The Expert May Matter More—and Less—at the Same Time

In customer support, generative AI raised productivity most for novice and lower-skilled workers. In Anthropic’s study of Claude Code usage, people with greater task-specific domain expertise were more likely to finish successfully and could direct longer chains of work, although the gap between intermediate and expert users was modest.

One reading says expertise is being compressed. Another says expertise earns a return. The contradiction softens when expertise stops being treated as one indivisible thing. Producing an answer, defining the problem, recognizing an error, and deciding whether the result is fit for use are different parts of performance.

AI can substitute for one layer while complementing another. Which layer matters most depends on the task, its representation, the quality of feedback, and the consequence of being wrong.

Brynjolfsson, Li, and Raymond studied 5,179 customer-support agents during a staggered rollout of a generative-AI assistant. Access increased issues resolved per hour by 14 percent on average and by 34 percent among novice and lower-skilled workers, with minimal effects for experienced and highly skilled workers.

In that environment, the system could suggest responses and retrieve relevant documentation during repeated textual interactions. Some performance once accumulated through experience became available inside the workflow. The novice did not need to generate every response or remember every policy unaided.

That is genuine substitution for parts of execution and retrieval. It does not show that every underlying capability transferred or that unusual, high-consequence work would produce the same distribution of gains.

Anthropic analyzed roughly 400,000 Claude Code sessions from about 235,000 people between October 2025 and April 2026 using privacy-preserving, model-based classification. In the typical session, the person made about 70 percent of planning decisions—what to do—while Claude made about 80 percent of execution decisions—how to do it.

The division is not a universal law of agentic work. It describes one product, period, and classification framework. Still, it gives concrete shape to a pattern executives often discuss vaguely: implementation can move toward the model while task direction remains substantially human.

The people bringing more task-specific domain expertise tended to achieve higher success and recover more effectively from errors or misunderstandings. Yet the difference between intermediate and expert users was modest. Deep mastery did not produce an unlimited advantage.

Anthropic’s classifier inferred expertise from how precisely a person framed directions, what they asked the system to verify, and whether the user corrected the model or the model corrected the user. A senior software engineer could be a novice in an unfamiliar domain; an accountant with little programming experience could be expert in the reconciliation rules the code had to enforce.

That distinction changes the workforce question. The scarce person may not be the one who can type the implementation. It may be the one who can state the true constraint, notice that a month-end edge case was lost, or recognize that passing tests do not answer the operational requirement.

As agents improve, some of those activities may also move toward the model. The study itself proposes tracking whether returns to expertise decline over time. Any claim that domain experts necessarily become more valuable should therefore carry a date and a measurement plan.

Consider a leader asking an agent to build a forecast. Retrieval and standardized production may become much cheaper. The agent can assemble data, draft code, run a model, and produce a presentable output. The human contribution may shift toward deciding which business mechanism the forecast represents, whether a discontinuity makes historical data misleading, and what consequence follows from an error.

That shift does not automatically make the human contribution more valuable. It can shrink the total labor required, move authority, or expose the framing itself to automation. Nor does cheaper production guarantee the organization retains enough skill to diagnose a strange failure when normal feedback breaks.

The useful operating question is granular: which layer of expertise is the tool absorbing, which layer is it amplifying, and which layer remains necessary for independent evaluation or recourse?

The customer-support study is causal evidence in a bounded workplace setting. Anthropic’s report is large-scale product telemetry analyzed by the company that makes the product, using model-based classifications of sessions and outcomes. It is descriptive rather than economy-wide causal evidence. The two sources deserve different weights and answer different questions.

For continuity planning, the implication is to preserve neither every old skill nor only the glamorous layer called judgment. Keep enough unaided and independent capability to define consequential requirements, test unusual cases, recognize material error, and govern escalation—where the task actually demands it. Let tools remove dependencies they can demonstrably remove.

Skagway will continue treating this as a research question. The balance among execution, framing, evaluation, and recovery may move quickly, and conclusions should move with the evidence rather than with a preferred slogan.