AI is reshaping work at Anthropic: entry-level roles lose ground as the company grows

One of the companies developing artificial intelligence has offered a close-up view of how the technology is changing employment. Anthropic, the developer of Claude, published two reports on Friday, October 16, examining its workforce and employees’ expectations. The finding with particular significance for people trying to enter the industry is the declining share of entry-level technical hires: their proportion fell from 27% to 9% over the period studied.

The share of junior employees in technical departments also fell, from 21% to 14% since early 2025. Those figures describe a change in the composition of the workforce, rather than evidence that the company laid off two-thirds of its junior technical staff. Anthropic says it is still growing and that technical hiring continues to increase in absolute terms.

The internal study, which combines company records and interviews, links these changes to the adoption of agents capable of carrying out tasks. Its authors acknowledge an important limitation: employees who adopt these tools may differ from those who have not yet done so. An association between adoption and results cannot establish that AI caused the entire change.

More output means more work to review

The company presents its reports as a window into an organization that makes intensive use of its own technology. According to the overview published alongside the studies, increased output extends to activities including coding, sales and contract processing. The comparison seeks to look beyond the amount of code produced, a measure the company itself regards as imperfect.

The problem it describes will be familiar to any organization automating some of its operations: the more material systems produce, the more important it becomes to decide what needs attention. Anthropic says its employees are putting more effort into building ways to monitor the work of agents.

That points to a shift in responsibilities. Assigning a task and receiving a result does not remove the need to check it, integrate it with other decisions and take responsibility for the consequences. Software can generate options quickly; an organization still needs judgment to choose among them. An AI laboratory’s experience, however, does not replicate the conditions of a hospital, a factory or a small business.

What employees expect by 2030

The second report helps distinguish observed changes from expectations about the future. In a survey conducted in August, Anthropic employees estimated, on average, that AI would be able to perform 75% of their current work by 2030. Among the U.S. workers surveyed, the average was 28%.

The comparison draws on a sample of 10,980 U.S. adults, weighted to represent the population, and responses from the company’s staff. It does not turn employees’ opinions into an official unemployment forecast. Nor does it mean that 75% of their jobs will disappear: the question concerns tasks within their current work.

The survey also identifies obstacles that technical capability alone cannot resolve. Americans cited customers’ preference for dealing with people, privacy, reliability and legal liability. A tool’s ability to carry out an operation does not guarantee that a company will want to delegate it—or be allowed to do so.

Productivity remains difficult to measure

Independent research offers a useful counterpoint. METR, an organization that evaluates AI systems, explained in February 2026 why it was changing an experiment on developer productivity. Participants’ tasks were randomly assigned to conditions in which AI tools were either allowed or prohibited.

Wider use of these tools complicated the experiment. Some developers did not want to participate if they would have to work without them; others avoided submitting tasks they expected to complete much faster with AI assistance. That excluded some of the users and work that stood to benefit most.

METR concluded that its new data provided an unreliable indication of the size of the effect. It also found it difficult to measure time spent when one person coordinated several agents while working on other tasks. Its conclusion was that measurement needed to improve, rather than that there was a universal productivity gain applicable to every worker.

The distinction matters when interpreting company announcements. Counting completed tasks helps describe activity, but assessing productivity also requires considering time, quality and the kinds of work being compared. The speed at which a response is produced is not, by itself, a measure of its value.

The challenge of entering the workforce

The International Labour Organization had placed that distinction in a broader context in May 2025. Its index estimated that one in four workers worldwide was in an occupation with some degree of exposure to generative AI. Exposure means that certain tasks may be affected; it is not a forecast that a quarter of all jobs will disappear.

The ILO emphasized that most occupations include functions that still require human input. It therefore considered changes to jobs more likely than full automation. It also identified greater exposure in clerical work and certain highly digitized occupations.

Anthropic’s new findings add a specific question to that debate: how do people learn a profession if entry-level positions account for a smaller share of its jobs? A company can continue expanding its workforce while offering fewer opportunities, proportionally, to those just starting out. Recognizing that distinction will be essential to assessing AI’s effects on employment without assuming that business growth creates opportunities equally for everyone.