Anthropic shipped a code review tool to enterprise customers last week. The announcement landed like most tool launches do: features, benchmarks, enterprise pricing. What nobody in the coverage noted is the question it sidesteps entirely.
When an AI system reviews code written by another AI system, and both are produced by the same company, the approval chain has no legal precedent. Nobody — not Anthropic, not the enterprise deploying it, not the engineer whose name is on the pull request — has established what it means when that approval fails.
Code Review went live March 9 for Teams and Enterprise customers. The product dispatches multiple agents to examine every pull request, verify bugs before surfacing them, and rank findings by severity. On large PRs over 1,000 lines, Anthropic's own data shows 84 percent get findings averaging 7.5 issues. Reviews run about 20 minutes and cost $15 to $25 per check. Less than 1 percent of findings are later marked incorrect by engineers, per the company's blog post. (Anthropic)
The numbers are worth putting in context. When Anthropic disclosed $2.5 billion in annualized run-rate revenue for Claude Code on March 9, it was a substantial figure on its own. But the company's total annualized revenue in February was $14 billion — and by April had reached $30 billion, up from $9 billion at year-end 2025. Claude Code's $2.5 billion run-rate in March thus represented a growing but still distinct minority of Anthropic's overall business. That distinction matters: the code review tool is being sold to enterprises as a standalone product, but it sits inside a company that has become a broad infrastructure play, not a niche tool vendor. (SaaStr)
The substantive PR review rate inside Anthropic went from 16 percent to 54 percent after deployment — a cultural shift, not just a tooling one. Code output per Anthropic engineer grew 200 percent in the last year. Subscriptions to the code review product have quadrupled since the start of the year, according to the company. Those are the inputs to the accountability problem. (TechCrunch)
The accountability problem lives in what comes next. Anthropic's product does not approve pull requests — that is still a human call. But the company is selling a system designed to catch what human reviewers miss. When a production bug reaches customers despite that review, the chain of responsibility is untested law. Did the engineer approve code they did not personally verify? Does the enterprise bear product liability? Does Anthropic, as the toolmaker?
The company's terms of service do not appear to include indemnification for code review findings. Professional liability insurance products have not been updated to account for AI co-signers. This is not unique to Anthropic — GitHub Copilot, Cursor, and other AI coding tools face the same gap — but Anthropic is the first to ship a multi-agent review system explicitly designed to be the authoritative second pair of eyes on AI-generated code.
The practical implication is straightforward. Engineering leaders deploying these tools are making judgment calls about acceptable risk without knowing what the liability framework actually is. The legal question will eventually be answered by a production incident, not a product blog post. When that happens, every company using AI-assisted review will need to know where they stand.
What Anthropic has built is impressive and probably useful. The question it created — who is responsible when the machine approves the machine's work — is one the industry has not answered yet.
