AI Exposure by Occupation: What It Measures and What It Doesn't

What does "AI exposure by occupation" actually mean?

You've seen a headline ranking the jobs "most exposed" to AI, and now you're trying to figure out what that actually means for your organization. Maybe a board member forwarded the article with a one-line question attached. Maybe a department head asked, half-joking, whether their team is on the list. Either way, you're now the person who has to explain what "AI exposure by occupation" measures — and what it doesn't.

This is a reasonable thing to be unsure about, because the term gets used loosely. Some outlets treat exposure as a synonym for "at risk of elimination." Others use it interchangeably with "automation risk," a related but distinct concept. Neither usage is quite right, and the difference matters a great deal if you're the one who has to present this analysis to leadership.

At its core, AI exposure by occupation is a measurement of task composition, not a forecast of job loss. It asks a narrow, answerable question: of the tasks that make up a given occupation, how many involve the kind of work that current AI systems can plausibly perform or substantially assist with? It does not ask, and cannot answer, whether an employer will choose to automate that work, whether budget exists to do so, or whether a specific person in that role will keep their job. Those are business decisions, not measurements — and conflating the two is where most of the public conversation about AI and jobs goes wrong.

This article walks through what exposure scoring actually measures, how a defensible framework builds that score from the task level up, and where the boundary sits between "exposure" as an analytical input and "outcome" as a leadership decision.

Inside the four-dimension exposure framework

Occupational exposure frameworks generally start from the same place: the U.S. Department of Labor's ONET database, the standardized taxonomy that has organized U.S. occupations by their component tasks, skills, and work activities for over two decades. ONET currently describes 1,016 occupational titles covering 923 data-level occupations, representing more than 55,000 individual jobs, using roughly 277 descriptors per occupation, refreshed on a rolling basis with a primary update each third quarter. Occupations map onto the 2018 Standard Occupational Classification, which organizes work into 23 major groups, 98 minor groups, 459 broad occupations, and 867 detailed occupations. This is public infrastructure, maintained by the U.S. DOL's Employment and Training Administration, and it is the reason exposure scoring can be consistent across employers rather than ad hoc.

This site incorporates information from O*NET. Used under the CC BY 4.0 license. O*NET is a trademark of USDOL/ETA.

O*NET gives you the raw material — the list of tasks that make up an occupation, and how important each task is to that occupation overall. What it does not give you is an exposure score. That requires a scoring layer on top, and the one we use organizes every task along four dimensions, each scored 0–100:

• Cognitive routine — how much of the task involves structured, rules-based information processing: data entry, scheduling, standard reporting, document review against a checklist.
• Physical routine — how much of the task involves repeatable physical actions in a stable, predictable environment.
• Social and judgment — how much of the task requires interpersonal negotiation, contextual ethical judgment, or reading a room — the kind of work that resists clean rule-based description.
• Creative — how much of the task requires generating genuinely novel output rather than recombining known patterns.

A task that scores high on cognitive routine and low on social/judgment — say, reconciling expense reports against policy — looks very different from a task that scores high on social/judgment and low on cognitive routine, such as mediating a dispute between two direct reports. Both might appear in the same job description. Scoring them separately, rather than assigning one number to the whole role, is what makes the output useful instead of just directional. For a deeper walkthrough of how this rubric is applied task by task, see our four-dimension task exposure framework breakdown.

From task to occupation: a worked example

Here is how the four dimensions roll up into a single occupation-level exposure figure, using a simplified, illustrative example — not a real scored occupation.

Suppose an occupation's O*NET task list includes five core tasks, each carrying a DOL-assigned importance weight (tasks are not equally central to the job):

To compute the occupation's cognitive-routine exposure score, you multiply each task's score by its importance weight and sum the results:

(0.30 × 85) + (0.25 × 70) + (0.20 × 25) + (0.15 × 80) + (0.10 × 15) = 25.5 + 17.5 + 5.0 + 12.0 + 1.5 = 61.5

Repeat the same weighted calculation for physical routine, social/judgment, and creative, and you get four dimension scores instead of one flattened number. That matters because a role can score high on cognitive-routine exposure while scoring low on social/judgment exposure — which is a materially different situation from a role that scores high on both. Averaging those into a single "exposure percentage" would erase exactly the distinction a workforce planner needs to see.

This is also why task-level, importance-weighted scoring behaves differently from headline-style claims about "which jobs are most exposed to AI." A ranked list of occupations is a reasonable starting point for orientation, but it necessarily strips out the task-level detail that determines whether a specific team's version of that role looks like the sector average — or not. If you want the full mechanics of how weighting, task matching, and dimension scoring interact, our task-based AI exposure methodology piece goes further into the underlying math.

What an exposure score does not tell you

This is the part of the conversation that gets skipped most often, and it's the part that matters most if you're the one presenting this to a room that includes a CFO or a board member.

