When AI Automates Your Tasks, What Remains Determines Your Fate

A recent working paper by David Autor and Neil Thompson (MIT, June 2025) argues that automation’s impact on wages and employment depends not on which tasks are removed but on whether those tasks were expert or inexpert relative to the occupation’s baseline. When AI strips away inexpert tasks, the remaining work concentrates around higher-skill demands — wages rise but employment falls as fewer workers clear the bar. When it removes expert tasks, the opposite occurs: entry barriers drop, employment expands, and wages decline. The framework resolves longstanding puzzles about why routine-task automation produced divergent wage outcomes across occupations, and it offers a concrete lens for assessing which roles AI will upgrade versus commoditize.

The Core Insight: It’s about “What Remains”

Conventional wisdom treats automation exposure as symmetric: two occupations losing the same types of tasks should fare similarly. Autor and Thompson show this is wrong. What matters is not which tasks are automated, but what the remaining tasks demand in expertise. Consider their running example. Accounting clerks and inventory clerks both had routine, codifiable tasks automated over the past four decades — recording transactions, reconciling statements, compiling inventories, verifying stock conformance. Standard task models predict similar outcomes for both. The expertise framework predicts the opposite — and the data confirm it.

When automation stripped away routine bookkeeping from accounting clerks, it left behind the hard stuff: problem-solving, judgment calls, specialized analysis. The occupation became more expert, wages rose, and employment fell as fewer workers could clear the higher expertise bar. Inventory clerks experienced the mirror image. Automation removed their most expert tasks (flagging items below government support prices, for instance), leaving behind stocking, counting, and weighing. The occupation became less expert, wages declined, but employment expanded as the barrier to entry dropped.

The counterintuitive result: expertise changes resemble labor supply shifts. Rising expertise requirements contract the qualified workforce, pushing wages up and employment down. Falling expertise requirements expand the qualified workforce, pushing wages down and employment up. These effects are distinct from — and for employment, opposite to — the effects of simply losing or gaining tasks.

Empirically, a 1σ rise in occupational expertise predicts an 18% wage increase and a roughly 5% employment decline per decade. Task quantity changes matter too but work through a different channel: losing tasks contracts employment (a demand shift), while gaining expertise contracts employment (a supply shift). The two forces are only weakly correlated, meaning both carry independent explanatory power.

The framework also resolves a longstanding puzzle — why routine-task-intensive occupations saw employment decline but wages did not uniformly fall. Routine tasks were expert in some occupations and inexpert in others, so their removal raised expertise demands in one group and lowered them in another, splitting wages accordingly.

The Anthropic Economic Index report (January 2026) corroborates this heterogeneity with real usage data. Claude tends to cover tasks requiring higher education levels — 14.4 years on average versus 13.2 for the economy-wide baseline — and removing those tasks produces a net deskilling effect across most occupations. But the pattern is not uniform. For some roles like property managers, AI handles routine bookkeeping and leaves higher-skill negotiation and stakeholder management behind, producing upskilling instead.

How to Think About Your Own Job

The key questions to ask yourself: (1) Which of my tasks are most exposed to automation — and are those my most or least specialized responsibilities? (2) After removing those tasks, does what remains require more or less expertise than my current average? (3) Would that shift expand or narrow the pool of people who could do my job?

Here is an example with my role as an economist at AWS. AI is already handling tasks that are supporting rather than core — data compilation, formatting, routine descriptive analyses, literature review. What remains and intensifies are the higher-expertise components: identifying the right analytical framing, interpreting ambiguous evidence, connecting technical findings to strategic decisions, communicating nuanced conclusions to non-technical leadership. At least for now this looks like expertise augmentation — the automation is stripping away the inexpert tasks and concentrating the role around judgment and synthesis. But the framework’s employment prediction follows directly: if the remaining task bundle demands more expertise then the qualified labor pool narrows and fewer positions are needed to cover the same output. I see early signs of exactly this at Amazon today: economist job openings have noticeably declined while the scope and responsibilities have not.

Open Questions

Two threads I plan to explore further. First, whether we can construct a practical AI risk index for a given profession by measuring the expertise level of tasks most exposed to current AI capabilities relative to the occupation’s average. The Autor-Thompson methodology — content-agnostic word expertise measurement combined with embedding-based task matching — provides a template, but adapting it to AI-specific exposure (rather than historical computerization) requires mapping frontier model capabilities onto task taxonomies in real time.

Second, the political economist in me cannot help but think about the political implications of deskilling. If AI systematically removes higher-skill tasks from many white-collar occupations — as the Anthropic data suggests is already the default pattern — the resulting compression of expertise demands could generate a large constituency of workers experiencing downward wage pressure even as aggregate productivity rises. The distributional politics of that transition, particularly for college-educated workers who have historically been better insulated from automation relative to blue-collar workers, could reshape the political landscape, likely with increased demand for redistribution and support for populist policies.