The board meeting question no one has a clean answer to
A People Analytics Lead at a 900-person professional services firm gets a note from the CFO three weeks before the next board meeting: the board wants to know which departments carry the most generative AI exposure, and what the plan is. The instinct is to reach for headcount data. The problem is that headcount tells you nothing about task composition — and task composition is where generative AI exposure actually lives.
Knowledge work is not one thing. A financial analyst, a paralegal, a customer support lead, and a UX researcher all sit under the "knowledge worker" umbrella, but their task mixes look nothing alike, and neither does their exposure. This article walks through how generative AI exposure shows up specifically in knowledge work, which task types drive it, and how to start assessing it for your own roles without guessing.
Why knowledge work reads differently than physical labor
Most public discussion of "AI and jobs" still defaults to a factory-floor mental model — robots on an assembly line. That model was never built for generative AI. Large language models do not manipulate objects; they manipulate language, structured data, and pattern-based judgment calls. That means their exposure footprint concentrates in exactly the tasks that make up most knowledge work: drafting, summarizing, classifying, researching, and producing first-pass analysis.
Brookings' 2024 research found that the highest-exposure sectors are concentrated in STEM, business and financial operations, engineering, and law — occupational families built almost entirely from cognitive tasks. The same research estimated that roughly 12.9 million workers, about a third of the workforce in those occupations, are highly exposed. That is a meaningfully different exposure profile than physical-labor-heavy occupations, where task exposure tends to concentrate in the routine (not the physical) tasks within an otherwise hands-on role. We go deeper on that contrast in AI exposure: white collar vs. blue collar jobs — the short version is that exposure tracks task type, not job title or collar color.
McKinsey's 2023 analysis put a number on the scale of the shift: it estimated that 60–70% of work hours across the economy could theoretically be automated with current and anticipated generative AI capability, up from roughly 50% in earlier estimates that didn't account for generative AI. That is a theoretical ceiling on automatable hours, not a forecast of who loses a job — an important distinction we'll come back to.
The task types that actually drive exposure
Occupation titles are a poor unit of analysis. The unit that matters is the task — which is exactly why ONET, the U.S. Department of Labor's occupational database, structures its data around detailed task statements rather than job titles alone. ONET currently documents 1,016 occupational titles (923 with full data-level detail) covering more than 55,000 job types, described through roughly 277 descriptors per occupation and updated on a quarterly cycle with its primary annual refresh in the third quarter. The Bureau of Labor Statistics' 2018 Standard Occupational Classification, which O*NET maps to, organizes work into 23 major groups, 98 minor groups, 459 broad occupations, and 867 detailed occupations.
That level of granularity is what makes it possible to separate a role into its component tasks and score each one, rather than applying a single label to an entire job. We use four dimensions to do that scoring, each rated 0–100:
- Cognitive routine — how much of the task is pattern-based analysis, classification, or drafting against known templates.
- Physical routine — how much of the task involves repetitive physical manipulation.
- Social/judgment — how much of the task depends on negotiation, persuasion, mentorship, or contextual human judgment.
- Creative — how much of the task requires genuinely novel synthesis or ideation, not recombination of known patterns.
For a deeper walkthrough of the first dimension specifically — which is usually the highest-scoring one for knowledge work — see cognitive routine tasks and AI automation. Each task within an occupation is weighted by its O*NET-documented importance to the role before the four dimension scores roll up into an occupation-level exposure profile. A role is never one number; it's a profile across four dimensions, task by task.
Augmentation and redeployment look different at the task level
The Anthropic Economic Index's 2025 research found that 36% of occupations, in early 2025 data, used Claude for at least a quarter of their associated tasks — rising to 49% in pooled data across the period studied. A 2026 follow-up from Anthropic broke down how that usage actually functions: 68% of usage occurred on tasks the researchers classified as fully feasible for the model to handle, while only 3% of usage occurred on tasks classified as not feasible. That gap — most usage sitting in "feasible," a small remainder in "not feasible" — is a more honest description of what's happening in knowledge work than a binary "AI replaces jobs" framing.
The World Economic Forum's Future of Jobs Report 2025 offers the most detailed public breakdown of how this shift is expected to redistribute work by 2030: 170 million jobs created, 92 million displaced, a net gain of 78 million, and 22% total workforce churn. The same report estimates that 39% of workers' current skills will be transformed or become outdated by 2030, and that of every 100 workers needing training, 59 will need it — 29 upskilled in their current role, 19 reskilled or redeployed internally, and 11 unlikely to receive the training at all. Separately, McKinsey's Global Institute projected that overall U.S. employment mix will shift toward healthcare, STEM, and managerial roles and away from customer service, office support, and food service roles by 2030.
None of this tells you which specific role in your organization should move from "monitor" to "review" status. It tells you the macro direction knowledge work is moving. The task-level view is what closes that gap — and it's the difference between an augmentation candidate and a redeployment candidate. We unpack that distinction fully in which jobs will AI augment vs. replace.
A worked example: scoring a knowledge-work role
Here's a simplified, illustrative walkthrough of how task-level scoring works in practice — not a claim about any real occupation's published score, just a worked example with round numbers.
Say a financial analyst role has four core O*NET-documented tasks, each with an importance weight (how central the task is to the job) and a cognitive-routine score (0–100):
| Task | Importance weight | Cognitive routine score |
|---|---|---|
| Compile and reconcile monthly variance reports | 0.35 | 85 |
| Build and maintain financial models | 0.30 | 70 |
| Present findings to business unit leaders | 0.20 | 25 |
| Advise on non-standard deal structuring | 0.15 | 15 |
A weighted cognitive-routine score for the role is calculated as: (0.35 × 85) + (0.30 × 70) + (0.20 × 25) + (0.15 × 15) = 29.75 + 21 + 5 + 2.25 = 58.
That single dimension score of 58 doesn't move the role into any category on its own — it sits alongside the role's physical-routine, social/judgment, and creative scores to form a full exposure profile. Two of the four tasks in this example (variance reporting, financial modeling) pull the cognitive-routine score up; the other two (presenting to leadership, advising on non-standard deals) pull it down because they depend on judgment and relationship context that current generative AI tools don't perform reliably. That's the mechanism — not a verdict that the role is "safe" or "at risk," but an input a workforce strategy team can use to decide whether the role warrants ongoing monitoring, closer review, or active redeployment planning.
Turning this into a workforce decision, not a workforce guess
Exposure is an analytical input to human judgment — never a substitute for it.
The organizations that get the most value from an exposure exercise are the ones that resist the urge to translate a score directly into a personnel decision. A high cognitive-routine score on a task means that task is a strong candidate for augmentation tooling, closer process review, or a redeployment conversation — categories we describe elsewhere as Monitor, Review, and Redeployment Candidate. It does not mean the person in that role is expendable, and it does not mean the organization has a forecast of who will lose their job. If your team is building its first department-level exposure heatmap, AI exposure by occupation, explained is a useful next stop — it covers how occupation-level scores roll up from individual tasks.
If you're preparing your own board-ready exposure review, our pricing page outlines the self-serve tiers built for exactly this kind of mid-market workforce assessment — no consulting engagement required. For a structured, printable framework you can walk a leadership team through before you run your first assessment, the Workforce AI Readiness Assessment Guide is available in our store.
Want more breakdowns like this one, including new task-type research as it's published? Subscribe to our newsletter for methodology updates and new exposure research as we release it.
This site incorporates information from ONET. Used under the CC BY 4.0 license. ONET is a trademark of USDOL/ETA.