Three legitimate escalation triggers:
- An explicit request for a human — stated directly, not inferred.
- A genuine policy exception or gap — the policy is actually ambiguous or silent on the case, not merely complex.
- Inability to make meaningful progress — after a real attempt, not preemptively.
Two proxies that feel plausible but are unreliable: sentiment-based escalation, and self-reported confidence scores. Neither correlates reliably with whether escalation is actually warranted.
The nuance the exam leans on hardest: an explicit demand for a human is honored immediately, without first attempting investigation — respecting the request outranks trying to resolve it anyway. But mere frustration on an otherwise straightforward issue is different: acknowledge it, offer the resolution path, and escalate only if the customer reiterates the demand. Candidates conflate these two cases constantly — treating frustration as if it were an explicit demand (over-escalating easy cases) or treating an explicit demand as something to investigate first (under-honoring it).
A third borderline case rounds out the set: genuine confusion. A customer writing "I don't understand what you're saying" is neither demanding a human nor expressing frustration — there's no negative sentiment and no explicit request. But if clarification and simpler restatement don't resolve the confusion after a real attempt, this becomes the third legitimate trigger from a different angle: inability to make meaningful progress, just diagnosed from the customer's side of the exchange rather than the agent's own. The handling is layered, same as frustration: attempt a clearer explanation first, and escalate only if the confusion persists despite that attempt — not on the first sign of confusion, and not never.
A related design point: when a lookup returns multiple ambiguous matches, the correct move is asking the customer for an additional identifier — not silently picking one match by heuristic.
The implementation fix for poor escalation calibration: explicit escalation criteria, paired with few-shot examples that show escalate-versus-resolve decisions on genuinely borderline cases. This mirrors Module 4's finding that explicit criteria beat vague instructions (§4.1) — the same principle, applied to a different decision.
The discriminator: escalate on policy and preference, not on feelings or self-assessed difficulty.
Escalation triggers at a glance:
| Scenario | Trigger? | Action |
|---|---|---|
| "I want to speak to a human." | Yes | Escalate immediately, no investigation first |
| "This is frustrating." (issue otherwise straightforward) | No, by itself | Acknowledge, offer resolution, escalate only if reiterated |
| "I don't understand what you're saying." | No, by itself | Attempt a clearer explanation, escalate only if confusion persists |
| Policy is silent or genuinely ambiguous on the case | Yes | Escalate |
| Agent has made a real attempt and cannot progress | Yes | Escalate |
| Low self-reported confidence score | No | Do not escalate on this alone |
| Negative sentiment, issue otherwise resolvable | No | Do not escalate on this alone |
Traps to recognize:
- A separately trained classifier for escalation — over-engineered before prompt-level criteria and few-shot examples have even been tried.
- A self-reported confidence threshold — the agent is already wrongly confident on exactly the hardest cases, which is the whole problem this is supposed to solve, not a fix for it.
- Sentiment analysis — solves a different problem entirely; sentiment doesn't correlate with case complexity or policy ambiguity.