The question every board is asking right now
A People Analytics Lead we'll call the reader is three weeks from a board meeting. The company's private equity owner has asked a version of the same question every portfolio company is hearing this year: which of our roles are exposed to AI, and what are we doing about it? The instinct is to reach for a headline — "AI could replace 300 million jobs" or "40% of work hours are automatable" — and paste it into a slide. That instinct is understandable and it is also the wrong move. Macro research describes the economy. It does not describe your org chart.
This article does two things. First, it lays out what credible research actually says about which occupations show higher AI exposure, and why. Second, it explains why that pattern — real as it is — cannot substitute for looking at your own roles, because exposure is a function of task composition, and task composition varies company to company even within the same job title. By the end, you'll know what the research supports, what it doesn't, and where to look next.
Why some occupations show up as high-exposure again and again
Exposure research keeps landing on the same broad pattern: work that is high in routine cognitive tasks — data processing, drafting, retrieval, structured analysis — shows up as more exposed than work built around physical dexterity in unstructured environments, or around judgment calls that carry legal, financial, or human stakes.
Goldman Sachs' 2023 analysis estimated that roughly 300 million full-time jobs (in equivalent terms) globally carry some exposure to generative AI automation. McKinsey's 2023 research on generative AI's economic potential found that 60–70% of work hours could theoretically be automated with current and emerging technology, up from about 50% in pre-generative-AI estimates — a jump driven specifically by AI's new ability to handle natural-language and knowledge tasks that used to be considered automation-resistant.
Anthropic's own usage data adds a task-level view rather than a theoretical one. Its 2025 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. A 2026 follow-up found that 68% of that usage fell into tasks the model could handle fully, versus only 3% falling into tasks it could not meaningfully assist with at all — a useful reminder that "exposure" is rarely all-or-nothing at the task level; it's a spectrum from full automation to light augmentation.
Brookings' 2024 research on generative AI and the American worker names the sectors where this shows up most: STEM, business and finance, engineering, and law post consistently higher occupational exposure than other sectors, with an estimated 12.9 million workers — roughly a third of workers in those occupations — classified as highly exposed. The Federal Reserve's 2026 note on AI adoption corroborates this from the usage side: work-related AI adoption ranges roughly 5% to 40% across the economy, with professional services and financial-sector firms showing both the highest current adoption and the strongest growth.
None of this means "white-collar work is disappearing." It means routine cognitive tasks inside white-collar occupations are the tasks large language models are currently best at handling. That distinction matters, and it's worth its own read if you want the fuller picture — see our companion piece on AI exposure across white-collar and blue-collar work.
The white-collar surprise: why exposure isn't about collar color
The popular framing — "AI is coming for white-collar jobs, robots already came for blue-collar ones" — oversimplifies what the data shows. Physical-routine work in structured, predictable environments (certain assembly and inspection tasks) has long carried automation exposure through robotics and machine vision, independent of generative AI. What's new is that generative AI extends exposure into cognitive-routine tasks that used to be considered safely knowledge-work: first-draft writing, structured research synthesis, code scaffolding, basic financial modeling.
McKinsey's Global Institute research projects the employment mix shifting by 2030 toward healthcare, STEM, and managerial roles, and away from customer service, office support, and food service roles — a mix of both physical-routine and cognitive-routine occupations losing share, for different underlying reasons. This is a sector-level projection, not a verdict on any specific role, and it's a useful frame precisely because it resists the "one collar color, one outcome" story.
PwC's Global CEO Survey data tracks the executive sentiment side of this. In the 27th edition (2024), 25% of CEOs expected AI to reduce their own headcount by 5% or more over the coming year, while 39% expected headcount to increase by 5% or more — both directions moving simultaneously, which is consistent with a redeployment story more than a pure-reduction one. The 28th edition (2025) found 13% of CEOs reporting an actual headcount reduction attributable to generative AI (16% in the most-affected sectors), against 17% reporting an increase. Executives, in other words, are not describing a uniform contraction. They're describing a redistribution, sector by sector and task by task — which is exactly why a single "most exposed jobs" list undersells the complexity your board is actually asking about.
What "exposure" actually measures — and what it doesn't predict
This is worth stating plainly, because it's the single most common misreading of this research: exposure is not a prediction of who gets laid off. It is a measure of how much of an occupation's task content overlaps with what current AI systems can plausibly perform. High exposure can just as easily mean "this role's tasks will shift toward oversight, exception-handling, and judgment" as it can mean "this role's headcount will shrink." Both outcomes are consistent with the same exposure score. The research measures overlap, not fate.
