Harvard study of 700+ firms finds AI code output rising without matching delivery

Harvard University conducted a study of 718 companies, drawing on data from the Jellyfish analytics platform covering January 2021 to March 2026. The dataset spans more than 700 companies, 700,000 employees, and roughly 300 million work events, combining sources including HR records, code version control systems, ticket management tools, calendar applications such as Google Calendar, and generative AI tools such as GitHub Copilot, Cursor, and Claude Code. The authors measured coding output along three dimensions: lines of code added, commits, and pull requests.

The researchers found that AI coding assistants lifted coding output only modestly: lines of code increased by 12%, commits by 9%, and pull requests by 5%, with only the rise in commits statistically significant. AI agents, by contrast, delivered large, statistically significant gains on all three metrics: lines of code up 30%, commits up 20%, and pull requests up 23%.

On the delivery side, the authors evaluated software delivery outcomes using the number of Jira tickets and large requests resolved per engineer per month. Coding assistants produced no increase in tickets resolved, while AI agents drove a 12% gain—well below the 30% rise in lines of code. To explain the gap, the team examined the full development workflow, which extends beyond coding to feature and scope planning, code review, testing, and deployment. Over the study period, the average review time for pull requests increased by 49% and the number of comments rose by 53%, pointing to later-stage review and integration as the bottleneck.