The board wants a list, not a feeling
A People Analytics Lead has three weeks before a board meeting. The company's PE owner asked a direct question at the last review: which roles in this organization are exposed to AI, and what is the plan? The honest answer right now is a set of opinions — the CFO thinks finance is fine, the VP of Customer Success is worried about her team, nobody has looked at operations. A feeling is not an answer a board will accept twice.
What the room actually needs is a defensible, repeatable way to say which roles carry the most AI task exposure, in this organization specifically — not a borrowed headline about "40% of jobs" from a research report that was never about this company's org chart. That is the gap between macro AI-exposure research and a usable internal answer, and it is the gap an O*NET-based framework is built to close.
This article lays out that framework: what it draws from, how it scores a role, and — just as important — what the resulting score is not entitled to say.
Why "which jobs will AI replace" is the wrong question
Every few months a new study estimates how much of the economy's work could theoretically be automated. McKinsey has put the technically automatable share of work hours at 60–70%, up from roughly 50% in earlier estimates. Goldman Sachs has estimated AI could affect the equivalent of 300 million full-time jobs globally. The World Economic Forum's Future of Jobs Report 2025 projects 170 million jobs created and 92 million displaced by 2030 — a churn rate of 22%, netting out to roughly 78 million new jobs, alongside a projection that 39% of core skills will be transformed or outdated by then.
These numbers are useful for understanding direction and scale at the level of an economy. They are not useful for telling a single mid-market company which of its roles should be reviewed first, because none of them were computed against that company's actual role list. A "jobs most at risk from AI" headline describes a labor market. A board needs an answer about an org chart.
That is also why "which jobs will AI replace" is the wrong framing to start from. Roles are not atomic units that get replaced wholesale — they are bundles of tasks, and tasks carry very different exposure depending on whether they are routine and well-structured or judgment-heavy and variable. A controller and a staff accountant can report to the same department and carry very different exposure profiles once you look at what each of them actually does day to day. The right question is narrower and more answerable: of the tasks that make up this role, how many carry characteristics that current AI systems handle well, and how many don't?
The O*NET foundation: how occupations become data
Any framework that scores occupational exposure needs a backbone of standardized occupational data, because without one, every analysis becomes a private judgment call. The U.S. Department of Labor's ONET database is that backbone for U.S. labor-market analysis. It currently covers 1,016 occupational titles, of which 923 have full data-level detail, mapped against roughly 55,000 real-world job titles, and it is described using approximately 277 standardized descriptors — things like tasks, required knowledge, skills, abilities, and work context — updated on a regular cycle, with the primary annual update released in the third quarter. [Confirm the current release and figures against the ONET Resource Center before citing externally, as these are revised periodically.]
ONET occupations sit on top of the Bureau of Labor Statistics' Standard Occupational Classification system, which organizes the U.S. labor market into 23 major groups, 98 minor groups, 459 broad occupations, and 867 detailed occupations. The practical effect of this structure is that a role like "Senior Financial Analyst" or "Customer Support Specialist, Tier 2" can be mapped to a specific ONET-SOC code, and from there to a documented list of the tasks that occupation actually performs — not a guess about what the title implies. If you want the fuller mechanics of how a role list gets matched to O*NET codes, see what O*NET is and how it works.
This site incorporates information from O*NET. Used under the CC BY 4.0 license. O*NET is a trademark of USDOL/ETA.
The four-dimension exposure framework
Once a role is mapped to its O*NET tasks, the question becomes: how exposed is each task to current AI capability? A single "automation score" per task collapses too much information, because a task can be routine in one respect and resistant in another. The framework scores every task across four separate dimensions, each on a 0–100 scale:
- Cognitive routine — how standardized and rule-based the thinking involved is (e.g., applying a known formula versus resolving a genuinely novel problem).
- Physical routine — how repetitive and well-defined any physical component is (most relevant for operations and field roles; often near zero for office roles).
- Social/judgment — how much the task depends on reading people, negotiating, persuading, or making a context-dependent call with incomplete information.
