for healthcare
Healthcare
HIPAA's minimum-necessary standard binds an agent the way it binds a person: a read from your EHR is a use of protected health information, limited to what the task needs. Retrieval scoping, the Section 1557 and HTI-1 decision-support obligations, SB 1120's human-final-decision rule, and the OCR-ready audit trail all follow from that.
Minimum necessary binds the agent
The minimum-necessary standard was written for people, and it applies to an agent the same way. A scheduling assistant has no business reading a full chart, and "the model needed context" is not a defense the Office for Civil Rights recognizes.
The control that matches the standard sits at retrieval time: scope what the agent can read to the task at hand, mask the fields that exceed it, and deny the requests that reach past both.
Every read is a use HIPAA regulates
Once an agent is wired to an EHR, a claims system, or a clinical knowledge base, each read is a use of protected health information in HIPAA's terms, and a disclosure the moment what it read leaves your organization. Minimum necessary governs both, and HIPAA expects a covered entity to enforce it through role-based access policies.
HIPAA does not demand a ledger of every internal use. The Security Rule does require audit controls over systems that hold ePHI, and an OCR inquiry into whether minimum necessary held is answered with a record or not at all. Logging the agent's final answer misses the exposure: the risk is in what the agent saw along the way, and only a record written at retrieval time captures it.
The decision-support rules have arrived
Section 1557's 2024 final rule prohibits discrimination through patient care decision support tools, with covered entities required to identify and mitigate that risk from May 1, 2025. ONC's HTI-1 rule makes certified EHRs disclose how their decision-support algorithms work.
California's SB 1120 goes further for utilization review: the AI must weigh the individual patient's clinical record, and a licensed professional makes the final medical-necessity call. An agent that drafts triage, scheduling, or coverage-relevant output sits inside this perimeter.
Answering an OCR inquiry
An OCR inquiry asks what was read, for which task, and why the access was necessary. A record written at retrieval time already holds the answers: it names the patient context and the permitting policy, and it shows where masking narrowed what came back.
The same record carries your 1557 mitigation evidence: what the agent was given, what it produced, and who signed off where sign-off was required.