Anthropic published a study on October 16 of how its own workforce changed as AI agents spread through the company. The share of junior-level employees in technical departments slipped from 21% to 14% since early 2025, and the share of junior technical hires fell from 27% to 9%. By August, interactive AI agent hours outnumbered employee hours six to one.
What the study measured
According to “How employment is changing at Anthropic”, by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory, the authors drew on internal data covering employment, AI use, output metrics, employee surveys, communication patterns, calendar data and hiring decision records, plus 17 employee interviews. Anthropic is an AI company with its workflows built around its own agents, so the paper positions it as an early indicator and also warns that many of its employment practices cannot be replicated across the economy.
The report says discussion of weaker demand for junior technical staff began in core technical functions in January 2026. By April, engineering and research teams had deemphasized hiring for lower-band technical positions, and about a third of machine learning managers and a fifth of software engineering managers said they were hiring less because of AI. Postings for junior roles have fallen especially sharply since December 2025. The study also reports that technical departments’ share of total employment fell from 61% to 56% over the year, while hiring in sales, finance and recruiting grew faster.
Output rose, and the study tested quality
The authors report that code commits per engineer are about five times their level of six months earlier, and that lawyers processed more contracts, accounts payable moved more invoices and salespeople closed more deals as coding agents spread. Because Anthropic’s business also grew quickly, they compared 192 engineers who adopted coding agents with 74 who did not. Output rose within a person after adoption. They also compared code written by a person with code written by that person’s own agent, in the same month and directory, and found no difference in the rate of lines later removed or reverted.
The study states its own limits. Tool adoption is not random, so an engineer about to start a large project may adopt agents for it, and more code does not necessarily mean better code or business impact. The report notes that Anthropic takes in two to four times as much revenue per employee as Meta and Google did at comparable points in their growth.
Human attention became the constraint
The most useful numbers for a CIO concern supervision. The report says about 80% of interactive Claude session time is spent waiting for human input, and that the output of non-technical employees’ interactive sessions would take over half a work week to read. In an April survey, 20% of respondents described their role shifting from doing work to managing Claudes. One engineer described work as facing “an army of Claudes, with every one asking for my attention.”
Employees responded by building oversight tooling. The report says “meta-work”, such as writing specs for multi-agent runs, looping one agent to review another and monitoring pipelines, rose from 10% of the average employee’s interactive sessions in fall 2025 to 22% in spring 2026. Written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later. Median meeting time rose 20%.
What it means for the buying committee
Three points follow for a committee budgeting agent deployments.
- Price the supervision. If a Claude agent session sits idle 80% of the time waiting on a person, the cost that scales is reviewer time, and the study’s own staff spend a growing share of their sessions on monitoring tools. Our look at Dataiku’s agent inventory product covers one vendor answer to knowing what is running.
- Track the entry-level pipeline. The Decoder, summarizing the study, cites researcher Nolan Lovett of the NATO Special Operations University, who describes a “tragedy of the cognitive commons” in which each employer’s cut to junior roles is cheap for it while the shortage of experienced specialists falls on everyone later. See The Decoder’s report. A committee can ask HR for its own junior share and hiring numbers before agents change them.
- Read the survey gap with care. Deloitte’s survey found 42% of respondents had tested agents and 15% had scaled them. Those survey bases differ from Anthropic’s, so the figures do not combine. The study adds that Anthropic’s employees expect AI to do 75% of their work by 2030, against 28% for other Americans, so treat its results as an upper-end case.
The report also names conditions that sit outside an AI vendor’s product: broad employee stock ownership, an early warning system that tracks jobs at risk, and internal redeployment. The authors say a way to share the gains of adoption is essential. Companies without equity to share should expect the same question from their own staff.
Actions for a CIO or chief AI officer
Ask each team running agents for its ratio of agent hours to reviewer hours and for the share of staff time spent on monitoring. Set a quality test like the study’s, comparing agent-written work with human-written work in the same system. Ask HR for junior hiring numbers by department now, so a later decline can be traced to a cause, and see Cognizant’s training of 30,000 people on Claude for one large retraining effort. Finally, set a rule for who reads agent output and how much of it, because in the study reading every agent output would take over half a work week.
Source: Anthropic Research
