Harvard researchers Fiona Chen and James Stratton analyze Jellyfish telemetry (300M+ work events; 700+ firms; 700k+ employees; 2021–Mar 2026) with difference-in-differences designs.
AI coding agents (not mere assistants): +30% lines of code, +20% commits, +23% pull requests; no significant change in Jira issue/epic resolution (software output proxy).
Review bottleneck: PR submit-to-merge time +49%; share of PRs with changes requested ~doubled; comments per PR +35%; share of workers doing reviews +14%. AI code review tools present at 80% of firms by Mar 2026 but only 23.3% of review comments and 10.8% of PRs from AI — limited relief.
Authors note observational design; employment effects near zero through sample end.