PRODUCTION SCENARIO
An e-commerce support platform receives 2,000,000 customer messages a day, all as free text with no ticket type attached. About 80 percent are order-status lookups answerable from one tool call; the rest are multi-order disputes and warranty claims that need several tool calls and real judgment. Customers expect a reply within seconds, and the platform is measured on first-contact resolution, so sending a customer away and waiting for them to write back is not acceptable.
Which architecture keeps cost and latency down without losing quality on the hard cases?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Classify at ingress on Haiku 4.5 and route only the disputes to a Sonnet 5 agent
A fast, cheap model at the ingress layer resolves the 1.6 million messages that need one lookup and hands the rest to a stronger agent, so the higher price and latency are paid only where judgment is needed. Routing on a customer-selected type is right when intake offers those types and customers pick them accurately, and these arrive as free text. Running the middle tier everywhere and escalating mid-conversation is right when the cheapest tier cannot handle routine lookups, and it pays Sonnet's price on 1.6 million of them. Escalating on the customer's reply is right when a second round trip is acceptable, and first-contact resolution is the metric here.
Official Source: https://platform.claude.com/docs/en/models/overview ↗