AI-generated code fails to boost productivity due to review bottlenecks

A Harvard study of 718 firms and 700,000 developers found AI assistants raise code volume by 12% but do not increase resolved issues. PR review time rose 49% and the share of PRs requiring changes doubled.

• Assistants add 12% lines of code, 9% commits and 5% pull requests
• AI agents add 30% lines of code and 12% more resolved issues
• PR review time rose 49%, comments per PR grew 53%
• Engineers spend only 30–40% of their time writing code