Routing AI chat requests

Route by intent: classify the question into a small set of buckets, send hard buckets to the strong model, easy buckets to the cheap one, and keep a fallback chain for uncertainty.

Route by intent, not by hope

Classify each message into a small set of buckets: factual question, troubleshooting, sales, complaint, chitchat. Then route:

• Hard buckets (complaint, legal, refund) go to the strongest model with the strictest guardrails.
• Easy buckets (hours, status, definitions) go to a small fast model.
• Uncertain goes to a fallback chain: retry once, then escalate.

Keep the buckets few. Ten intents is enough. Fifty is a taxonomy you'll maintain forever.