Non-use of enterprise data for training models is table-stakes for enterprise ML products. Google does the same thing, for example.

They'll want to climb the compliance ladder to be considered in more highly-regulated industries. I don't think they're quite HIPAA-compliant yet. The next thing after that is probably in-transit geofencing, so the hardware used by an institution reside in a particular jurisdiction. This stuff seems boring but it's an easy way to scale the addressable market.

Though at this point, they are probably simply supply-limited. Just serving the first wave will keep their capacity at a maximum.

(I do wonder if they'll start offering batch services that can run when the enterprise employees are sleeping...)