Who escalates?
A support agent hands the wrong cases to people: angry customers with easy problems go up, calm customers with hard ones stay. The tempting answers swap in a signal that seems objective: a mood score, the model's own confidence, a trained classifier. Each is an unreliable stand-in for what matters. The exam's answer is to say plainly when to escalate, and show a few examples.
D5 Context & ReliabilityD1 Agentic ArchitectureExam guide 5.2 and 1.4About 15 minutes
Start here: the new shop assistant
A shop hires a new assistant and says: “call the manager when you need to.” A week later:
You could give the assistant a mood meter, or ask “how sure do you feel?” before each sale, or build a system that learns from last year's calls. Each one measures something near the real question. None of them tells the assistant where the line is. And when they do call the manager, “someone needs help” is not enough: the manager needs who, what went wrong, how much, and what to do.
- The customer asks for a person: escalate at once, even if the agent could fix it.
- Policy exception or gap (over the limit, a case policy doesn't cover): escalate.
- The agent can't make progress: escalate.
- Upset but simple: resolve it. Mood is not a trigger.
- Unsure who the customer is (two accounts match): ask for an identifier. Don't guess, and don't escalate.
Words in the question that tell you which side you're on
Criteria + examples
- “escalates simple cases, attempts complex ones”
- “improve escalation calibration”
- “the prompt says escalate when needed”
- “asks for a real person” → escalate now
Unreliable proxies
- sentiment, frustration, tone of the message
- “confidence score”, “rate how sure it is”
- “classifier trained on past tickets” (overbuilt)
- “pick the most likely account” (heuristic)
Try it: replay a day of 100 tickets
Pick a ticket type to follow, then pick the escalation rule. The grid is one day of 100 tickets in arrival order; the type you picked is highlighted. The diagram above shows where your rule acts.
What the person receives
The rates are illustrative, chosen to show the shape. They are not measurements. The day's mix is 55 simple, 15 angry but simple, 12 policy exceptions, 8 calm but complex and 10 requests for a person, so the best possible first-contact resolution is 70%. The exam guide's point is that sentiment and self-reported confidence are unreliable proxies for complexity.
Things to try: Angry but simple + Sentiment score, then Calm but complex + Sentiment score: the same rule fails both ways. Then Policy exception + Self-rated confidence. Last, any type with Criteria + examples, and watch the day's resolution rate.
The three traps
| Trap | What the wrong option looks like | Why it loses |
|---|---|---|
| Unreliable proxy | Escalate on negative sentiment; escalate when Claude's self-rated confidence is low; pick the “most likely” account when two match | Mood is not complexity, and the model is often most confident on the cases it gets wrong. A heuristic pick sometimes acts on the wrong customer. |
| Overbuilt fix | A separate classifier trained on historical tickets | Needs labelled data and retraining, learns past habits, and misses new kinds of case. Criteria and examples in the prompt fix an unclear boundary for far less. |
| Losing the handoff | Resolve first when the customer asked for a person; escalate with “needs help” and a link, or with the whole transcript | An explicit request is a trigger on its own. The person needs a structured summary (customer ID, cause, amount, recommended action) to act without reading the chat. |
Practice: three exam-style questions
Answer first, then open “Think it through”.
Rule to remember
Exam guide 5.2: “Triggers: customer requests a human; policy exceptions/gaps; inability to progress.” “Escalate immediately on explicit demand; offer resolution when straightforward.” “Sentiment-based escalation and self-reported confidence are unreliable proxies for complexity.” Exam guide 1.4: “compile structured handoff summaries (customer ID, root cause, refund amount, recommended action) for humans without the transcript.”