Escalation

Escalation is the step in an automated workflow where a case leaves the fast path and goes to a more capable, more expensive handler: a human agent, a specialist queue, an on-call engineer, or a slower System 2 model that can reason at length.

Escalation is triggered by two different kinds of signal, and good systems keep them apart:

• Content signals: the input itself calls for a human. A customer threatens legal action, an invoice exceeds an approval limit, a security alert looks like credential compromise. These are answers to questions you ask.
• Uncertainty signals: the model is not sure. The input is ambiguous, off-distribution or in an unexpected language. This is the abstention case.

Escalation with Laya

With Laya you express content signals as typed questions and uncertainty through the returned confidence. A support triage call might ask a noul question such as "Does the customer ask to speak to a person?" and a score question for frustration, then escalate if either crosses a threshold. The typed-decisions workflows include ids built for this: needs_human in customer service and needs_review in agent-trace observability.

A typical policy looks like:

1. Escalate if a content question fires (for example needs_human noul above 0.5).
2. Otherwise escalate if confidence on the primary question is below a threshold fitted on labelled data.
3. Otherwise act automatically.

Do not build step 2 on action.act_probability. The model card reports it reads about 1.0 for nearly every input and its raw logits run against correctness (AUROC 0.30 vs 0.77 for confidence). See act probability.

Because Laya answers a question in tens of milliseconds, you can afford to ask several escalation questions on every item and still leave most of your latency budget for the escalated path. The act/escalate routing guide and the support ticket triage use case walk through concrete thresholds.