Would you mind fleshing this out a bit? I'm not following how it identifies the sensitive pieces without sending to a model, or do you mean letting the model see messages, calendar, email, etc. but not able to go beyond those?
So the model can see the results, but the access is whitelisted and temporary.