Anthropic has published research warning that computer programmers, customer service workers and financial analysts face the highest displacement risk from AI, but the data tells a more complicated story than the headlines suggest.
Thepaper, authored by economists Maxim Massenkoff and Peter McCrory and published on 5 March 2026, introduces a new framework for measuring AI's real-world impact on the labour market, one that deliberately separates theoretical capability from what AI is actually doing in professional settings today.
The findings offer some reassurance alongside several sharp warnings, and they arrive at a moment when anxiety aboutAI-driven job losseshas moved well beyond the pages of academic journals and into boardrooms, parliament buildings and household conversations.
The starting point for the research is a problem that has long plagued attempts to forecast AI's economic impact: past predictions have repeatedly overstated disruption. Anthropic's economists point to a well-known 2009 study thatidentified roughly a quarter of US jobsas vulnerable to offshoring, a finding that, a decade later, had simply not materialised in employment data. The lesson they draw is that measuring what AIcouldtheoretically do is not the same as measuring what it is actually doing.
To bridge that gap, the authors created a metric they call 'observed exposure.' The measure pulls from three sources: theO*NET database, which catalogues the tasks associated with around 800 US occupations; real-world Claude usage data gathered through theAnthropic Economic Index; and task-level exposure scores from a2023 academic study by Eloundou et al., which assessed whether an AI model could theoretically double the speed at which a given task is completed.
Crucially, the new measure weights automated use, where AI is replacing a worker's output, more heavily than augmentative use, where a human is simply being assisted. A job where AI helps a worker do more is treated differently from one where AI has started doing the job outright.
Computer programmers sit at the top of the exposure ranking, with 75% of their tasks now covered by observed AI use. Customer service representatives follow, driven largely by the rapid rise of AI-powered chat interfaces deployed directly by companies via API. Data entry workers, whose core task of transcribing and processing documents has long been ripe for automation, are 67% covered. Financial analysts also rank among the most exposed, given how much of their work; data retrieval, modelling, summarisation, maps directly onto what large language models do best.
Despite those headline numbers, the research lands with a significant caveat built in. AI is far from reaching its theoretical capability: actual coverage remains a fraction of what is technically feasible. Even in the most exposed occupational category; Computer and Mathematical roles, Claude currently covers just 33% of all tasks, against a theoretical ceiling of 94%.
Office and administrative roles sit at a theoretical 90%, but observed coverage lags far behind. Legal constraints, software requirements, human verification steps and simple inertia are all slowing the pace of adoption, even in sectors where the technology could, in principle, take over tomorrow.
That gap matters enormously when it comes to unemployment. The paper's authors are explicit: theyfind no systematic increase in unemploymentfor workers in the most AI-exposed occupations since late 2022, when ChatGPT's release began transforming public awareness of the technology.
Source: International Business Times UK
