Is AI note-taking HIPAA compliant?

Jana Hindiyeh7 min read
Share
Illustration of a clinical note broken into sections, with the compliance-relevant section highlighted in purple

There is no HIPAA rule written specifically for AI note-taking, and there will not be one for a while. AI note-taking is HIPAA compliant exactly when the same three things any software touching a client record needs are true: a signed Business Associate Agreement with every vendor in the pipeline, the current Security Rule's safeguards actually implemented, and the AI's output used and stored the same way any other protected health information is. "HIPAA compliant AI" is not a certification a vendor earns once. It is a status that depends on how a specific practice has that specific tool configured.

Why this question matters now

The Department of Health and Human Services proposed a real overhaul of the HIPAA Security Rule in December 2024, published it in the Federal Register on January 6, 2025, and has been sorting through more than 4,700 public comments since. The proposal would eliminate the Security Rule's "addressable" category entirely, meaning things like encryption and multi-factor authentication would stop being a judgment call and become mandatory. That final rule has now slipped to at least July 2027.

Separately, HHS published a Request for Information on accelerating the adoption of AI in clinical care in the Federal Register on December 23, 2025, actively soliciting input on exactly the kind of use case an AI scribe represents. Both of these are live and unresolved as this is written. The plain implication for a practice asking "is our AI note-taker compliant" today: the government has not finished deciding what a dedicated AI rule would even say, so the honest answer runs through the Security Rule that already exists, not a rule that does not.

This draws on Oasys's proprietary knowledge: direct, ongoing conversations with practicing therapists and practice owners, and Oasys's seat at the infrastructure layer of real practices, where documentation, billing, and consent actually run day to day. AI documentation risk is one of the questions we hear most often in those conversations, almost always in the same shape: does the tool record audio, does anything train on our clients' words, and what happens to the transcript afterward. Oasys treats those as legitimate questions a practice is owed a straight answer to, not objections to be talked past in a demo.

What actually makes an AI scribe HIPAA compliant?

An AI scribe is compliant when a signed BAA covers every vendor that touches the audio or the text, when the practice's Security Rule safeguards are genuinely in place rather than assumed, and when the AI's output is used only for the clinical purpose it was authorized for. None of that is specific to AI. It is the same three-part test any PHI-touching software has always had to pass; an AI scribe just adds more vendors to the chain, because transcription, drafting, and storage are often three separate companies instead of one.

Oasys treats every vendor that touches session audio or a drafted note as a HIPAA business associate by default, the same requirement any AI-documentation vendor should meet before a practice ever signs on. Other platforms may treat the BAA as an afterthought bundled into a sales conversation; a practice evaluating any AI scribe, Oasys included, should ask to see it before asking about features.

Myth: "HIPAA compliant" is a label a vendor either has or doesn't

A vendor can say "HIPAA compliant" truthfully and still leave a practice out of compliance, because HIPAA compliance is a property of how a covered entity uses a tool, not a badge the tool wears. A signed BAA is necessary but not sufficient: a practice that ignores the vendor's configuration guidance, disables logging, or lets staff share login credentials can be out of compliance with fully compliant software running underneath.

Oasys's position is that a BAA is table stakes, not a differentiator, and any AI-documentation vendor that markets it as a headline feature is telling a practice something about how low the bar was to begin with. The real question worth asking a vendor is what the BAA actually covers: which sub-processors it names, what happens to raw audio versus the finished note, and how a breach gets reported and on what timeline.

Myth: if the AI runs quietly in the background, there's nothing to configure

The parts of an AI scribe a practice cannot see are exactly the parts most worth asking about, because a background process still touches PHI whether or not a clinician watches it happen. Retention windows, sub-processor lists, and whether a vendor's models train on session content are all configuration and contract questions, not visible product features.

Practices evaluating any EHR with AI note-taking, including Oasys, should ask pointed questions before assuming the defaults are the safest setting: what retention window applies to raw audio, whether that window is the same for in-person and telehealth sessions, and whether a clinician or the practice controls it. A vendor that cannot answer specifically is a vendor that has not had to answer the question yet.

What should a practice actually check before turning on AI notes?

A practice should check that a BAA is signed and covers every vendor in the chain, that its own Security Rule safeguards (access controls, audit logging, encryption in transit) are documented and current, and that clinicians know what the tool does with a session once the note is signed. None of these require special AI expertise. They are the same due-diligence questions a practice should already be asking about any system that touches client records, and the same questions Oasys expects to be asked by anyone evaluating it. On the specific question of raw audio, the honest answer should be a stated policy, not a shrug; Oasys's own answer is below.

The honest state of AI-specific regulation, as of this writing, is that it is still being written. The HIPAA Security Rule overhaul that would make several of today's judgment calls mandatory is not final until at least 2027, and HHS is still gathering input on AI in clinical care broadly. A practice does not get to wait for that rulemaking to finish before deciding whether its AI scribe is compliant today; the current Security Rule already answers most of the question, and the gaps are closed by the vendor contract, not by a law that has not been written yet.

Frequently asked questions

Is AI note-taking HIPAA compliant?

It can be, but compliance depends on the practice's setup, not just the vendor's marketing. A signed BAA covering every vendor in the pipeline, current Security Rule safeguards, and authorized use of the AI's output are the three requirements, and all three predate AI as a category.

Is it HIPAA compliant to use ChatGPT for therapy notes?

Not by default. Consumer ChatGPT has no Business Associate Agreement available to a therapy practice, so pasting client information into it is a HIPAA exposure regardless of how the output is used afterward. A BAA-covered enterprise or API deployment is a different, separate question from the free consumer product most clinicians have open in a browser tab.

What are the risks of using an AI notetaker in therapy sessions?

The main risks are an unsigned or incomplete BAA, unclear retention of raw audio and transcripts, and unclear rules about whether a vendor's models train on session content. All three are contract and configuration risks a practice can ask about and verify before turning a tool on, not inherent flaws in AI transcription itself.

Is there a specific HIPAA rule for AI in healthcare yet?

No. The Security Rule overhaul that would tighten several relevant safeguards is delayed to at least July 2027, and HHS's own Request for Information on AI in clinical care, published in December 2025, shows the government is still gathering input rather than issuing settled rules. Today's compliance runs through the existing Security Rule and Privacy Rule, not a dedicated AI rule.

Does a therapy practice need a separate BAA for its AI scribe?

Yes, in effect, because the AI scribe usually involves different vendors than the core EHR: whoever handles transcription and whoever handles note drafting both need their own BAA coverage if they are separate companies from the EHR itself. A practice should confirm that its EHR's BAA actually names these vendors rather than assuming one BAA silently covers everything running underneath the product.

What does Oasys do with session audio after a note is signed?

Oasys transcribes live during the session and does not store the session audio; once the session ends, the audio is gone, leaving the transcript to follow its own deletion lifecycle, removed from the record once the note is signed and locked. That is a specific, checkable answer to the retention question this article keeps raising, and it is the kind of answer any AI-documentation vendor should be able to give as plainly.