Anthropic published a study last week that maps the gap between what AI can theoretically do to white-collar work and what it is actually doing. The gap is enormous. For computer and math workers, large language models can handle 94 percent of tasks. Claude currently covers 33 percent in observed professional use. Office and administrative roles show a similar split: 90 percent theoretical capability, a fraction in practice.
The paper, titled “Labor market impacts of AI: A new measure and early evidence,” was written by Anthropic researchers Maxim Massenkoff and Peter McCrory. It introduces a metric called “observed exposure” — a comparison of theoretical AI capability against real-world usage data pulled directly from Claude interactions. The finding that jumps off the page: AI is barely scratching the surface of what it is technically capable of doing. And when it does close that gap, the workers most at risk are older, highly educated, and well paid.
This is not the story most people expect. The most AI-exposed group is 16 percentage points more likely to be female, earns 47 percent more on average, and is nearly four times as likely to hold a graduate degree compared to the least exposed group. That is the lawyer, the financial analyst, the software developer — not the warehouse worker. Computer programmers, customer service reps, and data entry keyers are the most exposed occupations.
The researchers give a telling example. A fully exposed task commonly performed by doctors: the authorization of drug refills to pharmacies. AI can certainly automate this task. The researchers note they have not yet observed Claude performing it, even though it can theoretically be completed by a large language model. This is the pattern everywhere. The “red area” depicting actual AI usage is dwarfed by the “blue area” of what is possible. As capabilities improve and adoption deepens, the researchers write, the red will grow to fill the blue.
At the other end, 30 percent of workers have zero AI exposure — cooks, mechanics, bartenders, dishwashers. Jobs requiring physical presence that no LLM can replicate.
The paper names the scenario everyone in the knowledge economy should be thinking about: a “Great Recession for white-collar workers.” During the 2007–2009 financial crisis, the U.S. unemployment rate doubled from 5 percent to 10 percent. The researchers note that a comparable doubling in the top quartile of AI-exposed occupations — from 3 percent to 6 percent — would be clearly detectable in their framework. It has not happened yet. But it absolutely could.
This is not just Anthropic talking its book. Federal Reserve Governor Michael S. Barr laid out the possibility among three scenarios he sees for AI adoption in a speech last month. Jack Dorsey’s Block cut nearly half its workforce, citing AI as a reason. “We’re already seeing that the intelligence tools we’re creating and using, paired with smaller and flatter teams, are enabling a new way of working which fundamentally changes what it means to build and run a company,” Dorsey wrote.
Critics including Salesforce CEO Marc Benioff have noted that Block has particular issues of its own and may be “AI washing” — using this as an excuse to conduct necessary layoffs. The distinction matters. If every layoff gets blamed on AI, the signal about real displacement gets lost in the noise.
The research finds that for young workers, the problem is not layoffs but a slowdown in hiring within AI-exposed fields. There is a 14 percent drop in the job finding rate in the post-ChatGPT era compared to 2022 in exposed occupations. However, the researchers note those findings are just barely statistically significant. There has so far been no systematic increase in unemployment. Citadel Securities, not known for publishing market research, was moved by a viral doomsday essay to note that hiring for software engineers has actually increased in recent months.
For some young workers, that means skirting the labor market entirely. “The young workers who are not hired may be remaining at their existing jobs, taking different jobs, or returning to school,” the researchers said.
The U.S. Bureau of Labor Statistics reported a dismal jobs report on the same Friday. Employers shed 92,000 jobs in February and the unemployment rate ticked up to 4.4 percent. Some of that is AI. Some of it is not. The Anthropic paper gives a framework for telling the difference.
Anthropic CEO Dario Amodei last year said the technology could disrupt half of entry-level white-collar work. Microsoft’s AI chief, Mustafa Suleyman, made a similar prediction, estimating most professional work will be replaced within a year to 18 months. The researchers attribute the lag between capability and adoption to existing legal constraints and technical hurdles — model limitations, the necessity of additional software tools, and the need for humans to still review AI’s work. That is just temporary, they project.
The most useful thing about this paper is that it replaces speculation with measurement. Instead of asking “will AI replace jobs,” it asks “how many tasks in each occupation can AI actually do today, and how many is it actually doing?” The answer is a gap large enough to drive a policy debate through. The “observed exposure” metric gives regulators, labor economists, and company leaders a way to track whether that gap is shrinking — and how fast.
For AI builders, the implication is uncomfortable. The technology is already capable of automating a vast share of white-collar work. The reason it has not yet done so is not a lack of capability. It is a lack of integration, a lack of trust, a lack of legal infrastructure, and a lack of incentive for companies to reorganize around the technology. Those barriers are all solvable. The question is not whether the red area will grow to fill the blue. The question is how fast, and who gets caught in the transition.
The paper does not answer that question. It gives the tools to watch it happen.