AI Coding Gains Neutralized by Human Review Bottleneck
AI-generated image · US National Wire
New research suggests that while AI agents increase raw code volume, the resulting surge in human auditing prevents actual software output from rising.
The promise of AI-driven productivity in software development may be hitting a ceiling of human capacity. According to reporting from Ars Technica, a study conducted by Harvard University researchers Fiona Chen and James Stratton reveals that efficiency gains from AI coding tools are being absorbed by downstream constraints in the production process.
Analyzing aggregated analytics from Jellyfish—covering over 700,000 employees at more than 700 firms between 2021 and March 2026—the researchers found that AI coding agents significantly boost raw output. Specifically, the introduction of these agents led to a 30% increase in total lines of code, a 20% rise in total commits, and a 23% increase in pull requests.
However, this volume does not translate to more finished software. The study found no statistically significant change in the resolution rate for Issues and Epics tracked via tools like Jira. Instead, the burden shifted to the review phase: the average time between a pull request submission and its merge increased by 49%. Furthermore, the number of comments per pull request rose by 35%, and the share of requests requiring changes nearly doubled.
To manage this influx, the share of workers performing code reviews increased by 14%. Despite 80% of firms using some form of AI code review by March 2026, humans remained responsible for the vast majority of the work; AI agents accounted for only 23.3% of review comments and 10.8% of pull requests. Consequently, the researchers found little evidence that firms reduced employment or increased overall software output through the use of these tools.