An AI escalation policy is the set of rules that decides when an AI agent must hand a conversation to a human instead of resolving it itself. It defines the triggers, low confidence, sensitive intent, explicit customer request, policy boundary, that route a conversation out of automation and into a person's hands.
A good escalation policy is not a safety afterthought. It is a design decision about where the AI's authority ends. Without one, an over-eager agent pushes past its competence and a too-cautious one escalates everything, defeating the point. The right line is drawn per intent, not globally.
Aide, the agentic AI platform for customer experience, sets escalation at the intent level. Each intent gets its own threshold and conditions, and every routing decision carries a confidence score and a reviewable record. A billing dispute can escalate immediately while an order-status check resolves autonomously. Anything the AI has not been tested and verified on defaults to a person. That default does double duty: the hard, novel, judgment-heavy conversations keep landing with the team, so its exposure to real complexity stays intact rather than being skimmed away by automation. Escalation by intent, not by blanket rule.
Frequently asked questions
- What triggers an AI escalation to a human agent?
- Common triggers are low model confidence, a sensitive or high-risk intent such as a refund dispute, an explicit request for a person, or any intent the AI has not been verified to handle. The strongest policies define these per intent rather than as one global rule.
- How is an AI escalation policy different from a fallback message?
- A fallback message is a canned response when the AI cannot answer. An escalation policy actively routes the conversation, and its context, to the right human so the customer is helped, not just deflected with an apology.