Anthropic Measures How AI Is Changing Work

Anthropic is no longer just studying how AI will transform employment. It is also measuring how AI is changing the company itself. A survey of 132 engineers and researchers, 53 interviews, and an analysis of 200,000 Claude Code conversations show a clear increase in productivity, along with doubts about skills, collaboration, and the future of software developers.

More work, not just less time

Anthropic employees say they now use Claude for 59% of their daily work, up from 28% a year ago. They also estimate that the tool delivers an average productivity increase of 50%, up from 20% previously.

The improvement is not simply about finishing the same tasks sooner. In many cases, workers are producing more. Based on their own estimates, 27% of the work assisted by Claude would not have been done without the tool: documentation, testing, prototypes, interactive dashboards, or small fixes that previously fell to the bottom of the list.

That last category has a revealing name: “papercuts,” the small annoyances of day-to-day work. The usage analysis found that 8.6% of Claude Code tasks involve these kinds of improvements, such as reorganizing code, creating shortcuts, or fixing minor issues that would not normally make the priority list.

Claude starts with easy tasks and moves toward more complex ones

Engineers mainly turn to Claude for debugging and understanding existing code. Fifty-five percent say they use it daily for debugging, while 42% use it to understand how unfamiliar code works.

Delegation usually follows a simple rule: people hand AI tasks that are easy to check, clearly defined, repetitive, or low-risk. Design decisions, strategy, and tasks that require organizational judgment generally remain in human hands.

But that boundary is moving. Between February and August 2025, the share of Claude Code usage devoted to implementing new features rose from 14.3% to 36.9%. Design and planning tasks increased from 1% to 9.9%.

Claude Code also went from carrying out around 9.8 consecutive actions without human intervention to 21.2. The average number of human interventions fell from 6.2 to 4.1 per conversation. The tasks analyzed also increased in average complexity, from 3.2 to 3.8 on a scale of 1 to 5.

That does not mean employees are handing their work entirely over to AI. More than half say they can fully delegate only between 0% and 20% of their tasks. Claude still needs supervision, review, and corrections, especially when the code is important or the margin for error is small.

Everyone does more, but some practice less

One of the most visible changes is that engineers are becoming more “multidisciplinary.” A backend specialist can build an interface; a researcher can create visualizations; a security team can analyze unfamiliar parts of a codebase.

AI reduces the time and initial knowledge needed to enter a new area. That makes it possible to test more ideas, build prototypes quickly, and take on projects that previously were not worth the effort.

The potential cost is less obvious. Some employees worry about losing practice in deep skills, such as writing code from scratch, understanding complex systems, or spotting subtle errors. If someone always lets Claude solve the basic problems, they may learn less in the process.

A paradox emerges: to supervise AI well, you need technical judgment, but that judgment can weaken if you stop practicing. Some workers are already setting aside tasks to do without Claude so they can stay prepared.

Fewer questions for coworkers

Claude is also changing the company’s social life. For many employees, it has become the first option for resolving questions they used to ask a coworker. A significant share of those questions has disappeared from team conversations.

There are advantages: interruptions to other workers decrease, and each person can explore ideas at their own pace. But opportunities for mentorship, informal learning, and collaboration are also lost. Some engineers say younger employees get faster answers from Claude but consult senior colleagues less often.

In practice, AI is filtering human work. Claude handles routine questions, while people reserve problems that require context, experience, or strategic decisions.

From writing code to directing agents

Several employees describe a shift in their role: they spend less time writing code and more time reviewing, guiding, and coordinating AI systems. Some already work with several instances of Claude at once and see themselves as responsible for the work of one, five, or even one hundred agents.

The transition brings short-term optimism and long-term concern. Engineers can do more, learn new tools, and take on ambitious projects, but they are not sure which skills will be valuable in a few years. Some even feel they are helping automate their own jobs.

Anthropic acknowledges that its findings do not represent the entire market. Its employees have early access to advanced models, work at an AI company, and belong to a relatively stable sector. In addition, the productivity figures are self-reported and may be inflated by memory bias or by the desire to show good results.

Even with those limitations, the study offers a useful signal for any company: AI does not just eliminate tasks. It also expands the kinds of work a person can attempt, changes who they collaborate with, and redefines what it means to have experience.

What to watch now is whether organizations create new forms of learning, mentorship, and professional development. If AI takes on more and more production work, human value will depend less on carrying out every step manually and more on knowing what to ask for, what to review, and when to say the result is not good enough.