[1/2] How AI Is Transforming Work at Anthropic: Insights for Engineers (Category AI)

I found Anthropic’s internal study of how its AI assistant, Claude, affects the work of the company’s engineers particularly interesting. Of course, Anthropic is itself an AI company, so its employees have a privileged position: they are among the first to access the most advanced AI tools, and their work is directly connected to AI development. The authors therefore emphasize that the findings may not fully generalize to other organizations.

In this first post, I’ll try to summarize the insights for engineers. In the second, I’ll discuss the implications for managers.

Productivity has increased Developers now use Claude in around 60% of their work, compared with ~28% a year earlier, and estimate their productivity gain at ~50%. That is more than a twofold jump in a year. Measurements support the trend: for example, successful daily pull requests per engineer increased by 67% after Claude Code was introduced.

Claude helps with routine tasks Its most common uses are debugging and making sense of other people’s code; ~55% of engineers do this daily. Around 42% use AI to understand code and 37% to write new features. They ask for architectural design or data analysis less often, preferring to handle those tasks themselves.

Previously impractical tasks are now within reach 27% of the work engineers do with Claude would not have been done at all before. AI frees up time for things such as internal nice-to-have tools, including dashboards, and experiments that would be too costly to do manually. Claude also handles small improvements: ~8.6% of its tasks are minor bug fixes and refactoring that people had not got around to. Over time, these small changes add up to tangible gains in quality and speed.

Delegation, with caution Most estimate that only up to 20% of tasks can be entrusted to AI without checking. Claude has become a regular partner, but not an autonomous worker: developers still review and direct it, especially on important work. Engineers have developed an intuition for delegating tasks that are easy to verify, low-risk or boring, such as draft code and routine sections. They gradually trust it with more complex work, while retaining control over architecture and final design decisions.

Broader skills, uncertain depth With Claude, people are more willing to tackle tasks outside their main area of expertise. Everyone is becoming a little more full-stack. A backend developer, for example, can work on the frontend or a database with AI’s help instead of calling in specialists. There is a downside, though: when AI handles routine work, engineers get less practice with the fundamentals, and their basic knowledge may gradually atrophy.

Attitudes to writing code are changing Some are glad to focus on concepts and results rather than writing code. Experienced engineers compare the shift to the move toward higher-level languages. Many are willing to lose some of the enjoyment because productivity is now much higher.

Less direct communication Claude is increasingly the first place people take a question, ahead of their colleagues. This saves time—you do not bother a teammate with trivial questions—but mentoring suffers. Experienced developers note that juniors ask for advice less often because Claude can teach them plenty itself. Some dislike how commonplace “Have you asked Claude?” has become.

Careers and the future The engineer’s role is shifting toward managing AI systems rather than writing every line of code. Many already feel more like a team lead for a couple of AI agents than simply a developer. One, for example, estimates that 70% of their role has become reviewing or editing AI-generated code rather than writing it from scratch. Productivity is soaring, yet people are unsure what their profession will become over the long term. There is optimism about the near future, followed by a great deal of uncertainty.

Continued in the next post.

#Engineering #Software #Processes #Productivity #Economics