AI in Clinical Documentation: HIPAA Requirements and Best Practices

Clinical documentation is the highest-volume source of protected health information (PHI) in any health system. Every visit, every note, every dictated line generates it. AI is transforming how that documentation gets created, with ambient scribes now listening to patient encounters and drafting notes in real time. That shift raises new HIPAA questions at each step the audio takes, from the microphone in the exam room to the signed note in the chart. Those questions add to the general framework for HIPAA-compliant AI rather than replace it.

The appeal is obvious: physicians spend a large share of their working hours on documentation, and ambient AI tools give meaningful time back. But an ambient scribe captures far more PHI than the note that eventually lands in the record. It captures the raw audio of the encounter, an intermediate transcript and every draft the model produces before a clinician signs off, each of which is PHI, and each of which needs its own HIPAA safeguards, not just the version that gets signed and filed.

This article covers what counts as PHI in a clinical note, where AI documentation tools create risk, what to check in an ambient AI vendor’s BAA, a set of best practices for rollout and how de-identification adds a compliance layer on top of whatever documentation tool a health system chooses. The summarization step in particular runs into the same LLM deployment questions that come up anywhere an LLM turns raw text into a finished clinical note ready for a physician’s signature.

What counts as PHI in clinical documentation

The 18 HIPAA identifiers apply to a clinical note exactly as they apply to any other record, but a few show up constantly in documentation specifically and are easy to overlook: the patient’s name spoken aloud during the visit, any dates tied to the encounter or the patient’s history, the voice itself (biometric identifiers are on the list) and any unique identifying number the practice uses internally. A transcript of a patient encounter is PHI the moment it exists, whether or not it has been formatted into a chart note yet, and the underlying audio recording is PHI in its own right, separate from the transcript derived from it.

Family members and caregivers present during a visit add a layer that is easy to miss. If a spouse’s name or a child’s school comes up during the encounter, whether it’s clinically relevant or just small talk, it’s captured in the same audio and transcript as the patient’s own PHI and needs the same handling. HIPAA’s de-identification standard treats identifiers of a patient’s relatives and household members the same way it treats the patient’s own, even though those people were never the patient of record.

Where AI clinical documentation tools create HIPAA risk

Three stages of the ambient documentation workflow carry distinct risk.

Transcription. Converting audio to text is usually the first point where PHI is processed by a third-party system. If the transcription engine runs in a vendor’s cloud without a BAA covering that specific function, the transcription step itself is a potential violation before a single word reaches the note.

Summarization. An LLM turning a raw transcript into a structured SOAP note is generating new content from PHI, which means the summarization step needs the same access controls and audit logging as the transcript it draws from, and needs review, since a model can hallucinate or misattribute details that were never actually said.

Ambient listening. Always-on or session-long audio capture raises a scope question that text-based tools do not: what happens to audio captured before or after the clinically relevant part of the encounter, and how long is the raw recording retained after the note is finalized? A retention policy that is not clearly documented, and cannot be audited, is not a real control. A scribe left running while a clinician steps out to take an unrelated call, for instance, captures PHI (or non-PHI conversation entirely unrelated to the visit) that was never meant to be part of the record, and that audio needs the same retention discipline as the clinically relevant portion.

The BAA question for clinical AI vendors

Leading ambient documentation platforms, including Nuance DAX (part of Microsoft), Abridge, Suki and Nabla, publish BAAs for covered entities, and a signed BAA with one of these vendors is the floor for using their tool with PHI, not the finish line. Retention and training-use terms vary meaningfully across all four. Suki’s own developer documentation states that raw audio is permanently deleted after 30 days, and that any data used for model training is first de-identified, with audio chunked so the original cannot be reconstructed and PHI stripped from transcripts; Abridge’s privacy policy similarly describes a comparable deletion window and states it uses de-identified data to refine its models. Nabla’s own help center states that audio is not stored at all, processed in chunks and discarded, while transcripts and notes are retained for a configurable 14 days by default to give clinicians time to review and export to the EHR, with backups purging seven days after that. Nuance DAX, now sold as part of Microsoft’s Dragon Copilot, retains audio and other identifiable health data for up to 90 days before deletion or anonymization, per a covered entity’s own published patient privacy notice describing the product. The pattern across all four: retention windows range from zero (Nabla’s audio) to 90 days (Nuance DAX), so the specific number matters more than the vendor’s general “HIPAA compliant” marketing language, and it should still be confirmed against your organization’s own signed agreement, since these terms can change.

