Anthropic's AI coding assistant Claude Code is spreading rapidly beyond developers to non-technical jobs such as management, finance and law. As AI becomes more capable of writing code directly and performing software tasks, software development is moving beyond a specific profession and seeping into a range of workplaces.

Gigazine reported on Tuesday that Anthropic released research analysing about 400,000 Claude Code sessions left by about 235,000 users from October 2025 to April 2026. The study used an automated research tool called Cleo that can analyse real language-model usage patterns while protecting user personal data.

Anthropic assessed that agent-based coding is spreading quickly. The share of GitHub projects to which coding agents contributed more than doubled between late 2025 and early 2026. As of June 2026, Claude Code users were estimated to spend an average of 20 hours a week on the service.

Breakdowns by occupation also stood out. Researchers said they could estimate users' jobs in about 70 percent of all sessions. The biggest share, as expected, was computer and mathematics-related occupations. Business and finance followed, and users in arts, design and media, management, and life, physical and social sciences also accounted for a significant share. Among non-software occupations, the increase was most pronounced for people working in management, sales and law.

Uses are also changing. Of all sessions, code modification accounted for 26 percent and code writing for 25 percent. While more than half still involved traditional coding tasks, software operation made up 17 percent and document and presentation creation accounted for 10 percent, showing broader use. In particular, the share for code modification fell to 19 percent in April 2026 from 33 percent in October 2025, while shares related to software operation, data analysis and document writing rose steadily. Over the same period, the estimated economic value of an average session rose 27 percent.

Anthropic analysed that even as AI improves at generating code, the importance of expertise remains. Researchers divided user proficiency into five levels and compared them based on factors such as what users ask AI to check and how often they modify AI outputs. In beginner sessions, an average of five actions per prompt and output of about 600 words occurred. In proficient sessions, the number of actions was twice as high and output rose more than fivefold to about 3,200 words. It indicates that skilled users make more active use of AI agents' autonomy.

Differences in success rates were also clear. Under the strictest criteria, the success rate was about 15 percent in beginner sessions, but rose to 28 to 33 percent for intermediate-and-above user groups. The failure rate was 19 percent for beginners, compared with 5 to 7 percent for other user groups. Anthropic explained that less experienced users tend to struggle to get the results they want and give up easily.

Anthropic also assessed that the importance of having a coding background itself is lower than in the past. It said that in code-generation tasks, success rates for major occupational groups did not differ much from the software developer group. Anthropic analysed this by saying, "With the emergence of coding agents, a coding background appears to be becoming less important than before for programming success."

Anthropic forecast that if success rates for coding tasks among non-developer users rise further, software development may become part of the daily work of people across fields rather than the preserve of a specific profession. It explained that AI coding tools are evolving beyond a support means that raises developers' productivity into a general-purpose work platform that enables anyone to build software and automate tasks.