In January I wrote that people who resist AI tools aren’t being stubborn; a lot of them are grieving a version of themselves that used to be worth something. I still think that’s true. I wanted to see whether the numbers back it up, so I went and read the studies I’d been half-quoting at people for months.
Start with practice, because that’s the old premise. We all grew up with some version of the 10,000-hour rule: put in the time, suffer a bit, and competence is what you get back. In 2014 Brooke Macnamara, David Hambrick and Frederick Oswald pooled the research on deliberate practice and performance (the meta-analysis in Psychological Science). Practice explained 26% of the differences in performance in games, 21% in music, 18% in sports, 4% in education, and less than 1% in professions. So time mattered, but it was never the whole story. It was the best proxy we had for something harder to name, and a lot of us (me included) built careers and a fair bit of self-worth on the proxy.
Then the tools arrived, and the first good study of what they did at work is the one I keep coming back to. Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed 5,172 customer support agents as a company rolled out a generative AI assistant (Generative AI at Work, Quarterly Journal of Economics, 2025; free version on arXiv). On average agents resolved about 15% more issues per hour. The gain wasn’t spread evenly, though. Less experienced and lower-skilled agents got faster and better. The most experienced ones got a little faster and their quality slipped slightly.
That’s the bit that explains the grief for me. The tool lifted the people at the bottom of the curve and did very little for the people at the top, so the distance between them, the thing the experienced ones had spent years building, got shorter.
And it happened fast. Alexander Bick, Adam Blandin and David Deming at the St. Louis Fed found that by August 2024 about 39% of Americans aged 18 to 64 had used generative AI, roughly two years after ChatGPT came out (The Rapid Adoption of Generative AI). For work use they compare it with the PC, which sat at about 20% three years after the first mass-market machines. I don’t know of anything we’ve adopted that quickly, which might be why nobody, me included, has a good map for it yet.
The third study is the one I’d misremembered, and I want to correct that here. Eleanor Dillon, Sonia Jaffe, Nicole Immorlica and Christopher Stanton ran a six-month randomised experiment with Microsoft’s Copilot across 66 firms and more than 7,000 knowledge workers (Shifting Work Patterns with Generative AI). I’d been telling people it showed that only the workers willing to rewire their work pulled ahead. It doesn’t say that. What it found is smaller and, I think, more interesting: the people who used the tool spent about two hours less a week on email and did less work outside regular hours, but the overall mix of what they did didn’t change in any way the researchers could detect. People changed the parts of their work they could change alone. The parts that depend on other people stayed where they were.
So where does that leave me? Honestly, a bit less certain than the version of this post I first wrote. What I learned shaping digital transformations didn’t vanish; it moved. Less of it is remembering how things work, more of it is judging what comes back and knowing which piece goes to which tool. Those are skills too, and the experienced people I know who are doing well with this are leaning on exactly that.
And I can’t take much credit for whatever head start I’ve got. I didn’t earn it the old way, with the hours. I just kept playing with the tools after the novelty wore off for most people, the way I used to change one line of BASIC to see what happened.
The Dillon study is the one I’ll keep an eye on. If the parts of work that need other people stay put, I’d expect the next few years to get interesting there… or stuck.
If this poked at something you're building: talk to me.
MAKE YOUR CASE.