- Creative — how much the task requires generating a genuinely new idea, design, or strategy rather than selecting among known options.
These four scores are then weighted by each task's O*NET-documented importance to the occupation, not averaged blindly — a task that occupies 5% of a role's time should not move the needle as much as one that occupies 40%. The result is a role-level exposure profile, not a single number pretending to summarize a whole job. The full mechanics of how task scores are derived and weighted are covered in the task-based AI exposure methodology, and the rationale for scoring on four separate axes rather than one composite number is covered in the four-dimension task exposure framework.
A worked example: scoring two roles side by side
Consider two illustrative roles, scored with round example inputs to show the method, not asserted as measured figures for any real organization.
Accounts Payable Clerk. Core tasks: matching invoices to purchase orders, flagging discrepancies, processing routine payments. Example scoring: cognitive routine 80, physical routine 10, social/judgment 25, creative 5. Task-importance weighting (most time spent on matching and processing, which score highest on cognitive routine) pushes the role-level exposure profile toward the high end on the cognitive-routine axis specifically.
HR Business Partner. Core tasks: advising managers on performance issues, mediating team conflict, designing retention interventions. Example scoring: cognitive routine 20, physical routine 0, social/judgment 85, creative 40. The highest-weighted tasks here score low on cognitive routine and high on social/judgment — the opposite profile.
Both of these are real occupations with real O*NET task lists; the point of the worked example is the arithmetic, not a claim about every AP Clerk or HRBP everywhere. Two roles in the same department, or even the same reporting line, can land in very different places once the scoring runs task by task instead of title by title. For a deeper walkthrough of how a composite score is produced and how to read it, see what an AI exposure score actually means.
From task scores to role-level signal: Monitor, Review, Redeployment Candidate
A role-level exposure profile is only useful if it routes somewhere. Rather than labeling a role with a verdict, the output sorts roles into three tiers that describe what kind of attention they warrant next:
- Monitor — exposure is present but concentrated in a minority of low-weighted tasks; revisit on the normal planning cycle.
- Review — exposure is concentrated in higher-weighted tasks; warrants a closer look at how the role's task mix might be redesigned.
- Redeployment Candidate — exposure is high across the majority of weighted tasks; warrants an active conversation about what adjacent, lower-exposure work this person's skill set could move toward.
None of these tiers says a role "will be automated" or that a person is "safe." They describe where a conversation needs to happen, not how it ends. The distinction between this kind of exposure scoring and a prediction about job loss is worth making explicit, because it is easy to blur the two — see AI exposure versus automation risk for why they are not the same claim.
What a high exposure score does not tell you
A high exposure score is a measurement of task composition against current AI capability, generated at a point in time. It does not account for an organization's specific tooling, its appetite for change management, the tenure and institutional knowledge a particular person carries, or the legal and contractual constraints that attach to any workforce decision. It also does not predict what a company will choose to do — plenty of organizations identify high-exposure roles and choose to redesign the role around the person rather than around the task list.
This matters because of how uneven exposure already looks across the economy. Brookings research has identified STEM, business and financial operations, engineering, and legal occupations as carrying some of the highest generative-AI exposure, affecting an estimated 12.9 million workers — roughly a third of those in the occupations studied. That is sector-level research, not a company-specific finding, and the same caution applies at the organizational level: exposure data tells you where to look, not what to decide. For a broader view of how exposure clusters across occupations generally, see which jobs are most exposed to AI.
Building this into a repeatable practice
The organizations that will handle this well are not the ones with the most alarming headline for their board — they are the ones with a repeatable, defensible process for answering the question when it comes up again next quarter, and the quarter after that, as both the organization and the underlying AI capability keep changing. A one-time analysis answers one board meeting. A standing practice, built on a transparent task-level framework, answers all of them.
If you want this kind of O*NET-grounded exposure analysis explained as it develops — methodology notes, framework updates, and how mid-market HR teams are building it into regular planning — subscribe to the newsletter below.