The World Economic Forum's Future of Jobs Report 2025 frames this well at the macro level: it projects 170 million jobs created and 92 million displaced globally by 2030 — a churn rate of 22% of total employment, netting to +78 million jobs overall. It also finds that 39% of workers' current skills will be transformed or become outdated by 2030, and that of every 100 workers, 59 need training, but only 29 are likely to get upskilled in their current role, 19 are likely to be reskilled and redeployed internally, and 11 are unlikely to receive adequate training at all. That gap — 63% of employers cite skill gaps as their biggest transformation barrier — is where workforce strategy actually lives. Exposure tells you where to look. It doesn't tell you what to do next; that's a judgment call, made by people, informed by the data.
This is also why we're careful with language throughout this article and on the product itself. A role scoring high on exposure metrics gets flagged for review — not declared automated, not declared safe. Some roles land as candidates for closer monitoring. Others may surface as redeployment candidates, where task overlap with adjacent occupations suggests a reasonable internal move. None of that is a verdict. It's an input into a conversation your leadership team still has to have.
A worked example: scoring two occupations side by side
To make this concrete, here's a simplified version of how a four-dimension exposure rubric works — cognitive routine, physical routine, social/judgment, and creative task content, each scored 0–100, then weighted by how important each task is to the occupation (per O*NET's own task-importance ratings, not just how often it's performed).
Take two illustrative, rounded examples — not measured occupational averages, just a demonstration of the method:
- Occupation A (a data-entry-heavy administrative role): cognitive routine 85, physical routine 15, social/judgment 20, creative 10. Its highest-importance tasks are almost entirely cognitive-routine, so the weighted exposure score sits high.
- Occupation B (a client-facing account management role): cognitive routine 40, physical routine 5, social/judgment 75, creative 35. Its highest-importance tasks involve negotiation and relationship judgment, so the weighted score sits meaningfully lower — even though both roles might carry the same job title at different companies.
That last point is the whole reason a generic "most exposed jobs" list can mislead a specific organization: two people with the title "Account Manager" can have genuinely different task mixes, and therefore genuinely different exposure scores. Title-level research is useful for spotting economy-wide patterns. It cannot substitute for scoring your own roles against your own task descriptions. For a fuller walkthrough of how the scoring model maps roles to O*NET occupations and computes these weights, see how AI exposure by occupation is actually calculated.
Why the same occupation title can score differently between companies
ONET — the U.S. Department of Labor's occupational database — covers 1,016 occupational titles across roughly 923 data-level occupations representing over 55,000 individual jobs, each described by around 277 descriptors covering tasks, skills, knowledge, and work context. It's updated on a rolling basis, with a primary annual update typically landing in the third quarter. This depth is what makes task-level exposure analysis possible at all — but it's also a reminder that ONET describes an occupation in the aggregate, based on the typical task mix reported across the labor market, not your specific team's day-to-day workload.
This site incorporates information from ONET. Used under the CC BY 4.0 license. ONET is a trademark of USDOL/ETA.
That's the gap between macro research and organizational decision-making. Macro research (Goldman, McKinsey, Anthropic, Brookings, WEF, PwC, the Federal Reserve) tells you where exposure clusters across the economy. It cannot tell you whether your specific "Account Manager" role, with your specific task mix, sits closer to Occupation A or Occupation B above. That answer requires mapping your actual role list to O*NET occupations and scoring the resulting task content — which is a different exercise than reading a headline, and a faster one than commissioning a full workforce consulting engagement. For the fuller numeric picture behind the "how many jobs" framing specifically, see how many jobs will AI replace by 2030, and for the underlying displacement figures in one place, AI job displacement statistics.
Turning research into a decision your board can act on
Reading the research is step one. Step two is applying the same logic — task composition, not job title — to your own organization's role list, department by department. That's a different kind of exercise than a literature review, and it's the one your board actually wants an answer to.
If you want to see how this works on a single role before mapping a whole department, our AI Exposure Scorecard walks through the four-dimension scoring model on one occupation, with the same O*NET-grounded methodology described above. For teams ready to move to a full department or company-wide view, pricing outlines the tiers built for that scale.
Want the research summarized like this delivered before it becomes a board-meeting fire drill? Subscribe to our newsletter for grounded, sourced updates on AI workforce exposure — no invented statistics, no hype, just what the data actually supports.