Skip to content

AI tools and PHI: what the HIPAA Security Rule already requires

By Sam Rivera, Founder, SentinelPanda · November 7, 2026 · 4 min read · HIPAA

The question staff actually ask is simple: can I paste this into the assistant? HIPAA answers it with rules written twenty years before generative AI.

The first question is the contract

Under HIPAA, an organisation that creates, receives, maintains or transmits protected health information on behalf of a covered entity is a business associate. An AI service that summarises clinical notes, drafts patient letters or transcribes appointments does exactly that. Before PHI reaches it, the Security Rule requires satisfactory written assurances, in the form of a business associate agreement (164.308(b)(1)), with the contract terms set out in 164.314(a).

Many enterprise AI offerings will sign an agreement, usually for a specific product tier or configuration. Consumer versions of the same tools usually will not. So the question that matters is less "which AI tool" than "which account, under which agreement".

De-identified means de-identified

Data de-identified under the Privacy Rule is no longer PHI, and HIPAA does not restrict its use. But de-identification has a specific meaning (164.514(b)): either the Safe Harbor method, which removes eighteen categories of identifiers including dates more specific than the year and most geographic detail, or a documented expert determination that the risk of re-identification is very small.

Deleting the name from a note before pasting it into an assistant meets neither. Dates of service, ages over 89, rare conditions, ZIP codes and free-text mentions of relatives routinely remain. Treat anything short of a documented method as PHI.

The risk analysis has to include it

The risk analysis (164.308(a)(1)(ii)(A)) must assess risks to all ePHI the organisation creates, receives, maintains or transmits. An AI tool in use, sanctioned or not, is part of that picture. Record where PHI flows into the tool, what the vendor retains and for how long, where it is processed, and whether it is used to train models.

Risk management (164.308(a)(1)(ii)(B)) then requires measures that bring those risks down to a reasonable and appropriate level. For most organisations that means an approved-tools list, a configuration standard (enterprise tenancy, retention and training use switched off) and a clear rule for staff.

The technical safeguards still apply

  • Access control (164.312(a)(1)): only workforce members who need the tool for their role should have it, under their own accounts.
  • Audit controls (164.312(b)): keep a record of who used the tool with PHI; the vendor should be able to provide activity logs for your tenant.
  • Integrity (164.312(c)(1)): AI-drafted text that enters the medical record changes ePHI, so a person should review it before it does.
  • Person or entity authentication (164.312(d)): single sign-on and multifactor authentication on AI accounts, never shared logins.
  • Transmission security (164.312(e)(1)): encrypted connections to the service. Standard for reputable vendors, but worth confirming for browser extensions and plug-ins.

Minimum necessary still applies

Even under an agreement, the Privacy Rule's minimum-necessary standard (164.502(b)) applies to most uses and disclosures. An assistant drafting a referral letter does not need the full chart. Prompt templates that pull only the fields a task needs are a minimum-necessary control as much as a convenience.

Training and incidents

Security awareness and training (164.308(a)(5)) should now cover AI explicitly: which tools are approved, what may go into them, and how to report a mistake. Someone who pasted PHI into an unapproved tool needs a quick way to say so, because it may be a security incident under 164.308(a)(6) and, depending on the facts, a breach that requires notification.

HHS proposed substantial updates to the Security Rule in January 2025, including more prescriptive asset inventories and risk analysis. Whatever their final form, an organisation that already inventories its AI tools and records the PHI flowing into them will be ahead of them.

The short version

  • No business associate agreement, no PHI. Check the account tier, not just the product.
  • De-identify with a documented method, or treat the data as PHI.
  • Put AI tools in the risk analysis, including the ones staff adopted on their own.
  • Individual accounts, single sign-on and MFA, and activity logs you can retrieve.
  • Send the minimum necessary, and review AI-drafted content before it enters the record.
  • Tell staff how to report a mistake, and treat reports as possible incidents.
HIPAA business associate agreements HIPAA de-identification HIPAA security risk analysis AI security risks across frameworks

Run your compliance program in one workspace.