The question your board is actually asking
A People Analytics Lead sits down three weeks before a board meeting. The board's PE sponsor has asked a version of the same question every portfolio company is getting this quarter: how exposed is our workforce to AI, and what are we doing about it? The instinct is to open a browser tab and reach for a headline statistic — a percentage from a research report, a number that sounds precise enough to put on a slide. That instinct is understandable, and it is also the fastest way to present a number that means nothing about the company in the room.
The problem is not a shortage of research. It is a mismatch between the level at which that research operates and the level at which the board's question actually lives. This article walks through the credible ways to measure AI's impact on jobs — from economy-wide indices to sector studies to company-specific, task-level scoring — and explains where each one is useful, where it stops being useful, and how to choose the right approach for the decision in front of you.
Three levels of measurement, and why they don't substitute for each other
Most public discussion of "AI and jobs" collapses three distinct levels of analysis into one number, which is where the confusion starts.
Macro indices estimate exposure across the entire economy or a large labor market. Goldman Sachs has estimated that AI could affect roughly 300 million full-time-equivalent jobs globally, and McKinsey has estimated that 60–70% of work hours could theoretically be automated with current and emerging technology, up from roughly 50% in earlier estimates. The World Economic Forum's Future of Jobs Report 2025 projects 170 million jobs created and 92 million displaced by 2030 — a net gain, but with a churn rate (22%) that implies substantial disruption inside that net figure. These numbers are valuable for understanding direction and scale. They say almost nothing about whether your VP of Underwriting's role is a Monitor, a Review, or a Redeployment Candidate.
Sector and occupation-family research narrows the aperture. Brookings has identified STEM, business and financial operations, engineering, and legal occupations as the highest-exposure sectors, with roughly 12.9 million workers — about a third of those in the occupations studied — classified as highly exposed. The Federal Reserve's 2026 monitoring work similarly finds professional services and financial-sector employers showing the highest work-related AI adoption, in the range of 5%–40% depending on sector. This is closer to useful. If you run a professional-services or financial firm, this research tells you that your sector is a reasonable place to look first. It still doesn't tell you which of your 40 job titles carry the exposure and which don't.
Company-specific, task-level scoring is the third level, and it is the only one that answers the board's actual question. It requires mapping your real role list to standardized occupational data, then scoring the tasks within each role — not the job title as a whole — against a defined exposure rubric. This is measurably more work than pulling a statistic from a report. It is also the only approach that produces something you can act on role by role, department by department.
The board isn't asking what AI is doing to the labor market. It's asking what AI is doing to this organization's task mix — a question no macro index was built to answer.
Why occupation titles are the wrong unit of measurement
A common shortcut is to take a macro or sector exposure percentage and apply it to a job title: "financial analysts are estimated at X% exposure, so our twelve financial analysts are at X% exposure." This substitutes the wrong unit of analysis. A job title is a bundle of tasks, and within any real occupation, some tasks are highly routine and language- or pattern-based, while others require in-person judgment, negotiation, or physical presence. Two people with the same title in two different companies — or two different departments of the same company — can have meaningfully different task mixes.
This is exactly the gap that standardized occupational data is built to close. The U.S. Department of Labor's ONET database catalogs 1,016 occupational titles and 923 data-level occupations covering more than 55,000 job types, described through roughly 277 descriptors per occupation, updated on a regular cycle with a primary annual refresh. Critically, ONET doesn't just name occupations — it breaks each one into its constituent tasks and rates how important each task is to that occupation. That task-importance data is the raw material for a defensible, occupation-level exposure model. Our companion piece on task-based AI exposure methodology walks through the full mapping process in detail; this article covers what to look for in any methodology, task-based or otherwise.
