Build a support triage workflow with an escalation path

Support triage should prioritize and route cases, not hide uncertainty behind a confident response. Define which severity, confidence and policy states require a person.

Execute a routing fixture

The agent controls lab classifies three synthetic tickets.

from agent_cases import support_triage_case

result = support_triage_case()
assert result["human_ids"] == ["S2", "S3"]
assert result["high_severity_automated"] is False
assert result["low_confidence_automated"] is False

High severity and 0.41 confidence route to a human. Only a high-confidence low-severity fixture uses self-service. These are authored values, not a calibrated model.

Separate classification and action

Validate intake, detect urgent safety/security categories with deterministic rules where possible, then apply a versioned routing policy. A model score alone should not authorize refunds or close tickets.

The escalation record needs reason, evidence, priority, owner and service target. Preserve the original message and model proposal under appropriate access. Let reviewers correct labels and feed reviewed cases into evaluation.

Evaluate operational harm

Measure missed urgent cases, unnecessary escalations, route accuracy, delay and workload. Test language, ambiguity, abuse, missing account data and prompt injection. Add an outage fallback that queues work rather than losing it.

The Generative & Agentic AI course connects triage to typed state, abstention and human approval.

Exercise

Create 50 synthetic tickets with severity and route labels. Predeclare zero-tolerance categories, evaluate confusion counts and demonstrate reviewer correction plus replay.

Continue learning

This article is part of the Agent workflows and state sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Design a safe read-only research agent.
• Continue with Compare orchestration frameworks using the same acceptance tests.

Reference: NIST AI RMF playbook.