AI-generated PRs are quietly draining engineering review capacity

A Salesforce study co-authored by Shan Appajodu and Ravi Boyapati (January 2026) found that AI-assisted development raised internal code volume by roughly 30 percent, with pull requests frequently exceeding 20 files and 1,000 lines, pushing review latency higher each quarter. A separate 33-week panel study of GitHub developers found that AI coding agents significantly increase exploratory work, meaning more branches and prototypes are opened — and discarded — than under traditional human-authored workflows. Every pull request that enters a review queue consumes reviewer attention regardless of whether it ultimately merges, a cost that grows as agents inflate submission counts with short-lived, exploratory changes. Salesforce observed senior engineers context-switching across multiple large AI-assisted changesets daily, and linked declining per-PR review time to shallower scrutiny and erosion of its second-pair-of-eyes quality guarantee. Teams need to distinguish between two distinct problems: genuine reviewer overload, which requires more capacity or better triage, and rational disengagement from submissions that are unlikely to ever merge, which requires stricter upstream gating.