Which Jobs Will AI Augment vs Reshape?

The question a board member actually asks

A VP of People walks into a leadership offsite with a simple mandate: tell us whether AI is going to augment our people or replace them. It is a fair question to ask and a nearly impossible one to answer at the level it is usually asked. "Our customer success team" is not a unit of AI exposure. The tasks inside that team are. One rep spends her mornings drafting renewal emails and her afternoons talking a frustrated enterprise buyer off the ledge. The first task looks like automation. The second looks like judgment no model produces on its own. Same job title, two very different exposure profiles.

This is the trap in the "augment vs. replace" framing: it assumes a job is one thing. It isn't. A role is a bundle of tasks, and each task carries its own mix of routine and judgment. The more useful question — and the one this article answers — is not "will this job be augmented or replaced," but "which of this job's tasks look like augmentation, which look like reshaping, and what does that mean for how the role should be evaluated going forward."

Why the binary breaks down

"Augment vs. replace" implies two buckets and a clean line between them. Real occupational data doesn't support a clean line. McKinsey's 2023 research on generative AI found that 60–70% of employees' work hours could theoretically be automated with current and emerging technology, up from roughly 50% in pre-generative-AI estimates — a jump concentrated in tasks that used to be considered judgment work: drafting, synthesis, first-pass analysis. That's not a story about entire jobs disappearing. It's a story about the boundary between "routine" and "judgment" tasks moving, occupation by occupation.

Anthropic's Economic Index tells a similar story from the usage side rather than the theoretical-capability side. Across a large sample of real Claude conversations, roughly 68% of usage fell into tasks the researchers classified as fully feasible for the AI to handle end to end, while only about 3% were tasks the model could not meaningfully assist with at all. The overwhelming majority of real-world usage sits in between — a human directing, checking, and finishing what the model started. That middle ground is augmentation in practice, not a hypothetical.

Meanwhile the World Economic Forum's Future of Jobs Report 2025 frames the shift at the labor-market level: it projects 170 million jobs created and 92 million displaced by 2030, a churn rate of about 22% of today's total employment, netting out to roughly 78 million additional jobs globally. The same report estimates that 39% of workers' current skills will be transformed or become outdated by 2030. Read together, these numbers describe an economy where most roles are being reshaped in place — task by task — far more often than they are being deleted wholesale.

None of this tells you what's happening inside your organization's org chart. That requires looking at the task, not the headline.

The unit of analysis: tasks, not titles

This is where O*NET-grounded exposure scoring earns its keep. O*NET breaks occupations down into the discrete tasks that make up the job, each with a documented importance weighting to the overall role. WorkforceAnalysis scores each of those tasks against four dimensions, each rated 0–100:

• Cognitive routine — how standardized and rule-based the mental work is
• Physical routine — how standardized and repeatable the physical work is
• Social/judgment — how much the task depends on negotiation, persuasion, mentorship, or contextual judgment
• Creative — how much the task depends on original synthesis, design, or novel problem-solving

A task scoring high on cognitive routine and low on social/judgment and creative looks like a strong automation candidate. A task scoring low on both routine dimensions and high on social/judgment or creative looks like a strong augmentation candidate — the kind of work where AI tools accelerate a person's output without approaching a substitute for their judgment. Most real tasks land somewhere in between, which is exactly why augmentation and automation are better understood as two ends of a spectrum than two separate outcomes.

A worked example

Take a mid-level financial analyst role. Suppose its O*NET task list includes four representative tasks, weighted by importance to the role:

The first two tasks — reconciliation and templated forecasting — score high on cognitive routine and low on judgment. They are strong candidates for augmentation tools that draft, check, and flag anomalies, freeing the analyst's time. The last two — presenting and advising — score the opposite way. No template replaces the judgment call about which forecast assumption to challenge in front of a skeptical VP. Blend the four tasks by their importance weight and the role's composite exposure score sits in the middle of the range — not because the analyst's job is "half automated," but because the role is a genuine mix of routine and judgment work that should be read task by task, not title by title.

This is a worked illustration to show the method, not a claim about any specific occupation's real O*NET score — but it demonstrates why the same job title can house both augmentation candidates and judgment-anchored work that no scoring model should call "safe."

White collar isn't a proxy for exposure, and blue collar isn't a proxy for safety

One of the more persistent misreadings of AI exposure is that it maps neatly onto white-collar versus blue-collar work — that office jobs are all at risk and physical jobs are insulated. The occupational research doesn't support that shortcut. Brookings' 2024 analysis found that the sectors showing the highest exposure are STEM, business and financial operations, engineering, and law — categories that sit squarely inside "white collar" but vary enormously in how much of their task mix is routine versus judgment-heavy. The same analysis estimated that roughly 12.9 million workers, about a third of the workforce in the most exposed occupational categories, are highly exposed by current measures.

The more accurate dividing line runs through task composition, not collar color. A skilled trades supervisor who spends most of the week on scheduling, compliance documentation, and standardized inspection checklists may carry meaningful cognitive-routine exposure despite the job being physically hands-on. A corporate attorney who spends most of the week negotiating novel contract terms may carry very little exposure despite sitting at a desk all day. For a fuller treatment of why this distinction matters for workforce planning, see our breakdown of AI exposure across white-collar and blue-collar roles.

Reading exposure as an input, not a verdict

None of the scoring described here is a prediction of who keeps a job and who doesn't. It's a structured way to see where a role's tasks sit on the routine-to-judgment spectrum, so that HR and leadership can make an informed call about what to do next — not have the call made for them by a single number. In practice, that means treating exposure output as one of three signals: a role that lands in Monitor warrants periodic re-assessment as tools and task mix evolve; one in Review warrants a closer look at how the role's task list might be redesigned; one flagged as a Redeployment Candidate warrants an internal mobility conversation grounded in adjacent occupations and transferable skills — not an assumption about headcount. The distinction between augmentation and reshaping described throughout this companion piece is the same distinction: most roles are being changed in place, not deleted, and the task-level view is what tells you how.

Turning the framework into your own org chart

The framework above works on a hypothetical financial analyst. Applying it to your actual roles — with your actual O*NET mappings, your actual task weightings, and department-level rollups your leadership team can read at a glance — is a different exercise, and it's the one WorkforceAnalysis's pricing tiers are built around. If you'd rather start smaller, the AI Exposure Scorecard walks a single role through this same four-dimension method as a standalone deliverable.

If you want this kind of framework delivered before your next board cycle rather than after it, subscribe to our newsletter — we send one structured breakdown of AI workforce exposure research a month, no hype, sources attached.

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