Study notes · 2.5% of the exam

5.2 Design effective escalation and ambiguity resolution patterns

Escalate on explicit, observable triggers (customer asks for a human, policy gap or exception, no meaningful progress) and resolve ambiguity by asking for identifiers, instead of relying on sentiment, self-reported confidence or heuristic guesses.

Key points

1
The three core escalation triggers: the customer explicitly requests a human; the policy is ambiguous or silent on the customer's specific request (policy exceptions and gaps, not just "complex" cases); and the agent cannot make meaningful progress.
2
An explicit request for a human is honoured immediately, without first attempting investigation. Making the customer ask twice is the failure mode the rule exists to prevent.
3
Frustration without an explicit request is different: acknowledge the frustration and offer to resolve the issue when it is within the agent's capability, escalating only if the customer reiterates a preference for a human. Angry customers with simple issues are exactly the cases the agent should keep.
4
Policy silence is a trigger even when the case is easy. Example: the policy covers own-site price adjustments within 14 days; a competitor price-match request is not covered, so the agent escalates rather than extrapolating in either direction (granting or refusing).
5
No-progress should be an explicit, checkable criterion: the same tool error repeated, or no change in case state across a set number of turns. Escalate with a summary of what was attempted so the human does not start from zero. Raising turn limits or widening a tool's authority (for example the refund limit) removes controls instead of fixing the decision.
6
Sentiment-based escalation is unreliable because sentiment does not correlate with case complexity; self-reported confidence scores are poorly calibrated (the agent is confidently wrong on the hard cases). Both are common distractors.
7
The proportionate first fix for miscalibrated escalation (escalating simple cases, handling policy exceptions alone) is explicit escalation criteria in the system prompt with few-shot examples of escalate versus resolve, before building classifiers or ML infrastructure.
8
When get_customer returns multiple matches, ask for additional identifiers (order number, email, phone on file) and re-run the lookup before any order operation. Do not select by heuristic (most recent order, most plausible match) and do not escalate a case the customer can disambiguate in one message.
9
Never remove escalate_to_human from a category of sessions to cut escalations; that blocks legitimate triggers such as an explicit request or a policy gap discovered mid-conversation.

Read the source

Test yourself on 5.2 Design effective escalation and ambiguity resolution patterns

Ten questions, with the answer and explanation after each one.