Jevons paradox describes a trap: efficiency lowers the cost of each unit, but expanded demand increases total resource consumption. In AI engineering, cheaper interactions invite experiments, automation, refactoring, and test generation that were previously uneconomical. section 9.1.2
Cheaper calls, larger bill
After optimization, a team increased AI interactions from 50 to more than 80 per developer per day. Monthly requests rose from roughly 13,000 to over 40,000 as developers, CI pipelines, batch refactoring, and automated test generation expanded usage. Lower token prices reduced unit cost but did not reduce aggregate spending. section 9.1.2
The same dynamic can distort engineering metrics:
- More API calls do not necessarily mean more productivity; measure value returned for the spending. section 9.1.2
- More pull requests do not necessarily mean more work completed; agents can split one human-sized change into several smaller changes. section 10.7
- Treat falling per-change effort as a reason to inspect outcomes, not as permission to optimize throughput alone. section 10.7
