Support Router
Every ticket lands on the right desk with a first reply already drafted.
Tickets piled into one queue and my team spent half the day just deciding whose problem each one was. The hard questions waited behind the easy ones, and everyone felt slow.Head of Support, Series-A fintech
Anything touching money or account access skips the draft and goes straight to a human.
We began by reading a quarter of resolved tickets to learn the shape of the queue. Underneath the mess were a handful of real intents — billing questions, onboarding friction, bug reports, refund and dispute requests, and account-access issues — each of which belonged to a specific person. The router’s whole job is to make that routing decision correctly and quickly.
Classification runs two passes at once. Claude reads the ticket for intent so it knows which desk it belongs to, and for risk so it knows whether a machine is allowed to draft a reply at all. Low-risk, well-understood intents get a first draft in the team’s voice; anything touching money or identity is routed but left untouched, flagged for a human to own from the first word.
The drafts are grounded in the team’s own past replies and the current help centre, so they answer in-house rather than in generic support-speak. An agent still reads every one before it goes out — the router speeds the first response, it doesn’t replace the person behind it.
Shipping into a regulated product meant the guardrails came first, not last. Tickets never auto-close, the risky categories never auto-send, and a single switch drops the whole thing back to manual routing if a policy changes overnight.
Whether the router could resolve tickets on its own. It never closes a ticket and never auto-sends on anything involving refunds, disputes, or account access — those always wait for a person. In a regulated business, a fast wrong answer costs more than a slow right one.
On retainer we retrain the intent classifier as the product ships new features, adjust the risk rules whenever compliance policy changes, and prune reply templates that drift out of the team's voice.