Everyone has a hot take on which jobs AI will kill first. Anthropic just published the receipts. They complicate the story more than they confirm it.
In a study released in early March, Anthropic’s economic research team introduced a metric called “observed exposure.” The idea is simple. Instead of asking what Claude could theoretically do to a job, measure what people already use it for inside real companies. That distinction sounds academic, but it isn’t. It separates a headline claiming AI will replace 90% of office work from the messier truth sitting underneath it.
Capability Outpaces Adoption
Anthropic’s researchers combined three inputs to build the index. The O*NET database catalogs roughly 800 U.S. occupations down to their component tasks. Claude’s own usage logs show what people actually do with it. A 2023 academic framework scores whether AI can cut a task’s completion time in half. Every occupation ends up with a coverage score, a rough measure of how much of the job AI genuinely handles rather than merely could.
The gap between those numbers tells the real story. Office and administrative roles look, on paper, like some of the most automatable jobs around — AI could theoretically handle up to 90% of their tasks. Observed usage says something different. Even computer and mathematical occupations, the single most AI-penetrated category, show real coverage of only about a third of their tasks. Capability sprinted ahead. Adoption still lags behind it.
That gap matters for anyone planning a career, not just anyone writing headlines about one.
Where the Pressure Actually Lands
Three occupations top Anthropic’s observed exposure list, and they’re not the ones a doomsday thread would guess.
Computer programmers lead, at roughly 75% task coverage. Claude isn’t just helping developers ship faster here. Usage patterns point toward full automation of a meaningful share of that work, not augmentation. It’s the same shift toward autonomous execution that shows up in how AI agents are absorbing junior-level coordination work elsewhere in the economy.
Customer service representatives rank second. Much of that exposure doesn’t look like people typing into a chatbot. It shows up in first-party API traffic — the unglamorous technical signal that companies route entire support workflows through AI systems instead of human queues.
Data-entry workers land third, at around 67% coverage. Nobody who’s watched a language model read a scanned document faster than a person can open the file finds that surprising.
No Mass Layoffs Yet, But a Quieter Kind of Damage
The research pushes back against the panic narrative here instead of feeding it. Anthropic’s team found no measurable spike in unemployment among high-exposure workers since ChatGPT’s public debut. Researchers even modeled a “Great Recession for white-collar workers” scenario, where unemployment in exposed fields doubles from 3% to 6%. Their framework says that would show up clearly by now. It hasn’t.
A second number buried in the report tells a subtler story. Among workers aged 22 to 25, the monthly rate of landing a job in high-exposure occupations has dropped by roughly 14% since generative AI went mainstream. The finding sits right at the edge of statistical significance, so treat it as suggestive rather than proven. It matches what entry-level hiring data across the wider market has been showing separately: young people aren’t losing AI-exposed jobs so much as struggling to get hired into them in the first place.
That’s a different kind of disruption than the “robots took my job” story everyone braced for. It’s quieter. It looks like a closed door rather than a pink slip, and the damage may not show up in hard numbers for years — the same slow erosion of leverage workers elsewhere are already describing as their expertise gets documented into the systems built to replace them.
Why the Framing Matters More Than the Numbers
Anthropic has been unusually candid, some would say alarmist, about the disruption its own technology could cause. CEO Dario Amodei has floated a scenario where AI wipes out half of entry-level white-collar work within a few years, a warning that sits close to what Anthropic’s own researchers describe as AI co-workers reshaping entry- and mid-level office roles. That framing made Anthropic a lightning rod. Rival AI executives have accused the company of overstating catastrophic risk to position itself as the industry’s designated conscience.
This particular study doesn’t lean on prediction, though. It measures what’s already happening, using Anthropic’s own product data as the dataset — a company grading its own technology’s real-world footprint in public, with the caveat that the picture stays incomplete. The researchers call this a first step, not a verdict. The data backs that humility. Entire job categories that look automatable on paper stay barely touched in practice, while a handful of specific roles absorb disproportionate pressure right now.
AI isn’t coming for every desk job at once. The jobs facing real disruption form a narrower, more specific list than the discourse suggests, and the earliest casualties may not be people losing jobs at all — just young workers who can’t get a foot in the door to begin with.
Related: They Think AI Could Kill Us. They’re Building It Anyway.
