AI review can help with Laravel upgrades, but it should not make the decisions

Laravel upgrades are a prime area where AI may be overestimated. While large code diffs, framework alterations, and repetitive refactors are ripe for AI analysis, the critical upgrade decisions still require human oversight. The author finds AI review helpful during Laravel upgrades, primarily as a supplementary tool. It excels at catching updated methods, outdated configuration structures, stale imports, and inconsistencies at the framework level.

However, the most significant bugs in an upgrade often stem from judgment errors rather than syntax issues. These errors arise from a misunderstanding of how specific codebases utilize queues, middleware, authentication, caching, tenancy, validation, or exception handling. The author now recognizes that while AI assists in auditing the change set, the final judgment about migration should still be left to humans.

The most effective application of AI in Laravel upgrades is narrow and mechanical. By providing the tool with a diff, target Laravel version, and a specific question, AI can swiftly identify issues such as framework API drift, pattern mismatches, coverage gaps in upgrade guides, and low-level review fatigue. However, AI is less valuable when it attempts to infer business intent from a codebase it has only recently encountered.

AI review often produces plausible answers that sound senior enough to pass a quick read while still being incorrect in critical areas. AI typically misses understanding application-specific invariants, such as nuanced queue operations, tenant resolution, auth context, and request-scoped caching. It tends to normalize code toward framework defaults, even when those defaults are inappropriate for the application's domain.

To optimize the upgrade process, the author now structures the upgrade beforehand, creating an "upgrade notes file." This file outlines expected framework touchpoints, risky subsystems, and known application-specific deviations. By doing so, both human and AI review processes benefit from contextual awareness. The author emphasizes separating mechanical changes from behavior-preserving changes, conducting mechanical updates first and then addressing behavior-related modifications.

Architectural changes should only be considered if the upgrade necessitates them. AI excels at highlighting differences between the old and new codebases but struggles to determine which changes are worth addressing urgently. Consequently, the author recommends treating AI as a reviewer of prepared evidence rather than the entire upgrade process.

This approach allows humans to maintain control over critical judgment calls while leveraging AI's efficiency in identifying mechanical discrepancies.