How AI is transforming work at Anthropic

Studies of AI and work usually look across many occupations. In December 2025 Anthropic's Societal Impacts team turned the question on its own staff, on the grounds that people at an AI company are among the earliest heavy users and might show what other workplaces will meet later. The study combined a survey, interviews and usage logs from August 2025, when Claude Sonnet 4 and Claude Opus 4 were Anthropic's most capable models. The authors acknowledge that the setting is privileged: early access to tools, a relatively stable field, and staff who are themselves building the technology (introduction).

The post matters for this library as a mixed-methods study in which self-reports and logged behavior are reported side by side, with the gap between them discussed openly.

What the study did

The study had three parts:

• Survey. 132 engineers and researchers answered questions about how often they used Claude for eleven kinds of coding task, how much of their work involved Claude now and 12 months earlier, how it affected their productivity, and how much of their work they could "fully delegate" ("Survey data"). The questionnaire is published. Most answers were chosen from ranges, such as 0-20% to 80-100% for the share of work involving Claude, and bands from "I get less done than before" to "I get >100% more done in the same time" for productivity (survey questions 7, 8, 15 and 16).
• Interviews. 53 survey respondents took part in in-depth qualitative interviews ("Qualitative interviews").
• Usage data. Because respondents reported that most of their use was through Claude Code, Anthropic's command-line coding agent, the team used its privacy-preserving analysis tool, Clio, on 200,000 internal Claude Code transcripts from February and August 2025 ("Claude Code usage trends").

Recruitment combined posts in internal Slack channels (68 responses) with direct messages to 5 to 10 people in each of 20 teams chosen across research and product functions (207 contacted, a 31% response rate, 64 responses). The first 53 people to respond were interviewed (Appendix, "Limitations").

What it finds

Survey

• Most common uses. 55% used Claude daily for debugging, 42% for code understanding and 37% for implementing new features (Figure 1). Design and planning, data science and front-end work were less frequent.
• Usage and productivity. Respondents reported using Claude in 28% of their work 12 months earlier with a 20% productivity gain, and in 59% of their work now with a 50% gain ("Usage and productivity"; the key findings round these to 60% and describe a "2-3x increase"). The post says this roughly matches the 67% increase in merged pull requests per engineer per day that Anthropic saw when it adopted Claude Code across its engineering organization. 14% reported gains of more than 100%.
• Time and output. Asked about each task category, respondents reported, for almost all categories, a small net decrease in time spent and a larger net increase in output volume (Figure 2). The time responses split at both ends: some people spent much more time on Claude-assisted tasks, citing debugging and cleaning up Claude's code and the effort of understanding code they did not write; some also spent more time in an enabling sense, persisting on tasks they would earlier have abandoned or testing more thoroughly. The authors read the self-reports as suggesting that gains come mainly through greater output.
• New work. Respondents estimated that 27% of their Claude-assisted work would not have been done otherwise, such as scaling projects, nice-to-have tools and exploratory work. On average they said 44% of Claude-assisted work was work they would not have enjoyed doing.
• Delegation. More than half said they could "fully delegate" only 0-20% of their work. The authors note that respondents may interpret the phrase differently, from needing no verification to needing light oversight.

Usage data

Between February and August 2025 ("Tackling harder problems with less oversight"; Figure 3):

• Average task complexity, rated on a 1-to-5 scale from basic edits to expert work of weeks or months, rose from 3.2 to 3.8.
• The maximum number of consecutive tool calls per transcript rose by 116%, from 9.8 to 21.2.
• Human turns per transcript fell by 33%, from 6.2 to 4.1.

Implementing new features rose from 14.3% to 36.9% of transcripts, and code design or planning from 1.0% to 9.9% (Figure 4). In August, 8.6% of tasks were classified as "papercut fixes", small quality-of-life improvements. Teams differed (Figure 5): the Pre-training team used Claude Code for building new features in 54.6% of its transcripts, often extra experiments; Alignment & Safety and Post-training had the most front-end work (7.5% and 7.4%), often data visualizations; Security used it for code understanding in 48.9%; and non-technical staff used it for debugging in 51.5% and data science in 12.7%.

