A Harvard working paper by Fiona Chen and James Stratton examines AI coding-agent adoption across 718 firms and 725,938 workers from 2021 to 2026. After adoption, code added per worker rose about 30 percent, commits 20 percent, and pull requests 23 percent. Yet completed Jira issues did not measurably increase, and the researchers found no evidence that tasks simply became larger. Instead, the review queue grew: pull requests took 49 percent longer to merge, formal change requests rose by about 12 percentage points, and review comments increased 35 percent. AI review tools did little to relieve the pressure. The authors suggest that more submissions burden a limited pool of reviewers, while increased requests and comments may also indicate lower-quality drafts. They found no clear short-term employment decline, though the estimates have limits.
The findings offer a measurement framework for legal AI tools that also draft or prepare work for human approval. Harvey’s motion-drafting workflow and Cooley’s developing redaction agent both retain lawyer review, but neither has published figures for reviewer time, revisions, or approved output. The study therefore supports tracking those measures alongside drafting speed, while recognizing that software work is not legal work and the paper has not undergone indicated peer review. Legal review also carries duties that cannot be delegated: attorneys remain responsible for reasonable inquiry before filing. Finally, the researchers identified AI use through system traces such as bot accounts and authorship signatures, suggesting that collaboration metadata may matter for retention schedules and legal holds.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.