AI can draft the refactor, but a DBA needs to own the business logic calls.

The mechanical parts — detecting nested transaction patterns, reformatting code, adding comment scaffolding — are structurally repeatable. But each procedure likely encodes unique business logic that requires fresh judgment to interpret correctly, making the task non-uniform across all 50+ procedures.

Success criteria are poorly defined: 'clarity' and 'business logic comments' are subjective, and correctness of the refactored logic cannot be verified without knowing the original intent. An agent cannot reliably know when it has gotten the business logic right versus plausibly wrong.

The agent can be given the raw SQL source files, which is sufficient for static analysis and pattern detection. However, it lacks access to runtime behavior, query plans, lock graphs, or the contractor's original intent — all of which are critical for diagnosing real deadlock causes.

Incorrectly refactored stored procedures in a production PostgreSQL database can cause data corruption, silent logic errors, or new deadlocks that are harder to trace. These errors may not surface immediately and can be difficult to reverse if the original procedures weren't versioned carefully.

Reconstructing business logic from undocumented code written by a third party is a high-judgment task — AI will frequently hallucinate plausible-sounding but incorrect intent. Deadlock resolution also requires understanding transaction semantics in the context of the full application, not just the procedure in isolation.