I need to handle customer support messages that require distinct actions: routing, risk detection, and drafting a response. I decided not to force every step through the same LLM.
Customer support tasks required actions: Route the message to billing.; Recognise that cancellation risk is present.; Draft a useful response for a person to review.
Model comparison status: not_benchmarked
Prompt instruction enforceability: not_enforceable
Design process order: task → data/actions → approvals → monitoring/recovery → model