What the Harvard study found about AI and productivity
A new study from Harvard University looks at how AI tools change software work. The researchers used data from 718 companies. They tracked about 300 million work events from 700,000 workers. The data covers the period from January 2021 to March 2026. The study used tools like GitHub Copilot, Cursor, and Claude Code. It also looked at HR records, version control systems, and issue trackers. The goal was to see if AI helps teams ship more software.
The study found that AI changes how much code engineers write. It does not always change how much work gets finished. The researchers measured three things: lines of code added, commits, and pull requests. A commit is a saved change to the codebase. A pull request is a request to merge code into the main branch. The study shows that AI assistants raise code volume a little. AI agents raise code volume a lot. However, the increase in finished work does not match the increase in code volume.
The study also looked at how engineers spend their time. Engineers spend 30 to 40 percent of their time writing code. The rest of the time goes to other tasks. These tasks include planning, reviewing, and testing. When AI makes coding faster, the other tasks can become bottlenecks. The study shows that the development process has many steps beyond just writing code. These steps are where the slowdown happens.
The specific numbers for code volume and completed work
The study used the Jellyfish analytics platform to collect data. Jellyfish connects data from GitHub and Jira. It helps companies see how developers work. The researchers looked at three measures of output. They counted lines of code added, commits, and pull requests. They also measured the number of Jira issues and epics resolved per worker per month. A Jira issue is a single unit of work. An epic is a larger project or feature.
AI assistants caused small increases in output. Lines of code went up by 12 percent. Commits went up by 9 percent. Pull requests went up by 5 percent. Only the effect on commits was statistically significant. This means the change is likely real and not just luck. AI agents caused larger increases. Lines of code went up by 30 percent. Commits went up by 20 percent. Pull requests went up by 23 percent. All of these increases were significant.