AI Won’t Save Your Code Migration, But It Can Accelerate It

Large-scale code migrations are challenging due to their complexity, the loss of context over time, and the lack of a repeatable system, often leading to failures as they scale beyond manageable updates and security patches. Ankit Jain and Chris Westerhold, in a workshop, emphasized that fully automated AI migrations are unrealistic due to limitations like hallucinations and edge case blind spots, advocating instead for a human-in-the-loop model that combines automation with human feedback to ensure semantic correctness and adaptability. They introduced the concept of Aviator Runbooks, executable specifications that guide code migration by explicitly documenting transformation rules, assumptions, constraints, and edge cases, fostering multiplayer collaboration and reducing surprises during code generation. These Runbooks not only integrate with existing tools but also capture and version tribal knowledge, reducing future cognitive loads and creating a reusable foundation for consistent and efficient migrations. The approach shifts the perception of AI from a replacement to an accelerator, enhancing trust and system improvement through continuous human interaction and feedback.