In Depth
The Health Insurance Portability and Accountability Act sets the federal rules for how health data is used, shared, and protected. Two of its rules do most of the work for AI vendors: the Privacy Rule, which governs when PHI may be used or disclosed, and the Security Rule, which requires safeguards for electronic PHI. Both live in the Code of Federal Regulations at 45 CFR Parts 160 and 164. PHI itself is any individually identifiable health information, such as a diagnosis, a treatment, or a payment record, that can be tied back to a specific person.
HIPAA's obligations flow downstream. A hospital or insurer is a covered entity. A vendor processing PHI on that entity's behalf is a business associate, bound by the same disclosure rules through a business-associate agreement (BAA). If your AI agent reads charts, drafts notes, answers clinician questions, or routes patient messages, you are operating in that chain, and the rules apply to you directly, not just to your customer.
Enforcement has teeth. HIPAA carries tiered civil monetary penalties that rise with the degree of culpability, and serious violations can draw regulatory investigation and corrective-action requirements. The exact dollar figures move with inflation adjustments, so the durable point isn't a number. It's that an unauthorized disclosure of PHI is a federal-compliance event, not merely a contractual one.
What It Looks Like
A startup sells an AI agent that triages inbound patient messages for a clinic. To answer faster, the agent retrieves across the clinic's full message history. A patient asks a routine billing question, and the agent's answer includes a fragment of another patient's appointment and condition, pulled in by an over-broad retrieval step. That single response is an unauthorized disclosure of PHI. The clinic, as the covered entity, now has a compliance problem, and its BAA points the obligation straight at the vendor.
Why It Matters For AI Vendors
In healthcare, HIPAA is the gate. No covered entity will deploy your agent without a BAA, and their security review will press hard on how PHI is stored, accessed, and prevented from leaking. A vendor who can't answer those questions cleanly doesn't get the deal.
The harder reality is that HIPAA compliance and AI risk aren't the same thing. You can have a signed BAA, encryption, and access logs, the controls the law expects, and still ship an agent that discloses PHI through normal model behavior, because the law's safeguards were written for systems that store and transmit data, not for systems that generate it. The compliance posture and the AI failure mode have to be managed together.