Escalation Policy
Written rules for when an AI agent stops trying and hands the case to a human — so the handoff is consistent, explained, and adjustable, instead of a mood.
Think of a new employee with a clear rulebook: handle these situations yourself, and go get the manager for those — any refund over a certain amount, any customer this upset. Without the rulebook, they either interrupt the manager constantly or push through things they shouldn't. An escalation policy is that rulebook, for an AI agent.
The problem, in plain words
An agent with no clear handoff rules fails in one of two directions. Either it escalates everything, and your review queue drowns — humans spend their day re-checking work the agent could have finished alone. Or it escalates too little, and shaky answers go out the door quietly — you hear about them from customers. Worse: without written rules, when it hands off shifts from case to case, and nobody on the team can say why this ticket reached a human and that one didn't.
What we set up
A written policy per workflow, enforced by the system that runs the agent (the runtime) — not left to the agent's in-the-moment judgment. The rulebook has four kinds of rules: a confidence floor (below this level of certainty, stop and hand off), risk classes (some actions — money, legal, account changes — always require a human, no matter how confident the agent feels), a retry budget (this many attempts, then stop trying and escalate), and deadline pressure (when a promised response time, an SLA, is about to run out, hand off early rather than late). Every escalation records which rule fired. And changing the rulebook is a versioned, reviewed change — never a quiet tweak.
How it works, step by step
- The rules are written per workflow
Confidence floor, risky categories, retry budget, deadline pressure — each workflow gets thresholds that fit its stakes.
- The system enforces them
The runtime applies the policy on every case. The agent cannot skip the rulebook on a good-feeling day.
- Confident and safe: the agent proceeds
Routine cases inside the lines get finished without bothering anyone.
- Unsure or risky: a human gets the case
The handoff arrives with context attached — what was tried, what the agent found — not a bare cry for help.
- Every handoff says why
The trace records which rule fired: low confidence, a risk class, retries spent, or a deadline closing in.
- The rulebook evolves from evidence
You can see which workflows escalate too much or too little, and adjust the thresholds with a versioned, reviewed change.
What changes for you
Before: escalation volume was noise — a pile of handoffs nobody could interpret, from an agent that was sometimes too timid and sometimes too bold. After: every handoff carries its reason, so the pile sorts itself into patterns you can act on, and tightening or loosening a threshold is a small reviewed change instead of a philosophical debate. What it won't do: the policy decides when a case reaches a human, not how good the human's answer will be. The hard cases still need your team's judgment — the policy just makes sure they arrive on time, with an explanation attached.