What is Escalation Policy?
Also called escalation rules.
An escalation policy defines when an agent must stop acting on its own and route a decision to a person or a higher authority agent. It specifies the triggering conditions, who receives the request, what context accompanies it, and what the agent does while it waits, including whether the run pauses or continues on other work.
Escalation triggers fall into recognizable families. Low confidence in an answer, an action that exceeds an authority threshold such as a value or record count, an irreversible or externally visible effect, an ambiguous request with multiple plausible readings, and repeated failure on the same step all warrant handing the decision upward rather than guessing.
The quality of an escalation is judged by what it contains. A request stating only that the agent needs help forces the recipient to reconstruct the situation. A useful one states what the agent was doing, what it found, which options it considered, what it recommends, and precisely what decision is needed, so the reply can be a single choice.
Waiting behavior must be defined explicitly. An agent may block until answered, continue with unaffected work and return later, or time out into a safe default. Blocking is simple but fragile in unattended systems, since an unanswered escalation stalls the run indefinitely, so timeouts paired with a safe default are common in scheduled work.
Escalation rates are a health signal in both directions. Escalating too rarely means the agent is deciding things it should not, which surfaces later as errors nobody approved. Escalating too often makes review a bottleneck and trains recipients to approve without reading, which is functionally the same as no oversight at all.
Key points
- Defines triggers, recipient, context, and waiting behavior
- Common triggers: low confidence, thresholds, irreversibility, ambiguity
- Good escalations name the decision needed, not just the problem
- Unattended runs need timeouts with safe defaults
- Both too few and too many escalations are failure signals
In practice
A procurement agent may approve orders under a set threshold and must escalate anything above it. On an order that exceeds the threshold it sends the vendor, amount, matching quotes, budget line, and a recommendation, asking only for approve or reject. If nobody responds within four hours the request expires unapproved, the order is not placed, and the run records the timeout.