# Escalation and ambiguity resolution

*Claude Certified Architect - Foundations Study Guide · Context Management & Reliability · Last updated October 22, 2026 · https://www.certifyforce.com/study-guide/claude-certified-architect-foundations/context-management-and-reliability/escalation-and-ambiguity-resolution*

When a support or developer agent should hand off to a human, offer to resolve, or ask a clarifying question: the three escalation triggers, why sentiment and self-reported confidence fail as signals, explicit criteria with few-shot examples, and AskUserQuestion for ambiguity in Agent SDK apps.

- Certification: Claude Certified Architect - Foundations
- Exam section: Context Management & Reliability
- Study guide: https://www.certifyforce.com/study-guide/claude-certified-architect-foundations

**Official exam objective:**

> Design effective escalation and ambiguity resolution patterns

## What the exam expects

- Recognize the three escalation triggers: an explicit request for a human, a policy exception or gap, and inability to make meaningful progress.
- Honor an explicit demand for a human immediately, but acknowledge frustration and offer to resolve when the customer has not asked for a person and the fix is within the agent's capability.
- Reject sentiment scores and self-reported confidence as escalation signals and replace them with explicit criteria plus few-shot examples in the system prompt.
- Escalate when policy is silent on the specific request instead of extending or inventing a rule.
- Ask for another identifier when a lookup returns several matches; in Agent SDK apps, let Claude ask clarifying questions with `AskUserQuestion` through `canUseTool`.

Key terms: **escalation trigger** (a categorical condition that requires handing off to a human); **policy gap** (a request the written policy neither allows nor forbids); **escalation accuracy** (share of conversations correctly escalated); **disambiguation** (asking for an identifier that picks one record); **AskUserQuestion** (the Agent SDK tool Claude calls to ask the user multiple-choice questions).

## Key facts

| Fact | Detail | Source |
|---|---|---|
| Escalation triggers | Human request, policy exception or gap, no meaningful progress | [Exam Guide v1.0](https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542750%2FClaude+Certified+Architect+%E2%80%93+Foundations+Exam+Guide.pdf) |
| Explicit human request | Escalate immediately, no investigation first | [Exam Guide v1.0](https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542750%2FClaude+Certified+Architect+%E2%80%93+Foundations+Exam+Guide.pdf) |
| Frustration, no request | Acknowledge, offer to resolve, escalate if the customer reiterates | [Exam Guide v1.0](https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542750%2FClaude+Certified+Architect+%E2%80%93+Foundations+Exam+Guide.pdf) |
| Escalation accuracy target | 95 percent or higher | [Customer support agent](https://platform.claude.com/docs/en/about-claude/use-case-guides/customer-support-chat) |
| Deflection target | Typically 70–80 percent handled without a human | [Customer support agent](https://platform.claude.com/docs/en/about-claude/use-case-guides/customer-support-chat) |
| Sample interactions in the prompt | At least 4–5 "good" examples, can include guardrails and tool calls | [Customer support agent](https://platform.claude.com/docs/en/about-claude/use-case-guides/customer-support-chat) |
| Emotion over intent | Known misclassification; direct Claude on when sentiment matters | [Ticket routing](https://platform.claude.com/docs/en/about-claude/use-case-guides/ticket-routing) |
| `AskUserQuestion` trigger | Fires `canUseTool` with `toolName` `AskUserQuestion` | [Handle approvals and user input](https://code.claude.com/docs/en/agent-sdk/user-input) |
| Restricted `tools` array | Must include `AskUserQuestion` or Claude cannot ask | [Handle approvals and user input](https://code.claude.com/docs/en/agent-sdk/user-input) |
| Options per question | 2-4 choices, each with `label` and `description` | [Handle approvals and user input](https://code.claude.com/docs/en/agent-sdk/user-input) |
| Auto-approved calls | `canUseTool` never fires for them (allow rules, `bypassPermissions`) | [Handle approvals and user input](https://code.claude.com/docs/en/agent-sdk/user-input) |

## Check yourself

- Why is a sentiment classifier a poor escalation trigger?
- When should a frustrated customer be escalated even though the agent can fix the issue?
- When does an agent ask the customer instead of escalating?
- What must an Agent SDK app with a restricted `tools` array do so Claude can ask clarifying questions?

## Sources

- [Claude Certified Architect – Foundations Exam Guide v1.0](https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F6nizmqk8tpzpfjvt6qmmav7rh%2Fpublic%2F1783542750%2FClaude+Certified+Architect+%E2%80%93+Foundations+Exam+Guide.pdf)
- [Customer support agent](https://platform.claude.com/docs/en/about-claude/use-case-guides/customer-support-chat)
- [Building effective AI agents](https://www.anthropic.com/research/building-effective-agents)
- [Ticket routing](https://platform.claude.com/docs/en/about-claude/use-case-guides/ticket-routing)
- [Prompting best practices](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices)
- [Handle approvals and user input](https://code.claude.com/docs/en/agent-sdk/user-input)

## About the full unit

The "Core concepts", "Exam traps" and "Worked scenario" sections of this unit, its 7 linked practice questions, and unlimited timed mock exams are included with the Claude Certified Architect - Foundations pack (USD $19): https://www.certifyforce.com/practice-exams