AI can scaffold this state machine refactor, but a senior engineer has to own the hard calls.

Each service has idiosyncratic logic, inconsistent naming, and undocumented assumptions baked into conditionals. The refactor is a one-off forensic exercise, not a repeatable pattern — the agent must re-reason from scratch for each service.

Success criteria are deeply underspecified: which library to use, how to handle partial failures, what the canonical transition graph actually is, and whether existing behavior is correct or buggy are all open questions. An agent cannot know when it's done without human sign-off at each decision point.

If the agent is given full read access to the codebase, it can parse the existing logic. However, it lacks access to runtime behavior, historical incident data, product requirements docs, and the implicit knowledge of engineers who built the system — all of which are essential for a correct refactor.

A misrepresented state transition in order fulfillment can cause real financial damage: orders stuck in wrong states, refunds triggered incorrectly, or shipments blocked. Errors may not surface until production and could be difficult to trace back to the refactor.

Deciding which existing behaviors are bugs vs. intentional, choosing the right abstraction (Spring Statemachine vs. Temporal vs. custom), and validating that the new model preserves all business invariants requires senior engineering judgment that AI cannot reliably substitute.