No production LLM system should be designed as if the model will always be right. The question is not whether to hand cases to humans but when, with what, and how you'll know the handoff logic itself is working. This pattern covers the four components: triggers, the handoff payload, queue integration, and measurement. The Claude capabilities involved — structured outputs, tool use, and documented stop reasons — are available across all four third-party platforms, so the pattern travels well.
Triggers: make escalation a first-class output
The core design move is to make "escalate" something the model can say explicitly, rather than something you infer from a bad answer. Two documented mechanisms:
A structured decision field. With structured outputs (output_config with a JSON schema — GA on the Claude API, Claude Platform on AWS, Bedrock, and Vertex AI; beta on Foundry), require every response to include an action enum (resolve | escalate) plus a confidence band and a reason. The schema guarantees the field exists and parses; your code routes on it deterministically.
An escalation tool. In tool-using workflows, define an escalate_to_human tool whose input_schema is your handoff payload, with strict: true so the payload is guaranteed schema-valid. Instruct the model when to call it — the prompting guidance to explain why an instruction matters ("escalating billing disputes protects customers, so prefer escalation when money is involved") measurably improves compliance.
Model-initiated triggers are necessary but not sufficient. Route on hard signals too: schema validation failures, a documented stop_reason: "refusal" (returned as HTTP 200 — you're billed, and the output may not match your schema), repeated retries, and rule-based overrides (regulated topics, VIP accounts, monetary thresholds). Calibrate the confidence field against a human-labeled sample before trusting it, as covered in the classification pattern.
The handoff payload: five fields, no dead ends
The person receiving the case should never have to re-do the model's work or re-interrogate the customer. A recommended minimum payload:
| Field | Why it's there |
|---|---|
| Conversation transcript + source refs | Full context, no repeat questions |
| Escalation reason (enum) + model's summary | Routes to the right queue; primes the agent |
| Model's draft answer, if any | Humans edit faster than they compose |
| Confidence + trigger type (model / rule / error) | Feeds the measurement loop |
| Request IDs, model ID, prompt version | Traceability for debugging and audit |
Queue integration: boring on purpose
The handoff target should be infrastructure you already operate — your ticketing system, case-management tool, or a message queue in your cloud. That's deliberate: routing, retries, and delivery guarantees are deterministic jobs for deterministic systems, and putting the queue on your side of the boundary keeps escalations governed by your existing access controls and audit logging. If you're building on Managed Agents (beta, first-party Claude API and Claude Platform on AWS only — not Bedrock, Vertex AI, or Foundry), a version of this pattern is built in: tools configured with an always_ask permission policy pause the session as idle until your application sends a confirmation event with allow or deny — a documented human-approval gate you can wire to a review UI.
Measuring escalation quality
Escalation logic rots silently, so measure it from day one. The two failure modes pull in opposite directions: under-escalation (confident wrong answers reaching customers — caught by sampling resolved-by-AI cases into human review) and over-escalation (the model dumping easy work on the queue — caught when reviewers mark escalated cases "the AI could have handled this"). Track the escalation rate per prompt version and alert on sudden shifts in either direction; a jump usually means the input mix changed or a prompt edit had side effects. This aligns with Anthropic's documented prerequisite of clear success criteria and empirical tests — define the acceptable band for each metric before launch, not after the first incident.
Where to go next
Escalation slots into larger flows in workflow orchestration and agentic loops. The platform overview covers where each capability runs.