AI Labs
All examples
AWS
Live demo

Intent Routing on Amazon Bedrock

Classify user messages, then answer with intent-specific system prompts on Bedrock Converse.

Intent routingAgentsGovernance
Jump to demo

Try policy, technical, and escalation phrasing — watch the route label change first.

Technical notes

Two Converse calls: classify intent, then answer with route-specific prompt.

Intent routing on Bedrock

Classify the message, then answer with an intent-specific system prompt.

Live
Case studyArchitecture, governance, and how to adapt this pattern in a pilot

Business use case

Problem

One chat UI serves policy questions, technical how-tos, and escalation requests. A single generic system prompt mishandles at least one class.

Who benefits

  • Operations, predictable handoff behaviour for escalation intents
  • Enablement, different tone and guardrails per intent
  • AWS platform teams, routing without standing up a full Bedrock Agent on day one

Success metrics

  • Intent accuracy ≥ 85% on a labelled eval set before production
  • Escalation intent never claims to be a human supervisor
  • Route-specific prompts owned by the right policy owner

Solution

Two-step Bedrock Converse flow: classify into policy | technical | escalation | general, then generate with the matching system prompt from lib/demos/intent-routes.ts.

Technical implementation

Architecture

Classify first, answer second, cheaper than a monolithic agent when intents are well understood.

How it runs
Drawing the flow…

Outcomes and learnings

  • Start with explicit intents; add ML routing only when taxonomy stabilizes
  • Log intent and routedTo on every turn for quality review
  • Graduate to Bedrock Agents when tools and memory enter the picture
Delivery playbookDiscovery → pilot → scale
  1. 1
    Discovery2–4 wks

    Document intent taxonomy with owners; label 200 historical tickets for accuracy baseline.

  2. 2
    Pilot6–8 wks

    Route three intents in a single internal assistant; log misroutes weekly.

  3. 3
    Scaleongoing

    Promote to Bedrock Agent with tools per intent; add guardrails per route.

Where else this appliesIntent routing is how you keep one entry point (search box, chat widget, IVR text) without one prompt trying to do everything poorly.

Employee self-service portal

Route HR, IT, and facilities questions to different knowledge bases and tone guidelines.

Retail customer care

Separate returns, loyalty, and technical product intents before hitting CRM tools.

B2B partner helpdesk

Distinguish commercial terms questions from integration how-tos with audit-friendly labels.

Voice-to-text pipelines

Classification step after transcription before downstream automations run.

Bedrock Converse classifies cheaply; route to specialized prompts, Lambdas, or Agents per intent while staying in the same AWS account and CloudWatch logging.