A proxy metric is an observable measure that correlates with value without proving that an intervention caused the outcome. For AI-assisted coding, it replaces unreliable developer self-reports with evidence such as elapsed delivery time, merged work, or review rework. section 9.4.2
Useful engineering proxies include:
- Cycle time: elapsed time from starting a task to merging its pull request; it captures implementation acceleration but also includes design and review. section 9.4.2
- Pull-request throughput: merged pull requests per developer per sprint; interpret increases alongside pull-request size because smaller changes can inflate the count. section 9.4.2
- First-attempt success rate: the percentage of AI-generated code accepted without modification; it indicates quality and retry cost. section 9.4.2
A metric can improve while value remains unclear
After adopting AI tools, Marcus’s team reduced medium-complexity cycle time by 25–35% and increased pull requests from 6.2 to 8.1 per developer per sprint. Because pull-request size also decreased, more merges did not necessarily mean more features. section 9.4.2







