Escalation policy
- Agents
- Standard term
An escalation policy specifies when an AI agent leaves its normal path and what happens next, such as asking for information, stopping, using a specialist check, or handing the case to an authorized person.
Each trigger needs a destination with enough information or authority to handle it. An uncertainty trigger may call a specialist check, while a spending limit may require authorization. The handoff also needs the evidence and work already completed, so the next step can continue rather than start over.
Where you’ll meet it
Your team's AI assistant answers customer emails. Most ask where an order is, and it answers from shipping records. This week a few emails ask for refunds on damaged items, and one mentions a lawyer.
What to do
Define which refund requests need a refund approver and which messages raise a legal complaint. Route legal complaints to the responsible team rather than relying on the word “lawyer” alone. Each handoff should carry the email, order record, and any reply already sent. Try one email for each route before switching it on.
Easy to mix up
An autonomy budget sets how far the agent may go on its own. An escalation policy says which conditions change its route, and where the case goes next, such as to a person.
Check yourself
Your team's coding agent sends every change it makes to the payment code to the one engineer who reviews that code. He's on leave for two weeks, and those changes sit untouched. What should the policy add?
Go deeperA policy can fail by sending too much or too little
Software incident management provides a useful parallel: alerts pass through tiers of responders with different expertise and authority. An agent's policy can use the same structure, with destinations matched to conditions in the work rather than an assumption that every handoff requires a manager.
Overbroad triggers can bury reviewers in routine cases, while narrow triggers can miss cases outside the agent's tested range. Review escalated cases and samples of cases that stayed on the normal path each week to find both kinds of error.
Sources
- Escalation Policy Basics, docs.pagerduty.com
- Google SRE: Monitoring Distributed Systems, sre.google