AI Escalation

How it works

Escalation is the most sensitive moment in an AI support interaction. A well-executed escalation is seamless: the customer feels heard, the agent is immediately context-aware, and the transition is fast. A poor escalation — routing to a generic queue with no context — signals to the customer that the AI wasted their time.

Escalation triggers fall into two categories. Automatic triggers fire without customer input: confidence falls below threshold, sentiment turns negative, a topic flag matches (e.g., 'refund fraud,' 'legal action'). Manual triggers are customer-initiated: typing 'human,' 'agent,' 'real person,' or clicking an escalation button.

Escalation rate (the inverse of deflection rate) is a useful secondary metric. A healthy escalation rate for a mature AI deployment is 20–40%. An escalation rate over 60% suggests the knowledge base is incomplete. An escalation rate under 10% warrants scrutiny — the confidence threshold may be too high, causing the AI to claim resolution on tickets it handled poorly.

When you'll encounter it

AI escalation appears in escalation workflow settings, agent training, and support analytics dashboards. You'll tune escalation triggers during the first weeks of deployment and review escalation clusters to identify knowledge gaps.