An exposure score does not tell you:

• Whether the role will be eliminated. Exposure measures task composition, not organizational intent. A high cognitive-routine score on a task does not mean the employer has decided, or is even considering, automating it.
• Whether the person in the role is replaceable. Exposure is scored at the occupation and task level, not the individual level. Two people in the same title can carry very different mixes of responsibility, tenure-based institutional knowledge, and adjacent duties not captured in a standard task list.
• Timing. Exposure scoring says nothing about whether a shift happens this quarter, next year, or not within a planning horizon at all. It is a snapshot of task composition against current AI capability, not a projection curve.
• Budget or intent. Whether an organization has the capital, the change-management appetite, or the strategic reason to act on exposure findings is a leadership decision entirely outside the scope of the measurement.

Multiple research organizations — including teams at Goldman Sachs, McKinsey, the World Economic Forum, and Anthropic — have published economy-level and sector-level estimates of how AI might affect the volume and composition of work over time. These are valuable for understanding macro direction: which broad sectors and skill categories show the highest concentration of exposed tasks, and how skill requirements are shifting. But they are calibrated at the level of the U.S. or global economy, not your company. None of that research substitutes for looking at your own role list, your own task mix, and your own department structure — which is a different, narrower, and more actionable question. Our company-specific AI workforce exposure guide walks through why the jump from sector research to your own org chart requires its own analysis, not an extrapolation.

Exposure is an analytical input to human judgment. It is not a verdict on any role, and it is not a substitute for the decision a leadership team still has to make.

Exposure vs. automation risk: two different questions

"AI exposure" and "automation risk" get used as if they mean the same thing, and treating them as synonyms is one of the more common ways this analysis gets misread.

Exposure asks: how much of this occupation's task composition involves the kind of work current AI systems can perform or meaningfully assist with? It is a description of task content, measured against present-day capability.

Automation risk, as the term is typically used in labor economics literature, asks a broader and more speculative question: given technical, economic, regulatory, and organizational factors combined, how likely is it that this work gets automated within some time horizon? That second question folds in variables an exposure score deliberately excludes — cost of implementation, regulatory constraint, union agreements, customer preference for a human, and dozens of organization-specific factors that no occupation-level dataset can capture.

Treating the two as interchangeable produces exactly the kind of overclaiming this analysis is designed to avoid. An occupation can carry a high cognitive-routine exposure score and still carry low near-term automation risk, because the organization has no current plan, budget, or business case to act on it. Conversely, a role with moderate exposure scores might see rapid change if an adjacent process shift makes automation suddenly cheap or urgent. Exposure is one input among several a workforce strategy team needs — not the whole picture. We go deeper on this distinction, including where the two measures tend to diverge in practice, in AI exposure vs. automation risk.

Reading the output: Monitor, Review, Redeployment Candidate

Because exposure is an input rather than a verdict, the language used to report it matters as much as the scoring method itself. A defensible framework never labels a role "safe" or "automated" — both terms imply a certainty the measurement doesn't support, and both invite exactly the kind of reaction a workforce strategy team is trying to manage carefully. Instead, exposure output is typically organized into three categories that describe posture, not fate:

• Monitor — the role's current task composition shows low-to-moderate exposure across the four dimensions. No immediate action is indicated; periodic re-scoring as the role or the technology changes is appropriate.
• Review — the role shows meaningful exposure concentrated in specific tasks, warranting a closer look at whether those tasks could be restructured, reassigned, or supported with tools, without implying the role itself is in question.
• Redeployment Candidate — the role's task composition is heavily weighted toward high-exposure work, which makes it a candidate for a deliberate internal conversation about how the person's skills map to other roles in the organization — a conversation the organization chooses to have, informed by the data, not one the data has already had for them.

Notice what none of these categories say: they don't say who keeps their job, when a change happens, or whether the organization acts at all. They describe where a role sits on a task-composition spectrum, so that a workforce strategy team can decide, with better information, what to do next. For a full explanation of what the scores mean and how to talk about them internally — including with employees who may see their own role's category — see what an AI exposure score means.

Why occupation-level data is a starting point, not a company-specific answer

Occupation-level exposure research — including the public rankings you may have started from — is useful for orientation. It tells you, in broad strokes, which sectors and occupational families carry the highest concentration of exposed tasks, and it's a fair place to start a conversation with a board member who wants context before specifics. If your board member's question was some version of "which jobs are most exposed to AI," a general answer grounded in published occupational research is a reasonable first response — our overview of which jobs are most exposed to AI covers that ground.

But a general occupational ranking cannot tell you how your accounts-payable team's actual task mix compares to the ONET-defined average for that occupation, whether your customer-service function has quietly absorbed judgment-heavy escalation work that a generic job title doesn't capture, or which of your departments carries the highest concentration of Review or Redeployment Candidate roles once your specific role list is mapped and scored. That requires taking your organization's actual roles, matching each one to its closest ONET occupation, applying the four-dimension scoring to your specific task mix, and rolling the results into a department-level view your leadership team can actually act on.

That mapping and scoring work is what WorkforceAnalysis is built to do — transparently, at the task level, without ever converting a score into a verdict about a person or a role. If you want to see how the methodology described here applies to your own organization's role list, or you'd simply like to keep learning about exposure scoring, workforce planning research, and how to talk about this analysis with your leadership team, subscribe to our newsletter for ongoing methodology breakdowns like this one — and if you want a structured way to think through your organization's AI readiness before you run any scoring at all, the Workforce AI Readiness Assessment Guide is a good next step.