Which Jobs Could AI Replace? Anthropic Research Shows White-Collar Work Faces the Greatest Exposure
History shows that major inventions often eliminate entire professions. The spread of electricity removed the need for lamplighters and elevator operators, while the rise of computers gradually replaced jobs such as switchboard operators, data entry clerks, and file clerks.
Anthropic analysis reveals AI could perform most tasks in finance, tech, and administration, raising warnings about hiring slowdowns and potential risks to white-collar workers.
‘Labor market impacts of AI: A new measure and early evidence’/Anthropic
Artificial intelligence may now be preparing to reshape another segment of the workforce.
A new report from Anthropic examines which types of work artificial intelligence is already performing and which tasks it could theoretically handle in the future. The findings suggest that the technology’s potential influence on professional jobs is far larger than its current use.
The company, created by former researchers from OpenAI and known for developing the Claude model, has frequently addressed both the opportunities and risks tied to AI progress.
Measuring AI’s Impact on Work
In the study “Labor market impacts of AI: A new measure and early evidence,” researchers Maxim Massenkoff and Peter McCrory analyzed professional interactions with Claude to understand how AI is actually being used.
Their conclusion: current adoption represents only a small share of the work AI systems could potentially perform.
Tasks across business operations, finance, management, mathematics, law, computing, and administrative roles are largely within the reach of AI technology in theory. However, real-world usage remains far lower than those technical capabilities.
Warnings about this possibility have already emerged from technology leaders. Dario Amodei has said artificial intelligence might disrupt about half of entry-level professional jobs. Mustafa Suleyman similarly predicted that a large portion of professional work could be replaced within 12 to 18 months.
Researchers attribute the slower rollout today to legal barriers, technical limits in models, the need for additional software, and the requirement that human workers review AI outputs.
Which Workers Face the Greatest Exposure
The study introduces a metric called “observed exposure,” comparing theoretical AI ability with the tasks the technology is actually performing in professional environments.
The results show that AI’s current involvement in workplace tasks remains limited compared with what it could eventually achieve.
If that gap narrows, the workers most likely to feel the impact are older professionals with higher incomes and advanced degrees. The study finds the most exposed group is more likely to be female by 16 percentage points, earns about 47% more than the least exposed workers, and is nearly four times as likely to hold graduate degrees.
This means the most vulnerable occupations include knowledge-based roles such as lawyers, financial analysts, and software developers rather than physically intensive jobs.
Among the most exposed positions identified in the research are computer programmers, customer service representatives, and data entry workers.
Capability Far Exceeds Current Automation
Even in the fields most vulnerable to AI automation, widespread replacement has not yet begun.
The researchers highlight a task frequently performed by doctors—approving prescription refills for pharmacies—as an example. Although AI models could theoretically automate this responsibility, the study did not find evidence of Claude performing it in real-world professional settings.
In technical occupations, the gap is particularly pronounced. Large language models could potentially perform 94% of tasks completed by workers in computer and mathematics fields, yet only about 33% of those tasks are currently handled by AI in observed professional use.
Administrative and office roles reveal a similar difference: around 90% of their responsibilities are theoretically automatable, but actual implementation remains limited.
In visual terms, the study compares the small amount of current AI usage to a much larger pool of potential capabilities. The researchers suggest that as models improve and adoption spreads, the gap may shrink.
At the same time, around 30% of workers have no measurable exposure to AI tools. Jobs requiring physical presence—such as cooks, bartenders, mechanics, and dishwashers—remain largely outside the reach of large language models.
The Risk of a White-Collar Recession
The researchers describe a possible outcome that could reshape the knowledge economy: a “Great Recession for white-collar workers.”
During the Great Recession, unemployment in the United States climbed from 5% to 10%. If unemployment among the most AI-exposed occupations doubled—from roughly 3% to 6%—the researchers say their model would detect it.
Such a shift has not yet occurred, but they note that it remains a realistic scenario.
Discussions about this possibility have spread beyond academic research. Michael S. Barr recently outlined several potential paths for AI adoption, including one where it significantly alters employment patterns.
Hiring Slowdown Rather Than Layoffs
Recent employment data highlights mixed signals. The U.S. Bureau of Labor Statistics reported that employers eliminated 92,000 jobs in February while the unemployment rate rose to 4.4%.
Some companies have linked workforce reductions to artificial intelligence. Block, Inc. recently cut nearly half of its employees. CEO Jack Dorsey wrote on X that new intelligence tools combined with smaller teams were changing how organizations operate.
Others remain skeptical. Marc Benioff has argued the layoffs may reflect internal challenges rather than AI alone.
The Anthropic research indicates the most visible effect so far may be slower hiring rather than large-scale job losses. In AI-exposed occupations, the job-finding rate has dropped about 14% compared with 2022 levels following the emergence of ChatGPT.
The authors note the statistical significance of this change is limited and emphasize that unemployment overall has not systematically increased.
Younger Workers Adjusting to Change
Other organizations see different patterns. Citadel Securities recently reported that hiring for software engineers has increased in recent months.
Nevertheless, the researchers say the modest slowdown could signal broader changes in how workers enter the job market during the AI era.
Additional research has shown a 16% decline in employment in AI-exposed roles among workers aged 22 to 25. For some young adults, this may mean postponing entry into the labor market.
According to the report, those who do not secure positions may remain in their current jobs, pursue alternative roles, or return to school.
