aiqbench analysis
A Harvard study of 718 firms reports that AI coding agents raised lines of code by about 30%, but resolved Jira issues rose only about 12%, and coding assistants showed no significant gain in resolved issues. The gap is attributed to added work in later stages, especially code review, so raw code output is a poor proxy for delivered software. For AI readers, the practical question is whether review, testing, and integration capacity can keep pace with generated code.
Study Finds AI-generated Code Fails to Boost Productivity Due to Downstream Bottlenecks
A recent Harvard study of 718 firms, using the Jellyfish analytics platform, finds that although AI coding assistants and agents increase code volume, this does not translate into a corresponding increase in completed work, partly because AI-generated code increases the workload in later stages of development, particularly code review. By Sergio De Simone
A recent Harvard study of 718 firms, using the Jellyfish analytics platform, finds that although AI coding assistants and agents increase code volume, this does not translate into a corresponding increase in completed work, partly because AI-generated code increases the workload in later stages of development, particularly code review. The study was based on approximately 300 million work events of 700,000 workers across 700+ firms from January 2021 to…
Excerpt shown under the publisher's RSS terms · 69 words · full text and figures are on the publisher's site