AI occupational exposure index explained: why the board asked and Google can't answer it
A People Analytics Lead pulls up a well-known AI exposure index the night before a board deck is due. The chart shows exposure scores by occupation, sorted from highest to lowest. It's a good chart. It's also useless for the question actually on the agenda: which of our 340 roles, in our org chart, should the leadership team be watching over the next 18 months.
This is the gap almost every HR strategy team hits the first time they go looking for hard numbers on AI and jobs. The published indices are real research, built by serious institutions, and they say almost nothing about a specific employer's roster. This article explains what an AI occupational exposure index actually measures, how the major public ones are constructed, and where the line falls between "useful context" and "answer to my board's question."
Inside the major public indices: what they actually measure
Every well-known exposure index — whether from a bank's research arm, a consultancy, or an AI lab — is built at the level of the occupation, not the employer. Goldman Sachs' 2023 analysis estimated that roughly 300 million full-time jobs, in equivalent terms, carry some degree of exposure to generative AI globally. McKinsey's 2023 research on generative AI's economic potential found that 60–70% of employee work hours could theoretically be automated with current and anticipated technology, up from around 50% in earlier estimates that didn't account for generative AI. The Anthropic Economic Index found that 36% of occupations used Claude for at least a quarter of their tasks in early 2025, rising to 49% on a pooled basis as adoption widened.
These are population-level statistics. They describe a labor market, a technology curve, or a national economy — not a company. None of them, including the more recent Anthropic labor-market research showing that 68% of real-world AI usage maps to tasks the researchers judged "fully feasible" for the technology today (with only 3% falling outside feasible use), were built to tell an HR director whether a specific director-level role at a 900-person logistics company should be on a watch list.
That distinction matters more than it sounds. An occupational exposure index is an average across every organization that employs people under that occupational title in national data. A financial analyst at a five-person shop with no forecasting software and a financial analyst inside a 2,000-person company running enterprise planning tools both get folded into the same "Financial Analysts" line. The index tells you something true about the occupation in aggregate. It tells you nothing about which of those two analysts is actually doing routine spreadsheet consolidation versus judgment-heavy scenario work.
The task-level logic behind every credible index
The reason these indices are still worth understanding — rather than dismissing as too abstract — is that the credible ones are built from a real methodological backbone: the ONET occupational database maintained by the U.S. Department of Labor's Employment and Training Administration. ONET catalogs 1,016 occupational titles (923 of them at the data level, representing more than 55,000 individual jobs), described by roughly 277 descriptors covering tasks, skills, knowledge areas, and work context, and it's updated on a regular cycle, with a primary annual update typically landing in the third quarter. The Bureau of Labor Statistics' 2018 Standard Occupational Classification system, which O*NET aligns to, organizes work into 867 detailed occupations, 459 broad occupations, 98 minor groups, and 23 major groups.
This site incorporates information from O*NET. Used under the CC BY 4.0 license. O*NET is a trademark of USDOL/ETA.
Most serious exposure research starts here: it takes O*NET's task and descriptor data for an occupation and scores each task or descriptor against some model of AI capability, then aggregates back up to an occupation-level number. That's a defensible approach — it's the same underlying data infrastructure behind our own methodology at WorkforceAnalysis, described in more detail in AI exposure by occupation explained. The difference is what happens after that aggregation step, and for most public indices, aggregation stops at the occupation. It never comes back down to the specific tasks a specific role at a specific company is actually performing.
Where sector-level indices stop and company-specific questions begin
Brookings' 2024 research on generative AI and the American worker found that the highest-exposure sectors cluster in STEM occupations, business and financial operations, engineering, and law — and separately estimated that roughly 12.9 million U.S. workers, about a third of the people in the occupations Brookings classified as highly exposed, sit in that top-exposure band. The World Economic Forum's Future of Jobs Report 2025 projects a churn of 170 million jobs created and 92 million displaced globally by 2030 — a net gain of about 78 million — alongside a finding that 39% of core skills are expected to be transformed or become outdated in that window. McKinsey's Global Institute research projects the U.S. employment mix shifting toward healthcare, STEM, and managerial roles and away from customer service, office support, and food service roles by 2030.
Every one of these findings is directionally useful. None of them tells a CHRO which of the 40 roles reporting into a specific VP of Operations should move to a "Review" designation this quarter versus which stay on routine "Monitor" status. That's a different question, answered with a different unit of analysis — the actual task mix of the actual role, not the national occupational average.
This is also where "potential impact of AI on US occupations" as a research question and "how to measure AI impact on jobs" as an operational HR task diverge. The first is a macroeconomics and labor-policy question, and the indices above are genuinely good answers to it. If you want the macro framing in more depth, see potential impact of AI on US occupations and how to measure AI impact on jobs. The second is an org-design question that requires task-level, company-specific data no published index collects.
A worked example: from occupational average to role-level scoring
To make the gap concrete, consider a simplified worked example — round numbers, illustrative only. Say a national index reports a hypothetical "Financial Analysts" occupation at a moderate exposure level overall. Inside one specific company, the same job title covers two very different roles:
- Role A spends most of its O*NET-mapped task time on data consolidation and standard variance reporting — tasks that score high on cognitive routine (a rubric dimension scored 0–100 for how repeatable and rule-based the cognitive work is) and low on social/judgment (also 0–100, capturing negotiation, persuasion, and interpersonal judgment).
- Role B, same title, spends most of its time on scenario design for the executive team and cross-functional negotiation over assumptions — low on cognitive routine, high on social/judgment.
A four-dimension exposure rubric — cognitive routine, physical routine, social/judgment, and creative, each scored 0–100 and weighted by each task's O*NET-documented importance to the occupation — would separate these two roles into different exposure bands even though they share a job title and would be averaged into the same line on a published index. Role A might land in a "Review" designation; Role B might stay at "Monitor." Neither designation is a verdict on the person in the seat. It's an input meant to focus a manager's attention and inform a redeployment conversation, not a prediction that a role "will be automated" or is "safe."
Using published indices as context, not as an answer
None of this makes the macro indices wrong or unnecessary — they're the right tool for board-level context about where an industry is headed, and they're the right citation when a leadership team asks "is this a real trend or a fad." Gartner's 2024 research found that 86% of HR leaders have not implemented strategic workforce planning at all, and a separate Gartner HR Priorities survey found 66% describe their workforce planning as limited to headcount forecasting or say they struggle to demonstrate its ROI — evidence that most organizations are still missing the connective layer between macro awareness and role-level action. That connective layer is where task-level, O*NET-grounded scoring earns its keep, and it's covered in practical terms in which jobs are most exposed to AI.
The practical takeaway for a People Analytics Lead heading into a board conversation: cite the published indices for macro credibility, then be ready to explain that your organization's specific exposure picture requires mapping your own roles to O*NET tasks and scoring them individually — because no public index was built to do that for you. Our pricing page outlines the tiers available for teams ready to run that mapping directly, and the Workforce AI Readiness Assessment Guide walks through a structured way to prepare your role data before you start.
If you want these distinctions explained as they develop — new index releases, methodology shifts, and how to read them — subscribe to our newsletter for ongoing coverage.