A detailed report about how work is changing.
Summary:
This is an Anthropic Institute technical report by Wilmers, Catipon, Massenkoff and McCrory, published October 16, 2026. It asks how heavy AI use has changed work and hiring at Anthropic, which the authors see as an early indicator for the wider economy. The data are internal: headcount, agent telemetry, output metrics, surveys, Slack, calendars, hiring scorecards and 17 interviews.
Labor demand
- Fewer people per dollar of revenue. At about five years old, Anthropic earns 2–4× the revenue per employee that Google and Meta did at the same age, adjusted to 2025 dollars. The company tripled in size over the past year.
- Junior technical roles are shrinking. Since early 2025, juniors fell from 21% to 14% of technical headcount, from 27% to 9% of technical hires, and to 6% of open technical roles. The shift began in January 2026, when new models changed what teams were hiring for. By April, about a third of ML managers and a fifth of SWE managers said they hire less because of AI.
- Jobs where Claude handles the hardest tasks saw the weakest junior growth. This comparison across jobs (Fig. 5) is descriptive, not causal.
- Technical departments fell from 61% to 56% of headcount. Technical hiring still grew, but sales, finance and recruiting grew faster.
- Coding has spread beyond engineers. Non-technical staff went from 2 to 90 coding-agent sessions per month (Aug 2025 to May 2026).
- More people in technical departments now code. The share who aren’t coding nearly halved. There are about 17 engineers per PM, and commits from managers doubled.
- Administrative and clerical roles stayed small as the company grew. Agents, automated review and outsourcing absorbed that work.
- People are moving between teams more. Employees with a year or more of tenure were 50% more likely each month to move departments than in 2025. One content-writing team moved as a group into prompt and eval writing.
Productivity
- Agent hours passed human hours in March 2026. By August, agents in employee-directed sessions logged 6× human hours company-wide and 19× in research organizations.
- Output per person rose across functions (indexed to Dec 2025):
| Function | Change |
|---|---|
| Commits per engineer | 5.1× |
| Queries per data scientist | 3.1× |
| Contracts per legal employee | 3.1× |
| Deals per salesperson | 1.7× |
| Hires per recruiter | 1.4× |
| Invoices per AP processor | 1.4× |
- Engineers produced more after adopting agents. The study followed 192 engineers who adopted agents and 74 who didn’t, using two-way fixed effects. Code added rose sharply after adoption.
- Agent-written code held up as well as human code. Compared with the same person’s own code in the same month and directory, it showed no difference in lines removed, bug-fix deletions, reverts, lint suppressions, TODO/HACK notes or flaky-test markers.
- Not cleanly causal. Engineers who adopted fully autonomous agents were already on a steep upward trend before adopting.
- Sales reps also closed more clients after adopting agents. Selection into adoption may explain part of this. One account executive said automation replaced about 20% of their tasks; another said it let them talk to 30% more customers.
Bottlenecks and new kinds of work
- Expectations rise with each model. In a survey, the share agreeing their work processes make them as productive as possible went 44% → 58% → 75%, though the latest question was worded more weakly. Employees expect AI to do 75% of their work by 2030, versus 28% for other Americans.
- Each model release brings harder tasks. The share of very hard tasks rose about 30%, from 10% to 13%, after releases.
- Human attention is the constraint. About 80% of interactive session time is agents waiting on a human. Reading all agent output would take about 424 hours per person per week in engineering and research, 43 in finance, 35 in sales, 26 in legal and 12 in marketing.
- Meta-work has doubled. Building skills, writing specs, having agents review agents and monitoring pipelines grew from 10% to 22% of sessions, and from 29% in infrastructure. Standing agent instructions grew from a few hundred to 8,500 in a year.
- People meet more. Meeting time is up 20% for the median employee and 16% for technical individual contributors. Interviewees describe face-to-face coordination as indispensable.
- Agents add to Slack traffic rather than replacing people. They appear in nearly half of active public channels, mostly dedicated ones. Human-to-human communication didn’t change much.
- Agents need visible human work. Undocumented meetings and DMs leave agents working from stale context. As a result, the culture pushes toward deciding in the open.
- Hiring criteria are shifting. Scorecard mentions of AI-tool use tripled for technical roles and quadrupled for non-technical roles, from 5% and 7% in 2024. Autonomy and judgment gained weight, while domain expertise and execution skills lost weight.
What employers can do
The authors avoid claiming that Anthropic’s practices generalize, but point to three:
- Broad stock ownership, so workers share in the gains. They compare this to the 1950s union-management deal that tied pay to productivity.
- An early-warning system that tracks role-level demand and productivity against AI capability, to flag jobs at risk.
- Moving existing employees into new roles before hiring externally. The tradeoff is that this may slow outside hiring. The benefit is that it pushes the company toward AI that creates new tasks rather than “so-so” automation that cuts labor costs for little gain.
Their conclusion: none of these steps help the junior workers who were never hired. Employers acting on their own need to be paired with government policy on retraining and income support.