If you run a business where your judgment is the product, your people are almost certainly using AI by now. This is showing up in more licenses, more daily use, and a sense that work is moving faster than it did a year ago (not to mention Anthropic's share price). But if a customer asked you tomorrow what your business can do now that it could not do a year ago, what would you say? Almost all the leaders I talk to can point to a great deal of activity. Far fewer can name a capability they have actually gained.
This gap matters, and almost nobody is measuring it. Adoption is high, but capability is not. Most organisations can't tell the difference, so they measure the thing that is easy to measure, use, and treat it as a proxy for the thing that is hard, capability.
It is not the same, and the distance between them is wider than the usage numbers suggest.
This shows up with many of the clients I work with (although in less extreme ways than the "tokenmaxxing" you read about in the media). In a series of hands-on sessions with the leadership team of a professional services firm two patterns showed up early. The first is the enthusiastic daily user whose use never widens. One partner began as the most sceptical person in the room and became, by frequency, the fastest adopter of the lot. She told me she had been using it a lot more since we started, and she had. But when we analysed what she was doing, it largely stayed in one lane: looking things up and checking things, the same narrow task done more often. The volume was there but the capability (and therefore usefulness) had plateaued.
The second pattern is more subtle, and the firm's chief executive named it herself. She had built genuine personal value out of the tool, working with it most days. When she reflected on how, she said what she had been doing was "real piecemeal, little pieces of bits and pieces" - she would develop something useful, then forget to carry a part of it forward, so nothing compounded into anything the firm could hold and reuse. Months of activity. Very little that accumulated into capability the firm owned.
Both of these feel like progress from the inside, and both would show up as adoption wins on any usage dashboard. And for professionals early in their use of AI, it should be taken as the positive step that it is. What's important is that businesses begin to look at the usefulness of AI, not just its use. In a perfect world, we could run an AB test or a time series analysis to isolate the impact of AI on the P&L. But the world is rarely perfect.
The reason this persists is not that leaders are careless. It is that capability, in judgment work, is genuinely hard to measure, while usage is easy. Faced with a number they can collect and a number they can't, most organisations collect the one they can and hope it stands in for the other. For many leaders it is worse than a private misread: usage is the number they end up reporting upward, to a board or an owner who will take it for capability. It doesn't, and the obvious fallback of asking people how capable they have become is the least reliable signal of all. A 2026 study in Computers in Human Behavior found that when people used AI, the usual link between skill and self-awareness broke down. Everyone tended to overrate how well they had done, and the more AI-literate someone was, the more confident, and the less accurately, they judged their own performance. The people leaning on AI most heavily are often the least able to tell you what it is actually doing for them. So you are left without the two measures most organisations reach for: usage tells you how much, not how well, and self-report is misleading at best. Which raises the obvious question: what does AI capability actually look like, and how would you know it when you saw it?
From my experience, it shows up in how someone works, not in how often they reach for the tool. So you need to stop counting and start watching. Across the businesses I work with, capability shows up along a number of observable dimensions: how someone sets up the work (the context they give the tool, how they frame the task, how fluently they drive it), how they verify what comes back, whether they know where the tool reliably fails, how they relate to it (as an oracle to trust or a clerk to check), how their use is deepening over time, and how they handle client confidentiality in practice. Not scored. Observed. The point is not the framework. It is that each of these is something you can watch a person do, in real work, where a usage dashboard sees only that they logged in and how many chats they have had.
What you find when you watch is that capability and usage come apart, often dramatically. The clearest moment in the engagement came from the partner I had been warned was the resistant one. Her objection, it turned out, had nothing to do with whether AI worked. She had spent years getting her team to a standard she trusted - as she put it, if someone brings her work they did without AI, she is confident they researched it, thought about it, and used their own analytical skills. Her resistance was not anti-AI. It was a quality guardian protecting the thinking discipline of the people she leads. What she cared about was exactly the thing a usage metric cannot see: whether the judgment behind the work was still there.
The session that moved her did not start with a tool at all. It started with her actual daily friction, which was managing people, not the technical work her training was built around. We worked on her real problem, and the AI produced something concrete enough that she could push back on it - "that's good," she said, "it's concrete enough to argue with." The shift from arms-length scepticism to engagement happened inside a single exercise, and what moved was not her usage. It was her ability to judge what she was looking at. That is capability, and you could see it happen.
The honest caveat is that this is early. Two months in, I can show you the gap between usage and capability, and moments where capability visibly moved. What I cannot yet show you is the long arc - the practitioner who can reliably evaluate what she could not before, measured over time and visible in the metrics that matter to the organisation. That evidence takes longer to mature than a single engagement, but the direction is already clear.
Which is where most leaders get stuck. Evaluating your people one by one was never going to scale, and it asks you to grade a thing you are not yet equipped to see. So the question isn't how to measure each person. It is how to build the conditions in which capability develops, and how to make sure you would recognise it when it does. That starts closer to home than most leaders expect: the fastest way to learn what capable AI practice looks like is to do enough of the work yourself to recognise it, which is why the leaders getting real value are usually the ones who have put in the hours, not the ones who delegated it to a working group. From there, the questions become organisational rather than personal. What does the firm actually train for, beyond a tool rollout and a policy memo? What does it reward - the person who got the work out the door, or the person who can show how their judgment got sharper? What good practice does it make visible, so that one person's hard-won approach becomes something the firm holds rather than something that leaves when they do?
I do not have a tidy answer to those, and I am suspicious of anyone who says they do this early. But the firms treating capability as something to build deliberately - through how they train, what they reward, and what they choose to make visible - rather than something they hope arrives with the licences, will be the ones that can tell the difference between a workforce that is busy with AI and one that has actually become more capable. Right now, most can't tell which they have.