The Jobs AI Is Most Likely to Transform Aren’t a Mystery Anymore. There’s Actual Data.

For a couple of years, the answer to “which Jobs AI Is Most Likely to Transform” was mostly speculation dressed up as analysis — theoretical capability scores, expert surveys, confident predictions built on how a model performs on a benchmark rather than how anyone actually uses it at work. That’s changed. Companies building these tools now have real usage data, drawn from millions of actual workplace conversations, mapped directly onto detailed occupational task lists. The picture that emerges isn’t vague. It names specific tasks, specific occupations, and a specific, measurable gap between what AI could theoretically do and what it’s actually doing right now.

What “Exposure” Actually Means, and Why the Distinction Matters

Before looking at the list, it’s worth understanding a distinction the underlying research is careful about, because most casual coverage flattens it. There’s a difference between a job’s theoretical exposure — the share of its tasks an AI model could plausibly perform based on its capabilities — and its observed exposure — the share of tasks actually showing up in real usage data. Anthropic’s own labor market research found that while AI could theoretically handle a large share of tasks in categories like office and administrative support, observed usage covers a much smaller portion even in the most AI-penetrated occupational category, computer and mathematical work, where real usage currently covers only about a third of that category’s total tasks. Capability and adoption are two different curves, and the gap between them is exactly where a lot of overheated predictions go wrong.

That gap matters for how seriously to take any specific claim about a job being “at risk.” A theoretical exposure score tells you what’s technically possible. An observed exposure score tells you what’s actually happening in real companies today, and it’s the far more reliable signal for understanding where transformation is genuinely underway versus where it remains mostly hypothetical.

The Occupations Showing the Highest Real-World AI Usage Right Now

With that distinction in mind, the actual data is specific rather than vague. Anthropic’s Economic Index, built by mapping millions of anonymized Claude conversations onto the U.S. Department of Labor’s O*NET occupational task database, found that computer programmers show the highest observed task coverage of any occupation, with usage leaning toward full task automation rather than simple assistance, followed closely by customer service representatives, whose core tasks are increasingly appearing in AI usage as well — with data entry specialists rounding out the occupations showing the heaviest real-world AI involvement. These aren’t the occupations pundits guessed at two years ago. They’re the ones the actual usage data names directly, ranked by how much of the job’s real task list is already showing up in AI conversations.

It’s worth naming what these occupations actually have in common, because the pattern is more useful than the specific job titles. Each one involves work that’s highly language-based, follows recognizable patterns across similar cases, and can be evaluated for correctness relatively quickly. That combination — linguistic, patterned, and fast to check against a clear right answer — is close to a working definition of what current AI systems are genuinely good at today, independent of any particular industry, seniority level, or job title on a business card.

Which Roles the Government Separately Projects Will Shrink or Grow

The usage data lines up closely with the government’s own independent forward-looking projections, which adds real weight to both. The Bureau of Labor Statistics’ 2024–34 employment projections found that AI-driven efficiency gains are expected to reduce demand for occupations including customer service representatives, legal secretaries and administrative assistants, and procurement clerks and other clerks — with employment in these roles projected to decline over the coming decade — while computer and mathematical occupations, including data scientists at a projected 33.5% growth rate, are expected to grow strongly as AI adoption increases. Two independent methodologies — one based on real usage patterns, one based on economic modeling and employer surveys — converge on largely the same occupations, which is a meaningfully stronger signal than either source would be alone.

Why Coding and Customer Service Are Actually Different Stories

It’s tempting to read “computer programmers and customer service reps both show high AI exposure” as the same story twice, but the underlying dynamic is genuinely different between the two, and understanding why matters for anyone in either field. Programming work is showing up in AI usage data leaning toward full automation of specific tasks — but the BLS projections still show computer and mathematical occupations growing overall, because the same technology that automates narrow coding tasks is also expanding what a smaller team of programmers can build, increasing total demand for people who can direct and evaluate that work even as the routine parts get automated.

Customer service tells a more straightforwardly contractionary story. The tasks involved — answering common questions, resolving standard requests, following defined scripts — are both highly exposed in the usage data and independently projected by the BLS to decline in raw employment terms, without the offsetting demand growth seen in programming. The difference isn’t really about which job is “more exposed” in some single number — it’s about whether the technology is expanding the total work available in that field or simply doing more of the same fixed amount of work with fewer people. Programming, so far, looks like the former. Customer service, so far, looks more like the latter.

What This Means If Your Job Is Somewhere on This List

The honest response to seeing your own occupation on a list like this isn’t panic, and it isn’t dismissal either — it’s a more specific diagnostic question: within your role, which tasks resemble the pattern these occupations share — language-based, repetitive across similar cases, fast to verify — and which tasks depend on context, relationships, or judgment that doesn’t reduce to a pattern? Our piece on how AI is changing knowledge work more broadly goes deeper into this same task-level lens, and the conclusion holds here too: the occupations shrinking fastest are shrinking because of their task composition, not because of their job title, which means the same job title can carry very different risk levels depending on which specific tasks actually fill someone’s week.

This is also where deliberately building the skills these lists don’t capture pays off. Our piece on tasks worth keeping firmly in human hands rather than delegating to AI and our piece on why connecting ideas across domains is becoming more valuable than narrow specialization both point toward the same underlying strategy: the parts of a role that involve ambiguity, relationships, and judgment are exactly the parts the exposure data shows AI still struggles to touch, regardless of how exposed the overall occupation appears on paper.

The Occupations That Barely Show Up at All

The flip side of this data is just as informative as the top of the list, and it gets far less attention. Anthropic’s research found that out of nearly 18,000 individual task statements tracked across the entire O*NET occupational database, only a small fraction show any measurable AI usage at all, and the occupations at the bottom of that list share their own clear pattern: physical-world work like equipment installation, maintenance, and repair, along with hands-on trades and in-person care work, show minimal presence in AI conversations regardless of how much AI capability has advanced generally. This isn’t because these jobs are somehow protected by policy or union agreements — it’s because the work itself depends on physical presence, tactile skill, and real-time physical judgment that current AI systems simply have no way to perform, language-based tools or not.

This bottom-of-the-list pattern is worth taking as seriously as the top, because it reframes what “AI transformation” actually means across the economy as a whole. It isn’t spreading evenly across every occupation, gradually touching more of each job over time. It’s concentrated heavily in specific occupational categories — knowledge work involving language, code, and pattern-based analysis — while remaining almost entirely absent from large swaths of the physical and interpersonal economy. Anyone trying to gauge their own risk benefits more from understanding this concentration than from a vague sense that “AI is coming for everything eventually,” which the actual usage data doesn’t currently support.

How to Actually Read Your Own Job Against This Data

A useful exercise, adapted from how the underlying research itself works: instead of asking “is my job at risk” as a single yes-or-no question, break your actual weekly tasks into the same categories the data uses. Which tasks are language-based, follow a recognizable pattern from case to case, and can be checked for correctness quickly? Those are the tasks resembling what’s already showing up heavily in AI usage data. Which tasks require you to weigh competing priorities, navigate a specific relationship, or make a judgment call that wouldn’t transfer cleanly to a different, similar-looking situation? Those are the tasks the data consistently shows remaining resistant, regardless of occupation. Most real jobs are a mix of both, and the ratio between them, not the job title on your business card, is what the data suggests actually predicts how much transformation is coming and how soon.

Frequently Asked Question