Build an AI-powered support triage application

What you will build#

A support queue that classifies new requests and routes uncertain results to human review.

Part 1: Plan the application#

Define a SupportRequest with message, customer, status, category, priority, confidence, summary, and assignment.

Part 2: Create the intake workflow#

Build a request form and queue. Keep classification fields empty until the AI step succeeds.

At this point, a submitted request appears in the untriaged queue.

Part 3: Add structured classification#

Return category, priority, confidence, summary, and needs_review from a validated schema. Route low confidence and unsupported categories to review.

Part 4: Add human approval#

Let a reviewer correct the classification and approve assignment. Never send an external response automatically.

At this point, high-confidence requests route automatically and uncertain requests remain in review.

Part 5: Test and debug#

Use fixed examples for each category, ambiguous text, empty text, and malformed model output. Record model, token use, latency, and result status without logging sensitive messages.

Part 6: Deploy#

Deploy and submit a non-sensitive test request at the deployed URL.

Extend the application#

Add service-level targets, an MCP help-center search, or approved reply drafting.