Skip to content
Cheat sheets

CCAR-F · Domain 5: Context Management & Reliability · 15% of the exam

Task 5.2: Design effective escalation and ambiguity resolution patterns

When an agent hands over to a human (an explicit request, a policy gap, no progress) and when it should resolve the case itself, and asking for more identifiers instead of guessing between matches.

Study it

  • Escalation triggers and ambiguity: when to hand over and when to ask

    Partly covered by: Guardrails

Sample questions

Try each one before opening the answer. Every option is explained, with the Anthropic documentation page that proves it.

Question 1 · choose 1

A customer opens a chat with "I need to change my delivery address. I want to talk to a real person, not a bot." Changing an address is a task your support agent handles well and quickly. What should the agent do?

  1. ALook up the account and order details first so that the human gets a complete case, then escalate
  2. BAsk why the customer prefers a person and explain how quickly the agent can change the address
  3. CChange the address first, since it takes seconds, and then offer a human if the customer is still unhappy
  4. DHand the conversation to a human agent right away, without first trying to resolve the request itself
Show the answer and why

  • ALook up the account and order details first so that the human gets a complete case, then escalate

    Incorrect

    Investigating before honoring the request delays the handoff the customer asked for at the very start.

  • BAsk why the customer prefers a person and explain how quickly the agent can change the address

    Incorrect

    Persuading the customer to stay with the agent overrides the preference they stated.

  • CChange the address first, since it takes seconds, and then offer a human if the customer is still unhappy

    Incorrect

    Acting first ignores what the customer asked for: they want a person, and a quick fix does not change that.

  • DHand the conversation to a human agent right away, without first trying to resolve the request itself

    Correct

    The customer chose a person at the start; Anthropic's own buying agent lets customers make exactly that choice and routes them to a rep.

When a customer explicitly asks for a human, hand over at once rather than investigating or resolving first. Anthropic built its own buying agent as opt-in: customers choose at the start whether they talk to the agent or a sales rep.

Anthropic documentation

Task 5.2Report a problem with this question

Question 2 · choose 1

A customer gives only the name "Maria Garcia" and asks to cancel a subscription. get_customer returns three active accounts with that name in different cities. The agent has no other identifier yet, and cancelling the wrong subscription would be costly. What should the agent do?

  1. AChoose the account whose city matches the location of the customer's IP address
  2. BEscalate the conversation to a human agent because more than one account matched
  3. CAsk the customer for another identifier, such as the account email or an order number, before acting
  4. DChoose the account with the most recent activity, since it is the most likely to be the person writing
Show the answer and why

  • AChoose the account whose city matches the location of the customer's IP address

    Incorrect

    Location is a weak signal (travel, VPNs, shared names in one city) and still a guess.

  • BEscalate the conversation to a human agent because more than one account matched

    Incorrect

    A routine clarification resolves this; escalating hands humans work the agent can finish.

  • CAsk the customer for another identifier, such as the account email or an order number, before acting

    Correct

    Multiple matches call for clarification; an additional identifier resolves the ambiguity with certainty.

  • DChoose the account with the most recent activity, since it is the most likely to be the person writing

    Incorrect

    A heuristic guess can cancel someone else's subscription, which is exactly the costly mistake to avoid.

When a lookup returns several matches, ask for an additional identifier rather than selecting one by heuristics. Guessing risks acting on the wrong account.

Anthropic documentation

Task 5.2Report a problem with this question

Question 3 · choose 1

Your support agent escalates about 12 percent of conversations. Leadership wants to know whether those are the right ones, including whether some conversations that needed a person were never escalated. Which metric does the customer support guide describe for this?

  1. AThe daily escalation rate compared with last month
  2. BCustomer satisfaction scores on escalated conversations
  3. CEscalation accuracy: correctly escalated conversations against those that should have been escalated
  4. DAverage handle time of escalated conversations
Show the answer and why

  • AThe daily escalation rate compared with last month

    Incorrect

    A rate shows volume, not whether the escalated conversations were the right ones or which were missed.

  • BCustomer satisfaction scores on escalated conversations

    Incorrect

    Satisfaction on escalated cases says nothing about conversations that should have been escalated and were not.

  • CEscalation accuracy: correctly escalated conversations against those that should have been escalated

    Correct

    The guide tracks escalations made against escalations needed, which counts the misses too.

  • DAverage handle time of escalated conversations

    Incorrect

    Handle time measures speed, not whether escalation decisions were correct.

The customer support guide's escalation efficiency measure tracks correctly escalated conversations against those that should have been escalated but were not, with a target of 95 percent or higher.

Anthropic documentation

Task 5.2Report a problem with this question

Question 4 · choose 1

Your get_customer tool returns the first account whose name matches. When several customers share a name, the agent never learns that other accounts matched, and it sometimes acts on the wrong one. What should change in the tool?

