it quietly disables instead of what it loudly announces, still unmoved by the slide with the upward arrow, still pretty sure the distance between "licensed" and "usable" is the most honest number in the building.
— M.
This is one piece of a longer argument. The whole thing — The Meantime, my book for people learning to direct the AI we actually have — lands in early August. If you'd rather not wait alone, there's a room for people doing this work, comparing the small operator moves that worked this week. It's called the Village, over on Substack chat. → Come in.