What is escalation trigger?

An escalation trigger is a condition that makes an AI agent stop answering and hand the conversation to a person. The common ones are a visitor asking for a human, retrieval finding no relevant content, a tool call failing, a visitor showing frustration, a topic the agent is told never to handle, and a plan's reply allowance running out. Each product has its own list and its own wording for the moment it fires.

Strictly, escalation is the decision and handoff is the transfer, though most vendors use the two words interchangeably, and Intercom and Zendesk use escalation for both. It matters because the triggers define the agent's boundary: everything that does not trigger escalation, the agent will try to answer. A product with only one trigger, the visitor asking for a person, answers everything else, including the questions it gets wrong.

Where definitions split: some products let you write triggers as rules (keywords, sentiment scores, page URLs), others let the model judge from the conversation, and most mix both. Rule triggers are predictable and crude: "refund" escalates every refund question, including the ones a policy page answers. Model-judged triggers catch "this isn't working and I'm done" but cannot be audited line by line. Where the idea runs out: a trigger is only as good as what happens next. Escalating to an inbox nobody watches is a slower fallback message.