Escalation rate counts the share of conversations where the AI agent hands off to a human, by request, by guardrail trigger, or by its own decision that the case is out of scope. The headline number is meaningful only in context.
A high escalation rate is not automatically bad. An agent that escalates a payment dispute because it should not be touching that case is doing its job. An agent that escalates routine bookings because it cannot read the calendar is failing. The same number can mean either thing.
The useful breakdown is escalation reason, surfaced per call. Reason codes like out-of-scope-by-design, model-uncertainty, customer-requested, guardrail-tripped, integration-failure tell the buyer what is actually happening. A vendor that publishes the aggregate without the reason mix is hiding which kind they have.
Vorel surfaces escalation reason on every transferred call, with the conversation context the human picks up. The reason mix is part of the monthly business review.