What we learned scaling a support assistant

A pilot that delights ten users can buckle under ten thousand. Scaling a support assistant taught us as much about operations as about models.

Routing, fallback to humans, and honest confidence signals turned out to matter more than raw accuracy. Users forgive a system that knows when to ask for help.

The biggest win was not deflection rate — it was the data the system surfaced about what customers actually struggle with.