Support AI
𧬠A/B experiments (live router)
When the router detects a situation on a real incoming email, it serves ONE variant at random and records it. When the customer replies, it scores whether they accepted the offer (and at what tier β a lower accepted cashback is a cheaper win). Only fires on matching emails; everything else drafts normally.
β Agents are in observe/draft mode β variants are assigned & drafted, but acceptance outcomes only fill in once the reply actually reaches the customer (when you send the variant draft). The split itself runs now.
New experiment
β³ Working β¦