AI Workforce Risk Index: Which Occupations Carry the Most Exposure

Why an exposure number is not a risk number

The most widely circulated US figure on AI and work is that roughly 80 percent of the workforce has at least 10 percent of their tasks exposed to generative AI, with about 19 percent of workers facing exposure on more than half of their tasks (Eloundou, Manning, Mishkin and Rock, 2023). That number is real, and it is widely cited, but it does not measure risk. Exposure is the technical question: can a current model perform this task. Risk is the human question: what would happen to a particular worker, in a particular place, if the model did perform the task in production.

The Workforce Risk Index is built to translate exposure into risk. It multiplies three components: the technical exposure score from the Eloundou methodology, a wage-vulnerability component that captures how damaging a displacement event would be at the local wage and benefits level, and a reskilling-difficulty component that captures how hard it would be for an affected worker to move into an adjacent occupation. The resulting index is bounded between 0 and 100 and is published at the occupation level, with optional weighting to the local labor market through the Bureau of Labor Statistics OEWS tables.

The point of the Index is to identify occupations that policymakers, unions, community colleges, and Workforce Innovation and Opportunity Act (WIOA) intermediaries should treat as priorities. High exposure on a high-wage occupation with abundant reskilling pathways is not where the most damage will land. The most damage lands where exposure is high, wages are low, and reskilling is hard.

The three components, in detail

The exposure component uses the published Eloundou et al. occupation-level scores, anchored to O*NET task descriptions. The score for each occupation is the share of its constituent tasks that a current frontier model can perform at human-equivalent quality without requiring physical presence. This is the same series used by Brookings Metro in its Geography of Generative AI series, and it has the advantage of being reproducible from public inputs.

The wage-vulnerability component combines three sub-inputs: the occupation's median wage from OEWS, the share of workers in the occupation without employer-provided health insurance from the Bureau of Labor Statistics Employee Benefits Survey, and the share of workers without a college degree from the Census Bureau's American Community Survey occupation tabulations. The intuition is that a displacement event is more damaging when the worker had thinner wage margins to begin with, when the loss of employment also means the loss of health coverage, and when the worker's credentials offer less optionality in adjacent labor markets.

The reskilling-difficulty component uses the O*NET career-cluster transition matrix to ask, for each occupation, how many adjacent occupations exist that meet two conditions: they have a wage within 80 percent of the source occupation, and they require a skill overlap of at least 60 percent. Occupations with many such adjacencies score low on reskilling difficulty. Occupations with few or no such adjacencies score high.

The Index is the geometric mean of the three normalized components, multiplied by 100. The geometric mean is the right aggregator here because the components are conceptually multiplicative: an occupation with low exposure but high vulnerability is not high-risk, and neither is an occupation with high exposure but easy reskilling.

What the Index says about specific occupations

Several patterns are durable across plausible weightings of the components.

The highest-risk cluster contains administrative-support occupations with exposure scores in the top quartile, median wages in the lower-middle of the distribution, and limited reskilling adjacencies. Order clerks, file clerks, billing and posting clerks, and certain customer-service representatives sit in this group. The exposure case is well documented in the recent NBER paper on the labor-market impact of generative AI by occupation. The wage-vulnerability case is supported by BLS Employee Benefits Survey data showing that these occupations have lower employer-coverage rates than the overall private workforce. The reskilling case rests on the O*NET observation that the closest adjacencies for these occupations are within the same low-wage administrative cluster.

A second high-risk cluster contains specific knowledge-work roles at the lower end of the professional wage band. Paralegals and legal assistants, junior financial analysts in back-office functions, and entry-level marketing roles are exposed at near the top of the distribution. Their wage vulnerability is lower than the administrative cluster, but their reskilling adjacencies are narrower than is sometimes assumed, because the closest available adjacencies are often other exposed roles.

