Escalation Policy
The explicit rules that decide when an AI agent stops handling a task itself and hands off to a human (or a more capable agent) — and to whom, with what context, under what SLA. A good escalation policy is deterministic and testable: it names the trigger conditions (low confidence, a policy edge case, a repeated failure, an angry customer, a value threshold), the destination, and what the agent includes in the handoff. Escalation is what makes Human-on-the-Loop (HOTL) work in practice — the human only enters the loop on the exceptions the policy defines.
Written by Max Zeshut
Founder at Agentmelt · Last updated Sep 9, 2026
Example
A support agent's escalation policy: escalate to a human when (a) sentiment is negative for two consecutive turns, (b) the customer explicitly asks for a person, (c) the request involves a refund over $200, or (d) the agent's answer confidence is below 0.6. On escalation it passes the full transcript, the customer record, and its best-guess resolution so the human doesn't restart from zero.
Frequently asked questions
- What makes an escalation policy fail?
- Three common failures: vague triggers ("escalate if unsure" with no measurable threshold), a broken handoff (the human receives no context and has to re-interrogate the customer), and no fallback if the escalation target is unavailable. Write triggers as testable conditions, pass full context on handoff, and define what happens when no human is reachable.