Guardrails, escalation & auto-resolve

You decide how much the AI is allowed to do. Three controls matter most.

Confidence threshold

Every AI answer gets a confidence score. Below your threshold, the AI stops guessing and escalates to a human (or captures a ticket). Raise it to be conservative; lower it to let the AI handle more.

AI auto-reply (email & chat)

When enabled, the AI drafts or sends replies automatically above a minimum confidence (ai_auto_reply_min_confidence). Start in "draft" mode so agents approve replies, then switch to fully automatic once you trust it.

AI auto-resolve

For straightforward tickets, the AI can send the answer and close the ticket without an agent, above ai_auto_resolve_min_confidence. CSAT and reopens are tracked so you can audit quality.

Procedures (deterministic workflows)

For sensitive situations you don't want left to judgment, write a procedure: a fixed sequence of steps (e.g. add an internal note, then escalate to a human with high priority) that fires on matching intents/keywords. Procedures override free-form answering for those cases.

Always-escalate topics

Define topics (billing disputes, legal, security) that should always reach a person regardless of confidence.