PHI and health data

PHI under HIPAA

HIPAA defines protected health information as health information that can identify a person, sent or kept in any form. To count, the information must come from a health care provider, a health plan, an employer or a clearinghouse. It must be about a person’s health, care or payment for care. And it must identify the person, or give a reasonable way to identify them. Some records are left out, such as job records an employer keeps, school records covered by FERPA, and records of people who died more than 50 years ago.¹

The Security Rule then asks covered entities and their business associates to protect electronic PHI. Two safeguards matter most when software agents act on it. One is audit controls that “record and examine activity” in systems that hold electronic PHI. The other is protecting it from being changed or destroyed without permission.² The documents the Security Rule asks for are kept for six years from when they were made or last used, whichever is later.³

Health data under India’s DPDP Act

India’s Digital Personal Data Protection Act, 2023 has one category. “Personal data” means any data about a person who can be identified by or through that data.⁴ There is no separate “sensitive” or health class. So a diagnosis carries the same duties as an address. The hospital, clinic or app that decides why and how the data is used is the data fiduciary.

The DPDP Rules, 2025 list “appropriate logs, monitoring and review” among the minimum safeguards. They ask for those logs to be kept for one year. These duties start eighteen months after the rules were published in November 2025.⁵ Health data that moves over the Ayushman Bharat Digital Mission network also passes through a consent manager. The mission calls it “data blind”, with the person’s consent for every data transaction.⁶ India’s EHR Standards add that every action on health information, including viewing, should be recorded.⁷

Why it matters when AI agents touch health data

An agent that books visits, drafts notes or files claims sees health data in three places: its prompts, the details it sends to tools, and the results it gets back. Each of these can be copied somewhere new. It might land in the agent’s own trace, a gateway’s request log, an approval message in a team chat, or a monitoring tool in another country. Under both laws, every copy is health data. And each copy adds one more system that needs the same protection.

Keeping PHI out of the audit trail

An audit trail has to say what happened without becoming one more store of patient data. Three habits help:

• Record ids and digests, not content. The kind of record, the server’s id, a fingerprint of the request and of the result, the rule and the outcome are enough to check an action later.
• Salt the patient reference. A medical record number or a health ID has few possible values. So a plain fingerprint of it can be guessed by trying every value. A secret salt that only the hospital holds stops that.
• Keep full requests for a short time. If full requests are stored to fix problems, keep them for the shortest time that works. Keep them on servers in the country where the data belongs.