Is AI unbundling expertise?

Anthropic published a study this week based on roughly 400,000 Claude Code sessions, and it will be cited primarily as evidence that coding has become accessible across professions — that lawyers, managers and accountants now succeed at technical work at rates within a few points of software engineers. Although that finding is valuable, the data actually contains a more significant observation: that AI may be decomposing expertise itself.

For most of human history, expertise bundled three distinct functions in a single person:

• You had to know something — accumulating domain knowledge through experience, training and pattern recognition. 

• You had to reason through problems — applying that knowledge to produce judgments, plans and solutions. 

• You had to transmit your model — making your understanding legible to other minds well enough to coordinate with them, teach them or direct their work. 

Because there was no tool that could perform any of those functions independently, the bundle held together. We treated expertise as a unified phenomenon — and priced it that way.

Externalizing mental models

What the Anthropic study may have unintentionally captured is what happens when that bundle begins to come apart. When they rated "expertise" across 400,000 sessions, the researchers weren't measuring knowledge depth or professional credentials, but were instead measuring precision of instruction, verification behavior, error detection and the ability to redirect the agent when it had misunderstood. 

These are primarily measures of a person's capacity to externalize a coherent internal model — to make the structure of their understanding available to another intelligence, so that intelligence can act on it rather than approximate it. 

The accountant with no coding background who knows exactly which reconciliation rules the script must enforce — and catches the edge case at month-end close — succeeds not because she possesses more domain knowledge than the junior developer, but because she can transmit a precise model of what the problem actually is. The AI actually has something to amplify. However, in sessions where that something is absent, the AI infers and approximates. In turn, the user spends most of the interaction correcting rather than directing.

The transmission of knowledge

This makes transmission legible as an independent economic capability in a way it has never been before. The capacity to reconstruct a coherent model of reality in another mind — whether human or artificial — has always existed. It is what teachers do, what consultants do, what filmmakers and writers and public speakers have always done professionally. It’s what the tribal elder who maintained collective coherence across generations was actually exercising: not superior knowledge of hunting technique, but the highest-bandwidth transmission of the community's operating model — its history, norms, exceptions, relationships and causal chains. Those roles were never primarily defined by possessing knowledge; they were defined by making knowledge transmissible. The bundled nature of expertise meant we never needed to price this capacity separately. The same person who knew things also had to transmit them, so their value was attributed to the knowledge rather than to the transmission.

That attribution is now being tested. The Anthropic data suggests the people extracting the most from AI systems aren’t necessarily those who know the most, it’s those who can construct and deliver a model of sufficient structural clarity that an AI system has something to act on. 

Expertise helps because deep familiarity with a domain usually produces better internal models. But expertise isn’t the multiplier; transmissibility is. The expert who has never been asked to make their reasoning explicit, who has accumulated knowledge through experience but not through articulation, may find the tool unexpectedly disappointing. The gap between what they know and what they can make usable will not show up as a knowledge deficit. It will show up as a productivity ceiling.

Pre-structuring context

I wrote about the architectural dimension of this problem in late 2025, in a piece about teaching AI to skip stones — the argument that success in AI-assisted work doesn't come from better prompts, but from better context selection. The child learning to skip stones doesn't need the physics equation; they need the right contact point at the right angle, drawn from enough embodied understanding to feel the difference. The Anthropic study is now offering empirical confirmation of the same structural observation, at behavioral scale: the humans achieving the best outcomes are those who arrive with a model that is dense enough, structured enough and bounded enough that the AI can amplify rather than infer. What I didn't name clearly enough then is that this capacity — the ability to pre-structure context into a form another intelligence can act on — is separable from expertise itself. It is its own thing.

The strategic implication isn’t primarily about tool adoption. Organizations will move to improve AI fluency, invest in prompting capability, build technical literacy programs. Those investments aren’t wrong but they are downstream of a question most organizations aren’t yet asking: does the workforce entering these interactions have the capacity to construct and transmit a coherent model in the first place?

An organization that has systematically rewarded execution over explicability, where people have been valued for producing outputs rather than for making their reasoning legible, may find it has underinvested in the layer AI actually rewards. The bottleneck in those organizations will be the gap between what people know and what they can transmit. That gap has always existed. It has never, until now, had a direct economic cost attached to it.

Can we cultivate the ability to transmit?

The deeper question the Anthropic study quietly raises is whether transmission can be cultivated deliberately and what developmental experiences produce it. Teaching produces it. Writing produces it. Scientific training, which requires making reasoning independently reproducible, produces it. Leadership, at its best, produces it. 

If transmissibility is now a distinct economic capability — separable from knowledge, separable from execution, valued independently — then the question of how organizations and education systems have been developing it, or failing to, becomes considerably more urgent than it was when it was simply bundled into everything else. AI may be the first technology in history that makes transmission quality directly observable.

Expertise isn't disappearing, but it may be getting repriced. And the components that get priced separately may not be the ones most organizations have been investing in.