AI and Bottlenecks in Software Production [pdf] (fion.ac)

A recent study from researchers at Harvard University, Fiona Chen and James Stratton, delves into the impacts of AI coding assistants and agents on software engineering productivity, utilizing a unique dataset from Jellyfish that encompasses 300 million work events across 718 firms. The findings reveal that while AI significantly boosts coding productivity—by up to 30% in various coding measures—the improvements do not correspondingly increase firm-level software output or employment. This discrepancy is largely attributed to a bottleneck in the code review process, where adoption of AI tools prolongs review times and increases the need for revisions. This research is significant for the AI/ML community as it highlights a paradox where greater efficiency in coding does not lead to proportional enhancements in overall software production, suggesting that AI's benefits may be hindered by traditional workflows. The study emphasizes the importance of recognizing complementary tasks in software development, such as code review, that can become bottlenecks. It suggests that firms may need to adapt their organizational structures and processes in response to AI adoption to fully realize its potential, prompting further exploration into managing labor dynamics and improving integration practices in software production environments.