Anthropic has started publishing internal case studies on how its own product, engineering, and security teams use Claude Code and its Slack-native successor, Claude Tag, in daily production work. The headline number — 65% of the product engineering team’s pull requests now originate from the internal agent — has been circulating since Claude Tag launched, but the new case studies go further, describing the workflows, review gates, and team structures built around that figure rather than just the figure itself.
A dogfooding story, not an industry baseline
The honest reading of this is narrower than the headline suggests. Anthropic’s own engineers wrote the model, understand its failure modes better than any outside team could, and get early access to capability improvements before anyone else. That’s about as favourable an environment for AI-assisted development as exists anywhere, which is exactly why the 65% figure keeps getting cited uncritically. A regulated business — a healthcare provider, a bank, a public sector body — operating with less internal AI expertise, stricter compliance requirements, and code that has to survive an audit rather than just a code review, should expect a materially different number, not the same one.
What’s actually useful here
What’s worth taking from the case studies isn’t the percentage — it’s the operating model underneath it: proactive “ambient” follow-up on stalled work, tasks assigned and tracked in the same channel the team already works in, and human review kept firmly in the loop rather than removed. That’s a more transferable pattern than “let the agent write two-thirds of your code” is on its own. Teams that have tried to import the headline number without the surrounding process tend to get the productivity claim without the quality controls that make it sustainable.
So what
If you’re evaluating how much of your own build to hand to AI-assisted development, the Anthropic case studies are a useful reference point for what a mature setup looks like — but they’re a ceiling, not a target, especially in regulated sectors where the code has to satisfy more than a fast reviewer. We build that operating model in, calibrated to what your industry actually requires: see our AI-assisted development work, or our approach to healthcare software where the compliance bar changes the calculation entirely.