AI agent escalation rules
Explicit triggers that route an AI agent's work to a human, with a destination and a SLA per trigger.
An escalation rule is an explicit trigger that routes an AI agent’s work to a human, instead of letting the agent act autonomously. Escalation rules are the safety valve on agent autonomy — the way an agent says “I’m not the right one for this; bring in a human.”
A well-designed escalation rule has four parts:
-
The trigger. A specific, falsifiable condition. “If the customer message contains the word ‘attorney’.” “If the proposed refund exceeds $200.” “If the agent’s confidence on the policy lookup is below 70%.” Avoid vague conditions like “if the agent is unsure.”
-
The destination. Who or what handles the escalation. “Manager review.” “Senior support agent on shift.” “The CEO’s inbox” (rare but valid for some triggers).
-
The SLA. How quickly the destination must respond. “4 business hours.” “Same-day.” “Within 15 minutes during open hours.” Without a SLA, escalations sit and rot.
-
The fallback. What happens if the SLA is missed. “Auto-escalate to the next level.” “Send a holding reply to the customer.” “Flag for the next morning’s review.” Failure paths matter.
The most common mistake in production agent ops is having one or two crude escalation rules (“escalate if confidence is low”) and treating that as enough. A real escalation policy has eight to fifteen specific rules, each with all four parts, written down in the agent’s R.I.D.G.E. card. They’re reviewed weekly: which fired? Which should have fired but didn’t? Which fired too often and need tuning?
RidgeHQ ships every agent with an explicit escalation policy as part of the R.I.D.G.E. framework’s “E” letter. The policy is part of the contract; changes to it are logged and auditable.