Documentation stage	Primary HIPAA question	Who typically controls it
Audio capture and transcription	Is the transcription engine covered by a signed BAA specific to that function?	AI scribe vendor, under contract
Note summarization	Does the model retain or train on transcripts, and is clinician review required before signing?	AI scribe vendor and organizational policy
Retention and deletion	How long is raw audio kept after the note is finalized, and is that timeline documented and auditable?	AI scribe vendor’s contractual commitment

Procurement teams evaluating a new ambient scribe should ask for these three answers in writing during the vendor selection process, not after the contract is signed. A vendor that can point to a specific, published retention window and a de-identification step before training, the way Suki and Abridge do, is easier to evaluate than one that only offers a general assurance of “HIPAA compliance” without documentation behind it.

Limina de-identifies clinical documentation in real time—audio, text and structured notes.

De-identification as a compliance layer for clinical documentation AI

A de-identification step, placed between transcription and any downstream use of the note, reduces how much PHI is exposed to secondary systems, such as analytics platforms, coding tools and quality review processes, without requiring a fresh BAA negotiation with every system that touches the note afterward. This does not replace the BAA with the scribe vendor itself, since the vendor still needs to process the identifiable audio and transcript to produce the note in the first place. It does mean that everything downstream of the finished note can work with de-identified text, which shrinks the number of systems that need direct PHI access and a corresponding BAA. For teams already de-identifying call transcripts and other audio, extending the same pipeline to clinical encounters is a natural next step rather than a new build.

Best practices for HIPAA-compliant AI documentation

The requirements above define the floor. Six practices separate a rollout that holds up under audit from one that only looks compliant on paper.

• Confirm the BAA by product and tier, not by vendor name. A signed agreement with a scribe vendor covers the specific product it names; verify that the tier your clinicians actually use, not a pilot or trial version, is the one listed.
• Require clinician sign-off before any AI-drafted note enters the legal record. This is a patient-safety control as much as a compliance one, since ambient tools can hallucinate or misattribute findings.
• Set a documented retention window for raw audio, and confirm it against the signed agreement. Do not assume a vendor’s marketing page reflects the terms your organization actually signed.
• Route de-identified output to any system that does not need the identifiable version. Analytics, coding support and quality review rarely need raw PHI to do their job.
• Scope ambient listening deliberately. Capture only the clinically relevant portion of the encounter where the tool allows it, and document why any broader capture window is necessary.
• Add AI documentation tools to the organization’s existing HIPAA risk assessment and BAA inventory. Treat rollout as a compliance decision from day one, not an IT pilot that compliance reviews after the fact.

How Limina handles clinical documentation PHI

Limina AI de-identifies clinical documentation in real time, across audio, text and structured notes, so PHI captured during a patient encounter can be redacted before it flows into analytics, coding support or any other downstream system that does not need the identifiable version. The platform supports 52 languages and all 18 HIPAA identifiers, and reaches 99.5 percent or higher accuracy on real physician conversations, audio being the hardest modality to de-identify accurately given accents, cross-talk and clinical shorthand. Providence Health used this approach to de-identify physician conversations and EHR records entirely within its own environment after years of clinical data sitting unused over PHI exposure concerns, evaluating and rejecting major cloud providers first over how they would use the data. The same underlying approach extends beyond documentation to broader operational text, as covered in this overview of privacy risk across unstructured data sources.

See how ANALYZE puts this real-time de-identification pipeline to work for clinical documentation teams.

Vendor BAA terms and audio retention windows referenced above are current as of September 2026 but change on a rolling basis. Confirm current terms directly with each vendor before making a compliance decision.