Let models classify; keep user-visible messages deterministic

Use a model to map inbound conversation into bounded intent and confidence, but keep eligibility gates, escalation, and user-visible reply text deterministic. This narrows the model’s authority: misclassification can still occur, so low confidence, stop-contact language, requests for a person, calls, attachments, and protected states must route to a human. Fixed templates also make receipts and regression assertions stable. This pattern fits safety-sensitive support with a small set of safe replies; it is too restrictive for open-ended cases that genuinely require personalized reasoning.

Applies when: Automating inbound support replies where unsafe or ambiguous outcomes can be escalated. Keywords: inbound support intent classification deterministic templates escalation safety receipts