According to CoinMeta, a study conducted by Harvard University found that the code output of AI coding assistants increased by 30%, but the number of additional problems solved did not increase. Researchers Fiona Chen and James Stratton analyzed data from 718 companies and discovered that although AI assistants such as Claude Code and Devin increased the code output per employee by nearly 1,500 lines per month, there was no significant improvement in solving Jira problems or completing projects. The study also showed that the code review process became a bottleneck, with review times increasing by 49%. Although 80% of the companies used AI auxiliary review tools, only about 23% of the review comments were generated by AI, indicating that the burden of quality control still falls on humans.