A model does not make a system HIPAA ready. Safeguards do. If the rest of the product cannot show who accessed ePHI, the copilot will not save you. If it can, the copilot has to obey the same rules.
We design AI features inside the product boundary: same identity, same logging, same change control. Retrieval is limited. Prompts are treated as data. Vendors that see tokens need a BAA or they do not see ePHI. That is slower than pasting a chart into a consumer chatbot. It is the only version we will ship.
A practical bar
Before we wire a model, we ask where the data lives, who is allowed to see it, and what happens when the answer is wrong. Then we implement: role based access, audit trails, human review for anything clinical, and evals that fail the build when groundedness drops.
Your assessor still issues the opinion. We prepare the architecture. Independent of brand claims, the test is simple: can you reconstruct a session, show the sources, and revoke access without a war room?