CCAR-F - Context Management & Reliability - Section 5.2

Design effective escalation and ambiguity resolution patterns.

The valid escalation triggers are an explicit customer request for a human, a policy exception or gap rather than mere complexity, and an inability to make meaningful progress. Sentiment analysis and self-reported confidence are unreliable proxies for case complexity. Candidates should honour an explicit request immediately without first investigating, acknowledge frustration while still offering resolution where the issue is within scope, and ask for another identifier when a lookup returns multiple matches.

escalation triggerspolicy gaps versus complexitysentiment-based escalation as an anti-patternself-reported confidence calibrationmultiple match clarification

Practice question for this objective

Free sampleContext Management & Reliabilitymedium

On the Customer Support Resolution Agent, a returns contact arrives in which the customer says only that the trainers arrived damaged and they want their money back. The lookup_order tool returns three delivered orders for that customer that each contain trainers, at 89.00, 89.00 and 134.00, placed in three consecutive months. The agent selects the most recent of the three and calls process_refund against it. A review of 200 contacts with more than one matching order finds the wrong order refunded in 27 of them, and every one of those calls passed schema validation cleanly. Which change most effectively addresses this?

  • AReturn every matching order from lookup_order together with the attributes that tell them apart, and require the agent to put a disambiguating question to the customer naming those attributes, with a deterministic check blocking process_refund while more than one candidate order remains unresolved. Correct
  • BTighten the process_refund input schema so that the amount must equal the order total exactly and the order identifier must match the order pattern, rejecting any call whose amount and order do not agree with one another.
  • CState in the project CLAUDE.md that when several orders match the customer's description the agent should always refund the most recent one, so that the behaviour is consistent and auditable across every session.
  • DHave the agent retry lookup_order with a narrower filter restricted to orders flagged as damaged on delivery, so that the candidate set shrinks to one before any refund decision is taken later in the same session.
When several records match a request, resolve the ambiguity with the customer before acting rather than picking a plausible candidate. A multiple-match result is an ambiguity signal, not a result to be narrowed by heuristic. The distinguishing fact lives with the customer, so no schema, no house rule and no repeated query can supply it. Surfacing the candidates and asking one targeted question converts an unresolvable guess into a resolvable exchange, and blocking the payment until then keeps a wrong answer from becoming a wrong transaction.

Why A is correct: Correct because the missing element is a fact only the customer holds, so the fix is to surface the competing candidates, ask for the one detail that separates them, and prevent a payment while the ambiguity is still open.

Why B is wrong: Tempting because stricter validation feels like the natural guard on a payment tool, but all 27 faulty calls were internally consistent and already passed validation: a schema constrains the shape of the arguments and cannot know which order the customer meant.

Why C is wrong: Tempting because it makes the behaviour predictable and reviewable, but it standardises a guess rather than resolving it, so the same 27 contacts would be refunded against the wrong order in exactly the same way, only now by written policy.

Why D is wrong: Tempting because a narrower query often does reduce a candidate set, but the damage was reported by the customer in conversation and is not recorded against the orders, so a repeated call cannot recover a fact the source never held.

See more CCAR-F practice questions, answers explained.

More in this domain

Back to all Context Management & Reliability objectives, or the CCAR-F cert hub.

Examworthy is not affiliated with or endorsed by Anthropic. Original, blueprint-aligned practice material only.