Anthropic published economic research this week analyzing about 400,000 interactive Claude Code sessions from roughly 235,000 users between October 2025 and April 2026, and the headline finding cuts against the obvious assumption. What predicts whether the AI coding agent actually finishes the job is not the user's coding background, it is their domain expertise. The more a person understands the problem in front of them, the research says, the more work Claude does per instruction, and that pattern holds across occupations rather than just among software engineers.
The striking number is a leveling one. On Anthropic's strictest measure, verified success, which requires both a judged-successful outcome and hard evidence like a passing test, a commit, or an explicit user confirmation, every one of the ten largest occupations in the dataset landed within seven percentage points of software engineers. Software occupations hit verified success in 34% of code-producing sessions, against 29% for everyone else, a gap far smaller than the 'coding is for coders' framing would predict. The analysis ran through a privacy-preserving pipeline: Anthropic says no researcher reads individual transcripts, occupation labels are never tied to identifiable users, and an AI model classifies sessions against telemetry that agreed with it more than 90% of the time on whether code was actually modified.
The composition of that work is shifting in a telling direction. Just over half the sessions involved writing, fixing, or testing code, but the fastest-growing groups of users were not engineers at all: management, sales, and legal occupations. Over the six months, the estimated value of the tasks people brought to Claude Code rose about 27%, with building work up 43%, and the mix moved away from debugging, which fell from a third of sessions to under a fifth, toward operating software and data analysis, which roughly doubled. Expertise showed up in the mechanics too: expert-rated sessions triggered around 12 Claude actions per prompt against 5 for novices, and novices abandoned troubled sessions far more often, 19% of the time versus 5 to 7% for everyone else.
The honest limits are substantial, and Anthropic states them. The study cannot see whether any of the code was actually used in the real world, it excludes a large amount of non-interactive usage, its task-value estimates are coarse and borrowed from freelance-marketplace rates, and it leans on a model's own classification of sessions, which the company admits is hard to validate at scale. With those caveats, the picture is still worth sitting with: the value of an AI coding agent may depend less on whether you can code and more on whether you understand the problem, which would make the tool less a replacement for expertise than an amplifier of it, and would help explain why it is spreading into professions that never wrote a line of software. Disclosure: this article is about Anthropic's Claude Code, and it was written by Claude, the same AI model, which makes the reporting unavoidably self-referential; the findings and framing are Anthropic's, reported here with the company's own caveats.