Interviews

The interviews describe several patterns:

• How people choose what to delegate. Interviewees tend to hand Claude tasks outside their context and low in complexity, easily verified (where they "can relatively easily sniff-check on correctness"), self-contained, low in quality stakes, repetitive or boring, and not faster to do by hand. Many described a progression of trust, and kept high-level design and "taste" decisions for themselves, while calling that boundary a moving target ("AI delegation approaches").
• Broader skills, less practice. Engineers reported working more "full-stack", such as backend engineers building interfaces. Some worried about skills atrophying and losing the incidental learning that comes from solving problems by hand. The post names a "paradox of supervision": using Claude well requires supervising it, and supervision requires the skills that heavy use may erode. One interviewee said skill loss would matter most for "my ability to safely use AI for the tasks that I care about" ("Skill transformations"). Some practice without AI deliberately; others argued the skills are no longer needed.
• The craft of coding. Views split between people who miss hands-on coding and people who care mainly about outcomes.
• Colleagues and mentorship. Claude has become the first stop for many questions. About half reported unchanged collaboration; others reported less contact with colleagues, and one senior engineer said junior staff came to them with questions less often, while adding that the junior staff got answers more effectively and learned faster ("Changing social dynamics in the workplace").
• Careers. Many described moving toward managing AI agents; one estimated their work had shifted "70%+ to being a code reviewer/reviser rather than a net-new code writer". Long-term uncertainty was widespread, with views ranging from optimism about the profession to concern about putting oneself out of a job ("Career uncertainty and adaptation").

Limitations and open questions

The authors list the limitations in the appendix: convenience and purposive sampling with likely selection bias toward people with strong views; responses that were not anonymous, from employees of the company that makes Claude, and so open to social desirability bias; memory distortion in recalling usage and productivity 12 months earlier (which the authors call recency bias); and the general difficulty of measuring productivity. The usage data are proportionate samples from each period, so they show changes in the mix of tasks, not in total volume: a rise in feature work from 14% to 37% of transcripts "does not necessarily indicate that more total feature work is being done". The models available in August 2025 have since been superseded.

The post also cites a METR study in which experienced developers overestimated the speed-up AI gave them on familiar codebases ("Usage and productivity"). The authors note that the conditions METR linked to slower work, such as large, complex codebases and tacit knowledge, match what their interviewees say they keep for themselves, and suggest that task selection may explain part of the difference. The post does not describe how average percentages were computed from the range-based survey answers.

Research questions

1. Do the self-reported gains survive anonymous collection? The authors name social desirability bias and suggest anonymous data collection. Repeating the survey anonymously with the same questions would show how much the 50% figure depends on attribution.
2. Where are reported time savings reinvested? The post says it is unclear whether saved time goes to more engineering, reviewing Claude's output or other activities. Time-use diaries or calendar data for a sample of respondents would answer it.
3. Does the paradox of supervision show up in measured review quality? Interviewees worry that skills needed to supervise Claude are atrophying. Measuring how often engineers catch seeded errors in AI-written code, across levels of Claude use, would test it.
4. Is the rise in feature work a change in capability or in adoption? The authors offer both explanations. Comparing task mixes for the same users across the two periods, rather than proportionate samples, would separate them.
5. Does task selection explain the difference from controlled studies? The authors suggest their staff choose tasks Claude handles well. A trial that assigns tasks at random versus lets engineers choose would measure the effect of selection.
6. Do the patterns hold outside engineering roles? The post says Anthropic is extending the research beyond engineers. Running the same survey with non-technical staff would show which findings are specific to software work.

How it connects

• Policy. The post points readers to Preparing for AI’s economic impact: exploring policy responses (October 2025) for possible policy responses.
• Interviewing at scale. Introducing Anthropic Interviewer, published two days later, says Anthropic previewed findings from Anthropic Interviewer on how AI is transforming software engineering at Anthropic in its work with teachers.
• Team study. Project Fetch: Can Claude train a robot dog? (November 2025) had said its finding that a team without Claude asked more questions "resonates with some of our upcoming findings from interviews with Anthropic staff".
• Economic Index. The post frames itself against Anthropic's economy-wide work, beginning with The Anthropic Economic Index (February 2025).

References

• Huang, S., Seethor, B., Durmus, E., Handa, K., McCain, M., Stern, M., & Ganguli, D. (2025, December 2). How AI is transforming work at Anthropic. Anthropic. https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic
• Anthropic. (2025). Claude at Work Survey (questionnaire). https://assets.anthropic.com/m/6cd21f7d4f82afcb/original/Claude-at-Work-Survey.pdf
• The METR study of experienced open-source developers (July 2025) is cited in the post and was not read for this note.