Most businesses measure AI adoption by counting who has logged in. That number hides the gap that matters. Two people with the same licence can get very different value from it, and the one who already knew how to think through a problem gets far more. AI multiplies the skill you bring to it. Like any multiplier, it widens whatever gap it starts with.

Anthropic’s own data points the same way. Its Economic Index report in March found that people who’d been using Claude for six months or more were seven percentage points more likely to use it for work. They were more likely to iterate on what it gave them, and their conversations succeeded more often, even after allowing for the kind of task. Usage was also concentrated in high-income countries and, within the US, in places full of knowledge workers. The promised leveller is so far favouring the people who were already ahead.

Anthropic’s head of economics, Peter McCrory, told TechCrunch there’s no material difference yet in unemployment between the most exposed workers and everyone else. The divide is showing up in what people can do long before it shows up in who keeps their job. It’s a slightly odd warning to get from the company selling the tool, whose chief executive, Dario Amodei, said last year that AI could wipe out half of all entry-level white-collar jobs within five years. Still, I’d rather they published it than not.

Anthropic can’t tell whether experienced users do better because they started out sharper or because practice made them so, and I can’t either. If it’s practice, the gap is something a business can close. The people with most to gain are junior staff, who have the least judgement to multiply and are already facing a graduation cliff. Left to work it out alone, they’re the ones most likely to fall behind.

That makes the login count the wrong number to watch.

A login tells you somebody opened the tool. It doesn’t tell you whether they pushed back on the first answer, or whether the work came out any better. Ask a few people to share examples they’re happy to show, and look at two things: how often they went past the first response, and whether the finished task was better or faster than it would have been without AI. You’ll probably find a few people using it as a partner and a larger group using it as a quicker search box.

Pair them up so the casual users can watch what the others do. It’s the same show-don’t-train fix I argued for in March, when I wrote about people using a Ferrari like a bicycle. Then look again in a month.

Handing out access was only ever the first step. The businesses that measure how their people use AI, and help the rest catch up, will end up with one capable team rather than two tiers.