20 cases before launch: a minimal eval set

You do not need a platform to start testing an AI feature. You need about twenty real cases, a rule for passing, and a rule for when not to ship. Plenty of teams launch an AI feature after trying it o