I haven't shipped a dispatch tool with an AI builder myself, so I can't speak to that specific use case. The concern you raise about tangled logic seems worth testing directly rather than trusting generic chatter: partial shipments and rework on failed deliveries are exactly where state transitions multiply, and that's usually where generated code gets fragile. One practical approach is to model those edge cases as explicit states and rules before generating anything, then see whether the builder preserves them or flattens them. That should tell you early whether you'll be rewriting later.