I've shipped this in production twice: the email classifier in Mudroom, which users can correct when it gets one wrong, and the AI chat in FitMirage. In Mudroom, plain rules handle the easy cases first and the model handles the rest, which keeps it cheap and predictable.