The expertise framework is a task-level account of AI's labor-market effects advanced by economists David Autor and Neil Thompson in their 2025 paper "Expertise" and an essay in Digitalist Papers Vol. 2. It distinguishes the expert tasks within an occupation, which require specialized knowledge or training, from the inexpert tasks, which require common skills most workers possess, and predicts an occupation's wage and employment trajectory from which of the two AI automates.
Core distinction and prediction
The framework holds that the consequences of automating an occupation depend on which of its tasks are automated. Automating an occupation's inexpert tasks concentrates the remaining work on its expert tasks, raising the barrier to entry: wages rise and employment falls. Automating an occupation's expert tasks makes the remaining work more accessible to generalists, lowering the barrier to entry: wages fall and employment rises.
| Which tasks are automated | Effect on remaining work | Effect on wages | Effect on employment |
|---|---|---|---|
| Inexpert tasks | More concentrated on expert tasks | Rise (higher barrier to entry) | Fall (fewer qualified workers) |
| Expert tasks | More accessible to generalists | Fall (more workers qualified) | Rise (lower barrier to entry) |
Autor and Thompson present the framework as resolving a 20-year puzzle in the routine-task-automation literature: why some routine-task-intensive occupations saw wages rise while others saw wages fall. In their account, the determining factor is which routine tasks were automated — expert routine work (for example, assembly-line precision) versus inexpert routine work (for example, data entry).
Empirical evidence
Autor and Thompson base the framework on a four-decade analysis of 303 US occupations (1980–2018). A one-standard-deviation higher expertise requirement correlates with 16–31% higher wages even after controlling for education, and changes in expertise requirements predict wage changes (a +1σ change in expertise associated with an +18% wage change).
Two occupations illustrate the divergent paths. Accounting clerks (1980 → 2018) retained their expert tasks as routine-arithmetic and filing work was automated and complex reconciliation remained; wages rose 39% while employment fell 32%. Inventory clerks (1980 → 2018) had their expert tasks automated by inventory-management systems and pricing calculations, leaving mainly physical stocking work; wages fell 13% while employment rose 175%.
AI-era applications
Autor and Thompson frame the framework as recasting the question "will AI destroy jobs?" into "which tasks within an occupation is AI automating?" — with the answer determining whether the occupation becomes more selective and better-paid or more accessible and lower-paid.
They apply this to several occupations as illustrative scenarios rather than predictions. For radiologists, AI that automates image-reading might remove one expert task; if consultation, cross-disciplinary coordination, and patient communication remain, radiology could become more expert (higher wages, fewer slots), whereas if AI erodes the diagnostic moat entirely, radiology could become more accessible (lower wages, more slots). For software engineers, AI coding assistants that automate inexpert tasks (boilerplate, syntax lookup) could concentrate remaining work on expert tasks such as architecture and debugging complex systems, raising wages for senior engineers and reducing demand for juniors; alternatively, AI that automates some expert tasks (common algorithms, refactoring) could broaden access to the field. For lawyers, AI legal research that automates inexpert associate tasks could concentrate senior-lawyer judgment tasks, raising wages for partners and reducing associate demand.
Two scenarios for policy
Autor and Thompson frame two scenarios that, in their account, call for different responses. The first, which they label "fail to imagine the future of expertise," is a transition-management problem: workers need help adapting to changing expertise bundles within occupations. The second, "labor obsolescence," arises if AI eventually strips economic value from all human expertise — a state they connect to Herbert Simon's "intolerable abundance" — and which they describe as raising three surpassing challenges: social organization, income distribution, and democratic stability. They argue affluent countries can hedge against both scenarios while many developing economies cannot.
Relation to other frameworks
| Framework | Core lens | |
|---|---|---|
| Skill-biased technical change (earlier Autor work) | Technology complements high-skill, substitutes low-skill. | |
| Polarization/hollowing-out (Autor, Katz, Kearney 2006+) | Routine tasks automated; low and high ends grow. | |
| [[machine-platform-crowd-book | Machine-Platform-Crowd]] (McAfee/Brynjolfsson) | Cognitive substitution + platform dynamics + crowd sourcing. |
| [[bounty-and-spread | Bounty and Spread]] (Brynjolfsson/McAfee) | Aggregate gains with distributional inequality. |
| Expertise framework (Autor/Thompson 2025) | Task-level expert-vs-inexpert distinction determines within-occupation wage/employment dynamics. |
Autor and Thompson position the expertise framework as a finer-grained mechanism that builds on earlier polarization models rather than contradicting them.
Relationships
- supports: The Digitalist Papers (Stanford, Volumes 1–2) — Autor/Thompson Vol. 2 essay.
- supports: Applying AI to Rebuild Middle-Class Jobs — Autor's related work on middle-class jobs.
- supports: AI Labor Disruption — provides the mechanism; concept page should reference this framework.
- supports: Generative AI at Work / Generative AI and the Nature of Work — Brynjolfsson empirical work testing similar mechanisms.
- related: Bounty and Spread — complementary at the macro level.
- related: The Simple Macroeconomics of AI — Acemoglu's task-level framework is a close cousin.
- related: (Source: epoch.ai) / Anthropic Economic Index — January 2026: Economic Primitives / Anthropic Economic Index — March 2026: Learning Curves — empirical Anthropic data that can be interpreted through the expertise lens.
- related: GDPval Paper (OpenAI, Oct 2025) — benchmarks for AI performance on real professional tasks.