Skip to content

AI Governance & Ethics

Before an AI Medical Scribe Listens, Check These BAA Gaps

A signed BAA can govern protected health information without answering how long audio survives or whether transcripts train models. Trace the data path and test deletion before the first patient visit.

Irene VaskoGovernance & Ethics Writer

August 9, 2026 · 8 min read

A clinic-owned phone beside a laptop showing an AI scribe retention and access review checklist.
A clinic-owned phone beside a laptop showing an AI scribe retention and access review checklist.

Picture one clinic-owned phone on an exam-room counter. A clinician taps record, the medical scribe streams the conversation to a cloud service, speech recognition produces a transcript, and a language model turns that transcript into a draft note for review. The clinician edits the draft and sends the final note to the electronic health record, or EHR.

That short workflow can create more copies than the phone suggests: buffered audio, a server-side recording, transcript segments, the generated draft, support logs, analytics events, backups and the final EHR note. A business associate agreement, or BAA, sets contractual duties for a vendor handling protected health information on behalf of a HIPAA-covered organization. It does not, by its existence alone, tell the clinic which copies exist or settle every permitted use.

Before that phone hears a patient, the reviewer needs evidence at three levels. The signed contract says what the vendor may do. Product settings control what the deployed account does. Logs and deletion tests show whether those controls operated.

A security webpage is supporting documentation, while a roadmap promise is only proposed behavior.

This is a procurement and deployment checklist, not legal advice. Recording consent, professional obligations and state privacy rules need their own review; a BAA does not replace them.

Draw the visit’s data path first

Start with the phone and follow one synthetic visit, using invented patient details until the contract is complete. Record where the app sends audio, whether it stores a local copy, which service performs transcription, which model generates the note, and how the result reaches the EHR. Include browser extensions, mobile crash reporting, customer support tools and analytics services if they can receive identifiers or clinical text.

Do not accept “the data is encrypted” as a data-flow answer. Encryption protects information in transit or storage, but it does not determine who can decrypt it, how long it remains, or whether the vendor can use it for another purpose. Ask for an architecture diagram and a written description that names each storage location and processing vendor.

Return to the phone after drawing the map. If the clinician stops recording but the upload fails, establish whether audio remains in an application cache and who clears it. If connectivity drops, document whether the fallback is local buffering, a failed session or ordinary manual documentation. Offline buffering may improve reliability, yet it also puts another copy on a device that can be lost.

Read the BAA beside the commercial terms

HIPAA’s contract rule requires a BAA to “establish the permitted and required uses and disclosures” of protected health information by the business associate. That language matters. A document titled BAA can satisfy the need for a contract while leaving broad uses permitted through incorporated service terms, an order form or another policy.

Read those documents together and resolve conflicts in writing. Identify which document controls, whether the vendor may change online terms without a new signature, and whether the BAA covers the exact product tier, mobile application, transcription service and model endpoint being purchased. A BAA signed with a parent company does not automatically answer which affiliate operates a particular feature.

Separate treatment-related processing from product improvement. The vendor needs to process the visit to create the requested note. Training a general model, evaluating future features or allowing humans to label transcripts is a different use, even if the vendor describes it as improving service. Require the contract to state whether protected health information, recordings, transcripts, prompts, outputs and feedback may be used for training or evaluation.

If the answer relies on de-identification, require the vendor to identify when de-identification occurs, what data exists beforehand and whether the BAA permits that step.

Check recordings and transcripts as separate records

Audio and text should have separate retention entries because vendors may delete the recording soon after transcription while keeping the transcript, draft or derived metadata for longer. “No audio retention” says nothing about text. “Zero retention” may describe only the outside model provider rather than the scribe application itself.

Use the data map to complete these checks:

  • Recording control: Confirm whether recording starts only after a deliberate clinician action, how the patient’s recording preference is represented, and whether audio can be paused without ending the encounter. The workflow needs a visible recording indicator and a manual fallback for a patient who declines.
  • Retention clock: Put the retention period and its starting event in signed terms. “Deleted after processing” is incomplete unless processing has a defined endpoint. Cover abandoned sessions, failed uploads and notes that remain unapproved.
  • Transcript handling: Establish whether users can view, export or search the full transcript; whether administrators can disable those functions; and whether deleting audio leaves searchable text behind. Restrict bulk export if the clinic does not need it.
  • Derived data: Include note drafts, speaker labels, suggested diagnoses, quality scores and user feedback. A vendor can delete the raw recording while retaining outputs that still identify the patient.

There is a tradeoff. Short retention reduces exposure, while longer retention can help a clinician investigate a disputed transcript or let support reproduce a failure. Make that choice explicitly. If the clinic chooses short retention, the fallback is timely clinician review and ordinary EHR correction, not an undocumented vendor archive.

Follow every subcontractor

A subcontractor, sometimes called a subprocessor in vendor documentation, is another company that handles data to deliver the service. The BAA should require downstream entities handling protected health information to accept the same applicable restrictions and conditions. The clinic still needs the current list.

