Escalation and ambiguity resolution
What this covers
- Escalate on the right triggers rather than on difficulty
- Honor an explicit request for a human immediately
- Ask for identifiers instead of guessing between matches
- Acknowledge frustration and offer to resolve it when no human was requested
- Reject sentiment and self-reported confidence as escalation triggers
- Use few-shot examples of escalated and resolved contacts to fix the boundary
Key terms
- Few-shot examples
- A small set of examples included in a prompt to show the model the expected behavior.
When to escalate
- The customer asks for a human.
- Policy is ambiguous or does not address the request.
- The agent cannot make meaningful progress.
Note what is absent. Complexity is not on the list. An agent that escalates because a case is complicated escalates constantly and resolves nothing.
Explicit requests are honored immediately
If a customer asks for a human, escalate without first attempting investigation. Investigating first is well-intentioned and reads as ignoring what they asked for.
There's an important distinction here. If a frustrated user hasn't asked for a person, acknowledge the problem and offer to solve it when the agent can do so safely. Escalate if the user then asks for a person. Frustration deserves a response, but it doesn't always require a handoff.
Do not use sentiment or confidence alone
Sentiment-based escalation treats tone as difficulty. An angry customer with a simple problem gets escalated; a polite one with an intractable one doesn't.
Self-reported confidence is no better as a trigger. It doesn't track case complexity reliably, and using it means the agent decides how hard its own job is.
Use explicit criteria instead, with few-shot examples showing escalation versus autonomous resolution — the 4.2 technique applied to a judgment call.
Policy gaps
A clear reason to escalate is a request the policy doesn't cover. If the policy addresses price adjustments on your own site but says nothing about matching a competitor, the agent should escalate instead of inventing a rule.
Multiple matches mean ask
When a lookup returns several possible customers, request an additional identifier. Do not select by heuristic — most recent, closest name match. A wrong pick here means acting on the wrong person's account, and the customer can disambiguate in one question.
Reading all three records to work out which one is meant isn't a safer version of guessing. It opens two uninvolved people's account details in order to answer one person's question, and read-only access doesn't make that acceptable. Ask for the identifier.
escalation criteriafew-shot examplesclarification
Field note — common misconceptions
- MythThat frustration is a reason to escalate
- ActuallyFrustration without a request for a human calls for acknowledging it and offering to fix it.
- MythThat a confident agent should keep going
- ActuallyEscalate on an explicit request, a policy gap, or no progress; confidence is not a trigger.
- MythThat a complex case is by itself an escalation trigger
- ActuallyAn agent that escalates on complexity escalates constantly and resolves nothing.
Guided review
Review this lesson as a study deck
Review the lesson's main ideas in five guided slides, then test yourself with three flashcards.
Open Task 5.2 study deckCross-domain reasoning
Connect this idea
Tool access sets an authority boundary
Hooks, tool descriptions, access limits, MCP configuration, and escalation rules decide what an agent may do and when another person or system must decide.
Applied practice
Practice this lesson in a lab
Use a related lab to create a decision, implementation or diagram, evidence record, and review.
- Lab 3 · Deterministic prerequisite and human handoff
Enforce safety policy outside the prompt and produce a handoff another person can act on.
- Lab 10 · Reliability, provenance, and escalation evaluation
Evaluate autonomy, escalation, failure propagation, and claim-level provenance across a synthetic scenario suite.