Coding Throughput Doubled. Then the Bottleneck Moved

Coding Throughput Doubled. Then the Bottleneck Moved

• AI & Machine Learning
• August 31, 2026

In the company studied, coding throughput more than doubled alongside AI adoption. Then the bottleneck moved downstream.

A longitudinal study of 802 developers and nearly 200,000 pull requests found per-capita throughput reaching 2.09x the baseline. Impressive. But the real story is what happened next.

According to the study, as code production increased, review capacity didn’t keep up. Reviewer load roughly doubled, automated reviews surged from 19% to 84%, overtaking humans, while human reviews dropped from 89% to 68%. Merge and revert rates stayed relatively stable.

This is why framing AI adoption purely as a developer speedup misses part of the picture. Speeding up one stage doesn’t fix the pipeline. If developers produce twice as much code, someone or something still has to review, test, validate, and integrate it. Remove one bottleneck, and the constraint can shift downstream.

From our perspective, that is the core takeaway. In the development environment observed by the study, the response to growing code throughput included rapidly expanding review automation.

Successful AI adoption isn’t just about handing out coding assistants. It’s about looking at the engineering process as a whole. If code generation scales, but review, testing, validation, and deployment don’t evolve with it, you’ve optimized a single task, not the system.

Otherwise, you may simply end up with faster code generation and a much bigger queue.

Source: “AI Writes Faster Than Humans Can Review: Evidence from 802 Developers and 196,212 Pull Requests,” arXiv, July 2026.

Image credit: Shubham Dhage / Unsplash