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 dimensions worth scoring — and a worked example
A credible task-level exposure model scores each task against more than one axis, because "automatable" is not a single quality. A useful framework scores four dimensions, each on a 0–100 scale:
- Cognitive routine — how repetitive and pattern-based the mental work is
- Physical routine — how repetitive and pattern-based the physical work is
- Social/judgment — how much the task depends on human relationship, negotiation, or contextual judgment
- Creative — how much the task depends on original synthesis or ideation
Each task within an occupation gets scored on these four dimensions, then weighted by how important O*NET says that task is to the occupation overall, and rolled up into an occupation-level (and eventually role-level) score.
Here's a simplified worked example, using round numbers to illustrate the arithmetic rather than a real occupation. Suppose a role has three core tasks. Task A makes up 40% of the role's importance weighting and scores 80 on cognitive routine. Task B makes up 35% of the weighting and scores 60. Task C makes up the remaining 25% and scores 30. The weighted cognitive-routine score for this role is:
(0.40 × 80) + (0.35 × 60) + (0.25 × 30) = 32 + 21 + 7.5 = 60.5
Repeat that calculation across all four dimensions and you get a four-number profile for the role — not a verdict, a profile. Our explainer on AI exposure by occupation, explained covers how these dimension scores differ across occupation families, and why cognitive-routine and social/judgment scores often move in opposite directions within the same department.
What a good measurement gives you — and what it deliberately withholds
This is the point at which the distinction between an exposure score and an outcome prediction matters most. A task-level exposure score tells you how much of a role's task mix resembles work that current AI tools handle well. It does not, and should not, tell you that a role "will be automated" or is "safe." Those are judgment calls that depend on variables no data model has — your company's specific technology roadmap, budget, change-management capacity, union agreements, customer relationships, and leadership priorities.
That's why the useful output of a measurement exercise is a categorization that supports a decision, not a label that replaces one. A well-built model sorts roles into something like Monitor (low exposure across dimensions, low near-term action needed), Review (moderate-to-high exposure on one or more dimensions, warrants a closer look at the specific tasks driving the score), and Redeployment Candidate (high exposure concentrated in cognitive- or physical-routine tasks, paired with meaningful strength on social/judgment or creative tasks elsewhere in the person's skill set — a redeployment conversation, not a headcount decision). If you want to see how these categories play out across a full role list, which jobs are most exposed to AI breaks down the occupation families where Review and Redeployment Candidate outcomes cluster most often.
This distinction is not a rhetorical hedge. Gartner's 2024 research found that 86% of HR leaders have not implemented strategic workforce planning, and 66% say their planning work is effectively limited to headcount forecasting and struggles to demonstrate ROI. An exposure score that gets treated as a verdict feeds that same shallow headcount-only conversation. An exposure score that gets treated as an input — attached to a redeployment plan, a reskilling budget line, or a hiring-freeze decision on a specific req — is the difference between a compliance exercise and an actual workforce strategy.
Choosing the right approach for the decision in front of you
If you're preparing a board briefing on directional risk across the industry, a macro index or sector study is legitimate context — cite it as exactly that, one analyst firm's estimate, not a claim about your company. If you're deciding where to focus reskilling budget, hiring freezes, or redeployment conversations for your specific role list, you need task-level, company-specific scoring; nothing else answers that question. Many organizations do this today through a manual consulting engagement or an internal analyst project mapping roles to ONET data by hand in a spreadsheet — a legitimate approach, though it tends to be resource-intensive and hard to repeat as the org chart changes. A self-serve tool built specifically to run this mapping at mid-market scale is the other option, and it's the approach WorkforceAnalysis was built around: ONET-grounded, task-level, transparent about the four dimensions behind every score.
Whichever path you choose, the discipline that matters most is keeping the levels separate. Quote the macro research to explain trend. Quote sector research to explain where to look first. Then do the task-level work before you put a number next to any real person's role. If you want a structured way to walk your own leadership team through that process, the Workforce AI Readiness Assessment Guide in our store lays out the discovery questions worth asking before you run any scoring exercise, and our pricing page outlines how the self-serve model works if you decide task-level scoring is the right next step.
To keep learning how these methodologies evolve — including new occupational data releases and changes to sector-level research — subscribe to our newsletter below.