  1. AReturn every matching account with a few non-sensitive details that tell them apart, such as city and masked email
  2. BKeep returning the first match, adding a flag in the result that says other accounts also matched
  3. CReturn "customer not found" whenever more than one account matches
  4. DReturn every matching account with its full record, including payment details and order history
Show the answer and why

  • AReturn every matching account with a few non-sensitive details that tell them apart, such as city and masked email

    Correct

    With every match and the details that distinguish them in the result, the agent can see that the lookup was ambiguous and which accounts are in play.

  • BKeep returning the first match, adding a flag in the result that says other accounts also matched

    Incorrect

    The flag shows that the lookup was ambiguous, but the agent sees only one account and is still steered toward it.

  • CReturn "customer not found" whenever more than one account matches

    Incorrect

    That is false and sends the agent down the wrong path.

  • DReturn every matching account with its full record, including payment details and order history

    Incorrect

    Full records flood the context with data the agent does not need, some of it sensitive; tools should return only high-signal information.

An agent can only handle ambiguity it can see. Design lookup tools to return all matches with just enough meaningful, non-sensitive context to tell them apart, rather than silently picking one or dumping every field of every record.

Anthropic documentation

Task 5.2Report a problem with this question

Question 5 · choose 1

Midway through a task, your Agent SDK assistant reaches a decision only the user can make: which of two deployment targets to use. The application has a UI that can show choices. How should the agent ask?

  1. AWith the AskUserQuestion tool, answered in your app's UI
  2. BBy ending the run with a question in its final text
  3. CBy escalating the whole task to a human operator
  4. DBy picking the more common target and mentioning it in the final report
Show the answer and why

  • AWith the AskUserQuestion tool, answered in your app's UI

    Correct

    A clarifying question pauses the run until your application returns the user's answer.

  • BBy ending the run with a question in its final text

    Incorrect

    Ending the run loses the in-progress work's flow.

  • CBy escalating the whole task to a human operator

    Incorrect

    A simple choice from the user is not an escalation.

  • DBy picking the more common target and mentioning it in the final report

    Incorrect

    Guessing a deployment target can cause real damage.

When the agent needs the user to decide, have it ask with AskUserQuestion; the Agent SDK pauses until your application supplies the answer.

Anthropic documentation

Task 5.2Report a problem with this question

Question 6 · choose 1

A customer in your car insurance support chat asks for advice on a medication dosage. The question is far outside the product's scope, and the customer has not asked for a person. What should the agent do?

  1. AGive general information with a disclaimer to consult a doctor
  2. BDecline that topic and steer the conversation back to insurance
  3. CEscalate to a human insurance agent so that the customer is not turned away
  4. DEnd the chat politely, since the question is off-topic
Show the answer and why

  • AGive general information with a disclaimer to consult a doctor

    Incorrect

    A disclaimer does not bring the topic within scope; the agent is still giving medical advice it is not there to give.

  • BDecline that topic and steer the conversation back to insurance

    Correct

    The guide's example keeps the agent on topic by declining unrelated questions and redirecting to the product's scope.

  • CEscalate to a human insurance agent so that the customer is not turned away

    Incorrect

    A human insurance agent cannot answer a medical question either; this only adds a wait.

  • DEnd the chat politely, since the question is off-topic

    Incorrect

    Ending the chat leaves the customer without guidance or a way back to what the agent can help with.

For questions outside the product's scope, the agent declines the topic and redirects to what it can help with, as the customer support guide's guardrails describe; escalation is for cases a person can resolve.

Anthropic documentation

Task 5.2Report a problem with this question

Question 7 · choose 1

A customer asks your support agent to return three items from one order: one within the return window, one past it but defective (policy: defective items are refundable for a year), and one bought on sale (policy: final sale, store credit only). Each part takes some reading, but the policy covers all three. What should the agent do?

  1. AEscalate, since a three-part request with different rules is too complex to handle alone
  2. BHandle the first item and ask the customer to contact a human about the other two
  3. CRefund all three in full so that the customer leaves satisfied
  4. DResolve all three under the policy: refunds for the first two and store credit for the sale item
Show the answer and why

  • AEscalate, since a three-part request with different rules is too complex to handle alone

    Incorrect

    Complexity on its own is not a reason to hand over when the policy answers every part.

  • BHandle the first item and ask the customer to contact a human about the other two

    Incorrect

    This sends the customer away for parts the agent is able to resolve.

  • CRefund all three in full so that the customer leaves satisfied

    Incorrect

    The policy says the sale item gets store credit; overriding it is not the agent's call.

  • DResolve all three under the policy: refunds for the first two and store credit for the sale item

    Correct

    The case is complicated, but the policy answers every part, so the agent can resolve it at first contact.

Complexity alone is not a reason to escalate. When the policy covers every part of a complicated request, the agent resolves it at first contact.

Anthropic documentation

Task 5.2Report a problem with this question

Practise domain 5 →Practise all domains →