Anthropic published a study showing the gap between what AI can do across every profession, and what workers actually use it for.

The chart is striking. A huge blue area, theoretical coverage, sprawling across management, legal, business, education, healthcare. A tiny red shape in the middle; actual observed usage almost entirely concentrated in software and coding.

The intended read: look at all this untapped potential. The opportunity is massive.

But I've been sitting with the other interpretation. The one the authors don't dwell on.

What if the gap isn't opportunity?

Coders adopted AI because the feedback loop is brutal and honest. Code compiles or it doesn't. The test passes or it fails. There's no ambiguity. When AI makes you faster, you feel it immediately. You keep using it.

For everyone else, the GTM lead, the lawyer, the financial analyst, the nurse, the feedback loop is diffuse. The output is a memo, a strategy, a recommendation. Quality is judged by humans, delayed, political. When AI saves you 30 minutes on a deliverable nobody reads carefully anyway, you don't feel it. You go back to your old workflow.

This isn't a capability problem. It's a legibility problem.

And the study can't really distinguish between: "people haven't tried AI here yet" versus "people tried AI here and found it wasn't worth it." Both produce the same gap in the data. Both look identical on a radar chart.

There's a third interpretation too, and this one is uncomfortable for everyone selling enterprise AI:

A lot of knowledge work is not actually task execution. It's relationship management, judgment under ambiguity, political navigation, and trust-building. None of that shows up in a job task taxonomy. It doesn't get a spoke on the radar chart. But it's often the majority of what a senior person in business, law, or healthcare actually does all day.

AI is genuinely excellent at the legible, decomposable parts of knowledge work. It's still largely irrelevant for the rest.

The honest version of the chart would include a third line: tasks that are in principle AI-addressable but where humans have evaluated AI and chosen not to use it. My guess is that line would sit much closer to the red than the blue.

None of this means enterprise AI adoption won't grow. It probably will. But the honest timeline looks more like a decade than a year. Adoption follows trust, and trust follows demonstrated ROI in high-ambiguity environments, which is exactly the hard part.

The companies that will actually close the gap aren't the ones selling the chart as evidence of opportunity. They're the ones doing the boring, patient work of figuring out where the feedback loops are tight enough for AI to prove itself, one workflow at a time.

That's less exciting than a radar chart. But it's the actual job.