Match each name to a job in the exam-room workflow: hosting, speech recognition, language-model inference, customer support, authentication or error monitoring. Record the processing location and the data categories each receives. A provider that sees only an account identifier presents a different exposure from one receiving the complete transcript.

Set a review path for list changes. Vendor documentation may offer advance notice, an objection window or neither; the signed agreement should establish the mechanism the clinic expects to use. An objection right with no practical migration plan has limited value, so preserve a fallback such as disabling recording, switching to manual dictation or pausing the feature while a new provider is assessed.

Make access controls observable

The phone should open into a named clinician account, not a shared exam-room login. Require single sign-on where practical, multifactor authentication for privileged users, and role-based access control, which limits actions according to a person’s assigned role. Define who can replay audio, read transcripts, export encounters, change retention settings and invite users.

Then inspect the audit trail. It should record sign-ins, recording access, transcript views, exports, deletions, administrative changes and support access with timestamps and account identifiers. Ask how long those logs remain and whether the clinic can export them without opening a support ticket. A log that exists only for the vendor’s engineers is weak evidence during an internal investigation.

Support access deserves its own control. Determine whether staff can open a patient encounter by default, whether access requires clinic authorization, and whether the event appears in the clinic’s audit log. Screenshots and ticket attachments can create additional copies, so the support workflow belongs on the same data map as the transcription model.

Run a deletion test before approval

Create a synthetic encounter on the same device and account configuration intended for clinical use. Give it a distinctive invented phrase, allow the scribe to generate a note, export whatever the product permits, then delete the encounter through the normal interface. Search for the phrase in the app, administrator console and available exports. Review the audit log to confirm who requested deletion and when.

Next, ask support to describe what remains in active storage, logs and backups. Backup deletion often means expiration under a rotation schedule rather than immediate removal from every copy; the contract should state that schedule’s governing limit and prevent restoration for ordinary product use. Also test account termination, since deleting one encounter may follow a different path from ending the service.

The test cannot prove that an inaccessible backend copy is gone. It can expose mismatches between the documentation and the product, missing logs, undeletable transcripts or support procedures that depend on informal promises. Preserve the screenshots, exported logs, contract version and vendor response as the approval record.

On the exam-room phone, deployment should remain blocked until that record names an owner for settings, periodic subprocessor review, access-log checks and termination. If the vendor cannot state where the recording goes or what deletion does, manual documentation remains the workable fallback.

Questions people ask

Does signing a

BAA make an AI medical scribe HIPAA compliant?

No single document makes a deployment compliant. A BAA establishes contractual duties, while the clinic still has to configure access, limit permitted uses, train users, manage recording consent where applicable and verify the deployed workflow. The product tier and subcontractors covered by the agreement also need to match the service being used.

Can a medical scribe vendor train its model on patient data?

The answer depends on the signed terms, the permitted uses in the BAA and the data involved. Do not infer a ban from “HIPAA compliant” marketing. Require explicit language covering recordings, transcripts, prompts, outputs, feedback and de-identified derivatives, including whether training is opt-in, disabled by default or unavailable for the contracted account.

Is deleting the audio enough?

Usually not for a full deletion review. The vendor may still hold a transcript, generated note, metadata, audit records, support attachments or backup copies. Treat each as a separate record with its own retention rule, then verify what the user interface deletes and what expires later under the vendor’s backup schedule.

What should a clinic do if the vendor will not disclose its subprocessors?

Do not let the scribe receive patient information until the clinic can evaluate the downstream data path and obtain appropriate contractual coverage. Keep the exam-room phone disabled for recording, use the existing EHR template or manual dictation, and record the missing disclosure as an unresolved approval condition rather than accepting a sales assurance.

ShareFacebook
privacy and data rightsai governanceai medical scribeshipaabusiness associate agreementshealth data privacyvendor audits

One story a day

The story of the day, in your inbox

One real story about AI each morning — no hype, no alarm, just company for the road.

Read next

Laptop displaying a cropped airport image beside metadata fields and a Content Credentials verification panel.

AI Governance & Ethics

What an AI-Generated Image Label Can Actually Prove

A visible badge, file metadata, generation log, and signed Content Credential answer different questions. Cropping and reposting expose the gaps between them.

Irene Vasko · 8 min read

A support chat labeled Automated assistant beside a phone displaying an incoming customer-service callback.

AI Governance & Ethics

When a Customer-Service Bot Has to Say It Is a Bot

There is no blanket U.S. disclosure rule. A practical answer depends on where the customer is, what the bot is doing, and whether chat becomes an AI-generated call.

Irene Vasko · 8 min read

A laptop displaying a hiring bias-audit table beside a printed job notice and handwritten calculation notes.

AI Governance & Ethics

How to Read NYC’s Hiring-AI Bias Audit Before You Apply

A public audit can reveal which hiring system was tested, whose outcomes were counted, and where selection rates diverged. It can also conceal job-level differences and omit demographic groups.

Irene Vasko · 8 min read