Artificial Intelligence in the Firm: Bottlenecks in Software Production
Artificial Intelligence in the Firm: Bottlenecks in Software Production
Chen and Stratton study anonymized development events from 718 firms, comparing earlier adopters of coding agents with later and nonadopters. The data run from January 2021 to March 2026; this version of the paper is dated August 4, not October. They find roughly 30% more lines and 23% more pull requests per developer after adoption, but no statistically detectable increase in resolved Jira issues or epics. The upper bound of the estimate for resolved issues still allows a gain of about 12%, so the result is not proof of no benefit.
The pressure appears downstream. Pull requests spent 3.45 more calendar days between submission and merge, against a baseline of 7.03 days; formal requests for changes and review comments also rose. That interval is not hours a reviewer actively spent reading code. It can include waiting, iterations and coordination. The paper's model argues that writing faster does not finish features if verification cannot keep up. Its measurement does not show which proposed changes were agent-authored or why the queue lengthened. An October press account mistakenly described the share of pull requests receiving an AI-generated *review comment* as the share written by agents.
This is an observational firm-adoption comparison, not a randomized test of code quality. Adoption detection may miss private use, and Jira completion is not customer-confirmed behavior. There is no per-change model bill or human review-minute ledger. A smaller study of an AI coding *assistant* reports shorter peer review in a different setting, so 'AI always slows reviews' is not the conclusion. The practical next measurement follows one request through discarded drafts, review touch time, accepted behavior and later repair, rather than treating submitted pull requests as delivered value.