Escalation & Ambiguity Resolution
When Claude should escalate to a human, when it should ask for clarification, and when it should proceed. Designing explicit escalation criteria, handling ambiguous policies, and resolving multiple-match scenarios.
Quick Reference
- →Always honor explicit customer requests for a human agent -- immediately, without delay or persuasion
- →Escalation triggers must be defined in the system prompt with specific criteria, not left to Claude's judgment
- →Sentiment-based escalation is unreliable -- frustrated customers may have simple problems, calm customers may have complex ones
- →When multiple customers match a lookup, ask for additional identifiers -- never guess based on heuristics
- →Ambiguous policy situations (e.g., competitor price match when policy only covers own-site) should escalate, not improvise
- →Few-shot examples in the system prompt are the most reliable way to teach escalation boundaries
- →Acknowledge frustration first, then offer resolution -- only escalate if the customer reiterates after a genuine attempt
- →Confidence scores from Claude are not calibrated for case complexity -- do not use them as escalation thresholds
- →Log every escalation with the trigger reason for downstream quality analysis
- →Design your escalation path before building the happy path -- it's the safety net
Why Escalation Design Is Critical
An AI agent that never escalates will eventually cause harm -- it will make up policies, approve exceptions it shouldn't, or frustrate customers who need human judgment. An agent that escalates too often is just an expensive routing layer. The architect's job is to define the boundary precisely: which situations the agent handles autonomously, which require clarification, and which require immediate human involvement.
Scenario 1 (Customer Support) is the primary testing ground for escalation and ambiguity questions. Expect questions about when an agent should escalate vs. attempt resolution, and how to handle policy edge cases.
The exam distinguishes between three types of escalation: explicit (customer asks for a human), policy-based (situation exceeds the agent's authority), and complexity-based (agent cannot make progress). Each type has different design requirements and different failure modes.