Anthropic’s workplace study puts human supervision in the spotlight
Three studies offer different windows into AI-assisted work, and reasons to distinguish more activity from better outcomes.
3 min read
Anthropic’s new examination of its own workplace highlights a question that deserves attention alongside AI’s output: what happens to the people responsible for directing and checking it? The company’s findings offer a case study, rather than a forecast for every employer.
Anthony Ha covered the research for TechCrunch on October 17.
What the new study actually measures
The October 16 report combines internal employment and usage records, surveys, calendars and hiring information with 17 employee interviews. It reports that junior staff’s share of technical departments fell from 21% to 14% since early 2025. Median meeting time per employee rose 20% over a year.
These figures describe a rapidly changing company studying itself. A shift in a workforce’s composition is not, on its own, a count of jobs eliminated by AI. Neither a meeting increase nor faster production establishes whether people’s overall work became better.
Expertise is different from a job title
A separate Anthropic study published in June examined roughly 400,000 interactive Claude Code sessions, covering October 2025 through April 2026. Its model-based analysis associated greater task-specific expertise with more successful sessions. Expertise here meant signals in the conversation, such as precise instructions and useful corrections, rather than a person’s formal seniority.
That distinction matters when interpreting workplace change. Knowing what a result must satisfy is different from being able to produce its first version. A person can lack coding experience while understanding the problem a program needs to solve. Conversely, a senior title does not guarantee familiarity with every task.
The June researchers also acknowledge an important boundary: they could assess transcripts and signals such as tests or commits, but could not observe the resulting work’s full real-world value. Their findings concern those sessions and their measurement system, not a universal score for workers.
Earlier employee accounts add context
The company’s December 2025 research used an August survey of 132 engineers and researchers, alongside 53 qualitative interviews and usage data. It described employees broadening the tasks they attempted, while some worried about reduced practice in deeper skills. These are earlier employee accounts, not fresh results from October’s study.
That earlier report explicitly recognized the unusual position of staff inside an AI developer, with early access to tools. Reading it beside the newer research helps keep a useful distinction visible: an internal experience can suggest questions worth investigating without representing ordinary workplaces.
What to ask about your own workflow
Sirlero’s practical suggestion is to examine one recurring task. Record what a successful result means, how you verify it and who can resolve an ambiguous decision. Keep the time spent reviewing and correcting alongside the time spent drafting. If a faster first attempt creates a longer checking process, both belong in the assessment.
The aim is a clearer account of the whole task. A completed draft, a checked result and a useful outcome are different milestones; a sensible evaluation gives each its place.
