What changed inside Anthropic
Anthropic has published an internal study of how AI is changing its own workforce. The researchers combine employment records, usage and output measures, employee surveys and 17 interviews. Their picture is of expanding activity alongside a changing mix of people and tasks.
The share of junior technical hires fell from 27% to 9%, while junior employees' share of technical roles fell from 21% to 14% since early 2025. That does not establish that total technical employment contracted: the report describes continuing technical hiring and faster growth in business functions. A smaller junior share and fewer jobs overall are different findings.
Agent hours need a careful reading
By August, the report estimates agent activity at six times human working hours. Its calculation counts inference and tool activity, excludes long waits, and assumes a 40-hour human week. It is a measure of activity, rather than evidence that each employee became six times as productive.
The researchers also compare engineers who adopted interactive agents with those who did not. They explicitly acknowledge that adoption was not random. Differences between those groups therefore cannot, by themselves, establish what AI caused. Code volume also leaves questions about the usefulness and quality of that output.
The bottleneck moves to judgment
The practical issue running through the report is how people supervise growing volumes of generated work. Employees describe building monitoring processes, reviewing output and deciding what deserves attention. Hiring places greater emphasis on autonomy and judgment as the work changes.
For a team considering similar tools, Franklin's reading is to track completed, reviewed work alongside generation volume. Record who checks the result and how much correction it requires. Anthropic offers a detailed case study from an unusually intensive AI user; it does not establish a universal forecast for junior employment elsewhere.