A third cluster, which most exposure-only rankings miss, contains medium-exposure occupations with very high vulnerability and very narrow reskilling adjacencies. Cashiers in non-retail settings, customer-service representatives in regulated industries, and certain teaching-assistant roles fall here. Their exposure scores are not at the top, but their wage vulnerability is at the top, and their reskilling options are few.

The lowest-risk cluster contains exposed but high-wage occupations with abundant reskilling adjacencies. Lawyers, financial managers, software developers, and management consultants have high exposure but also the deepest reskilling pathways and the strongest wage margins. Exposure does not equal harm when workers can absorb it.

How geography enters

The Workforce Risk Index is published at the occupation level, but it can be weighted to a metro or county by multiplying by local OEWS employment shares. A metro whose top-ten occupations are dominated by high-risk roles will have a much higher metro-level risk weight than a metro whose top occupations are concentrated in high-exposure but low-risk roles like software development or legal services.

This is where the Index becomes operationally useful. The metro of Phoenix, with very large concentrations of customer-service and back-office occupations, scores meaningfully higher on the population-weighted risk metric than the metro of San Francisco, despite San Francisco having higher headline exposure on the occupation-level Eloundou score. The same logic produces high metro-level risk weights in parts of metropolitan Atlanta, Tampa, Dallas, and Cleveland. The geography of risk is not the geography of exposure.

Methodological caveats worth printing

Three caveats matter for anyone using the Index in a real decision.

The exposure scores are a snapshot of current model capability. They will move year over year as models improve. The exposure component should be re-estimated annually, not assumed to be stable, and any policy intervention that takes more than 24 months to deploy should be calibrated against a forward expectation, not the current score.

The wage-vulnerability component is a relative ranking inside the United States. Cross-country use of the Index requires reweighting against local wage floors and welfare-state coverage. Comparing US administrative-support risk to German administrative-support risk on this Index without that reweighting is a category error.

The reskilling-difficulty component depends on the ONET task and skill database. ONET is the best public source available, but its skill tagging is updated on a slow cycle and does not yet reflect changes in skill bundling that the rise of AI tools is producing. As AI itself changes the skill content of work, the adjacency calculation will need recalibration. The Acemoglu and Restrepo research program has the most rigorous treatment of how this recalibration should be approached (Acemoglu, 2024).

What this means for policy

Three operational moves follow from reading the Index seriously.

First, target the high-vulnerability mid-exposure cluster, not the high-exposure cluster. The administrative-support and customer-service occupations have the worst combination of all three components. They are also the easiest political constituency to support, because their geographic concentration is in mid-sized metros with active workforce-development institutions. The places identified in our Geography of Opportunity Rankings are not random; they correlate with where the Workforce Risk Index runs highest.

Second, redesign reskilling around adjacencies, not occupations. A community-college program that trains displaced billing clerks to become medical billing clerks moves them from one high-risk role to another. A program that trains them to become licensed practical nurses, ultrasound technicians, or HVAC apprentices moves them out of the high-risk cluster entirely.

Third, sequence wage-floor and benefit reforms to the timing of displacement. If the displacement cycle is multi-year, as the OECD Employment Outlook expects, then unemployment-insurance reform, portable benefits, and wage-insurance pilots have time to be designed and tested. The OECD 2023 Employment Outlook covers the policy literature on this in detail.

For broader context on the structural conditions that shape where these risks land, see our companion pieces: AI Adoption Index for the demand side, State AI Readiness Rankings for the local infrastructure picture, and Human Opportunity Index for the upstream conditions of mobility and access.

Sources

• Eloundou, Manning, Mishkin and Rock, *GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models*, 2023.
• Felten, Raj and Seamans, *Occupational, industry, and geographic exposure to artificial intelligence*, NBER 2023.
• Bureau of Labor Statistics, Occupational Employment and Wage Statistics, Employee Benefits Survey, and Occupational Projections.
• US Department of Labor, O*NET Task and Skill Database.
• Acemoglu, *The Simple Macroeconomics of AI*, NBER 2024.
• OECD, *Employment Outlook 2023*.
• Brookings Metro, *The Geography of Generative AI*.