MedWrite's engineers accept only 36% of AI-generated code, finding strong programming skills essential even with 97.1% AI contribution to commits

At healthcare startup MedWrite, AI now accounts for up to 97.1% of committed code, with agents contributing 26.2K edits. However, the team rejects more AI code than it accepts, with accepted output lower than suggested output in their internal review metrics.

AI-generated code often looks reasonable in isolation but fails in real-world context. Example: AI suggested raising an exception if a monitoring server was unavailable, which would interrupt a doctor's patient session in an ambient listening tool—the wrong choice despite being technically correct code.

Enterprise software in regulated industries (healthcare, banking, defense) requires human review because bugs can become safety, security, or compliance issues. Hidden failures live in backend logic, APIs, permission rules, and integrations—places where coding skill remains critical. The team still requires challenging programming tests for new hires despite heavy AI adoption.

MedWrite uses Cursor for day-to-day coding, Cursor Bugbot for pull request reviews, Replit for prototyping, and Claude code for CLI automation.