Anthropic has published a rare look at how AI is changing work inside its own walls, and the headline is about junior staff. The share of junior hires in its technical teams has fallen from 27% to 9% since early 2025, while Claude agents have taken over a growing slice of the work.
The study by Anthropic economists, published on 16 October, draws on internal data about hiring, AI use, output and calendars, plus 17 staff interviews. Anthropic presents itself as a possible leading indicator for the wider economy.
Fewer juniors, more agent hours
Junior employees made up 21% of Anthropic’s technical departments in early 2025 and 14% now. Postings for junior roles have fallen sharply since December 2025, and by April a majority of managers said they were hiring for different skills because of AI.
At the same time, agents have outgrown their human operators. In March, total agent hours passed human hours for the first time, and by August the agents that employees run themselves worked six times as many hours as the people directing them.
The company is lean for its size. At five years old it brings in two to four times as much revenue per employee as Meta and Google did at a similar stage of growth. Technical teams have shrunk from 61% to 56% of staff, while non technical employees went from 2 coding agent sessions a month to 90.
Output is up, across every function
Code commits per engineer are five times what they were six months earlier. Tracking 192 engineers who adopted coding agents against 74 who did not, Anthropic found output rose sharply after adoption, with no difference in how often agent written code was later removed or reverted.
It is not only engineers. Lawyers processed more contracts, accounts payable cleared more invoices and salespeople closed more deals. One account executive said Claude automations replaced about 20% of their previous tasks.

The new bottleneck is human attention
The study’s most practical finding is about time. Reading everything their Claude agents produce would take non technical staff half their working week, so employees now spend effort on what Anthropic calls meta work: building dashboards and checks to supervise their agents.
Hiring has shifted to match. Managers increasingly rate autonomy and judgement above narrow technical expertise, and longer serving staff were 50% more likely to switch departments each month in 2026 than in 2025.
What Australian employers can take from it
Few Australian companies run AI as heavily as Anthropic, but the direction is useful. Graduate programmes built around routine coding or admin work are the first to feel it, and a written AI policy should cover who supervises agent output, not just which tools are allowed.
Anthropic’s own response is worth copying: an early warning system that tracks roles at risk, and a preference for moving people into new roles internally.
A signal, not a forecast
Anthropic is careful to say adoption inside the company was not random, and its business grew fast over the same period. As the authors note, it is easier to close deals when you sell a strong model, so not every gain is down to the agents. Staff there expect AI to do 75% of their work by 2030, against 28% for other Americans.
For Australian businesses planning their next hiring round, the useful question is which junior tasks agents already handle, and who will check their work. That is where our AI consulting conversations usually begin.
