Study Finds AI-generated Code Fails to Boost Productivity Due to Downstream Bottlenecks

AI coding tools may increase output without improving delivery speed, especially when review and integration capacity stay fixed. A Harvard study covering 718 firms found that coding assistants and agents generated more code, but the added volume did not produce a comparable increase in completed work. The research, based on engineering data from the Jellyfish analytics platform, points to downstream bottlenecks as the main constraint. AI-generated code increases the burden on later stages, particularly code review, shifting work rather than eliminating it. For engineering leaders, the practical metric is not lines of code or pull requests created, but cycle time, review load, rework, and completed features. Teams adopting coding agents may need to redesign review practices and invest in testing and validation automation before expecting organization-wide productivity gains.