A browser sandbox for the configuration surface: persona and tone, the knowledge it can draw on, escalation triggers, and how it handles a booking. Then run your own scenarios against it and watch where it fails. Nothing is saved and nothing is billed.
Configure it deliberately badly first. Leave out an escalation trigger and watch it answer something it should have handed off. That is more instructive than a clean configuration, and it is the fastest way to understand why escalation design matters more than persona.
No phone leg, no live calendar, and no CRM write. Those are the parts that need credentials and an eval pass. The sandbox is honest about the boundary rather than simulating a booking that would not happen, because a fake confirmation is the least useful thing a demo can show you.