Study finds AI coding tools raise output in code, not in shipped software
The gains shift work into review and revision, leaving overall software delivery and staffing unchanged.
A Harvard-led study of software teams says AI coding agents raise code output but do not translate that into more shipped software or lower headcount, because review and revision work slows the rest of the pipeline. Using Jellyfish analytics from more than 700,000 employees at over 700 firms, the researchers found that gains in the coding stage were offset by longer code reviews, more requested changes, and more reviewer comments. The dataset covered 300 million work events, including commits and pull requests, across 2021 through March 2026. The result is a caution for companies betting that AI assistants will automatically increase engineering throughput without adding downstream process capacity.
Why it matters
For teams using AI assistants, the bottleneck moves downstream instead of disappearing. That means the measure that matters is not how much code is produced, but whether review and revision capacity can keep pace with it. For companies expecting faster delivery or smaller engineering teams, the study says those outcomes do not follow automatically.
Keep or strike?
Does this story matter, or is it hype? Mark it before you see what everyone else did.
Sources
- Ars Technica