Some of the clearest published returns from AI in engineering come from unglamorous work: upgrades, migrations and incident investigation.
Why migrations fit AI well
A migration has a clear target ("move from Java 8 to 17", "replace this test library"), many similar changes, and a built-in check: does it compile, do the tests pass? That lets you wrap the model in a pipeline: generate the change, build, test, lint, and retry or escalate when it fails.
- Amazon used Amazon Q Developer's transformation agent to upgrade tens of thousands of production applications from Java 8 or 11 to Java 17, and reported savings of over 4,500 years of development work and $260 million in annual cost savings (AWS, August 2024; Amazon called the estimates conservative).
- Google described AI-assisted migrations in a 2025 experience report: in one migration, 80% of the code modifications in landed changes were fully AI-authored, and engineers estimated total time fell by about 50% compared with a similar migration without LLM help.
Legacy understanding
AI can summarize unfamiliar code, explain business rules buried in old modules and draft documentation. This speeds onboarding and modernization planning, but summaries need checking by someone who knows the system before they drive decisions.
Operations and incidents
AI can triage alerts, summarize incident channels and suggest likely root causes. Meta described a system that narrows thousands of code changes to a short list and uses an LLM to rank the top five candidates; it achieved 42% accuracy in identifying root causes at investigation creation time for its web monorepo. Meta stressed the risk of misleading engineers and relied on explainability and confidence thresholds.
The pattern
Bounded scope, automatic verification, a human owner, and metrics on time saved and defects introduced.