AI coding tools increased software output in a study of 718 firms, but the gain shrank when researchers measured completed work rather than generated code. The analysis covered roughly 300 million work events from about 700,000 workers between 2021 and March 2026.
Coding assistants were associated with 12% more added lines, 9% more commits and 5% more pull requests, although only the commit increase was statistically significant. Agents produced larger changes: 30% more added lines, 20% more commits and 23% more pull requests. Yet resolved Jira issues increased only 12% with agents, and assistants did not increase resolved issues.
The researchers point to later development stages as a bottleneck. Average pull-request review time rose 49%, comments per pull request increased 53%, and the share requiring changes doubled. Fourteen percent more workers participated in reviews, while AI-assisted review did not show an improvement.
The Harvard study used operational data assembled through the Jellyfish analytics platform, including version control, issue trackers and workplace tools. Its results do not establish that every team will see the same effect, but they show why code volume alone is an incomplete measure of engineering productivity.