Helpin
Dark mode
Escalation rules

Escalation rules

Most escalations are judgment calls by the AI. Hard escalations are different: certain message categories bypass answering entirely and go straight to a human, regardless of confidence or settings.

Hard escalation categories

  1. Profanity or abuse — hostile or abusive language toward the assistant or team. The AI does not engage; it escalates.

  2. PII exposure — the customer sends sensitive personal data (payment card numbers, government IDs, credentials, health details). The conversation is escalated rather than processed through AI answering, minimizing sensitive text flowing through AI responses.

  3. Safety topics — threats of harm, self-harm indications, legal threats, or other categories you should never let an AI handle. Escalate immediately to a human.

The principle: when a message touches these categories, the right move is a human — fast, without the AI attempting to answer first.

What the customer sees

The customer sees the configured escalation message (default, or the busy/after-hours variant if those apply) and their conversation continues with your team. Hard escalation does not expose that the AI "refused" — the handoff message is the same calm, professional copy you set elsewhere.

How it works mechanically

  • The conversation is marked as escalated and waiting for a human, and lands in the handoff inbox per your routing settings.

  • The thread records the escalation as a system event, and the "Escalated by AI" badge appears.

  • Human takeover proceeds as with any escalation. Agents should handle the sensitive content with normal care — messages containing personal data are visible to agents per your workspace permissions.

Customer-requested humans

Related but distinct: when a visitor taps Talk to a person or explicitly asks for a human, the conversation escalates immediately, regardless of confidence, and the "Requested Human" badge appears on the row.

Guidance and policy

Hard escalation categories are product-level safety behavior. You influence tone and scope through guidance on your support agent configuration — for example, instructing the AI to hand off any account-security concern immediately:

"If the customer mentions unauthorized access, account compromise, or data loss, hand off to a human immediately."

Was this article helpful?