Does an AI dental scribe store recordings of my patient visits?

It depends on the product, and there are three possibilities: the audio is uploaded and kept indefinitely, uploaded and deleted on a schedule, or never written to storage at all. A marketing page may not say which one applies. Ask directly whether audio is written to disk or object storage at any point, and get the answer in writing before you sign.

Is a recording of a dental visit considered PHI?

Yes. A recording of an operatory conversation contains a patient's voice plus clinical and often financial detail, which makes it protected health information under HIPAA. That means it belongs in your BAA, counts toward breach notification if it leaks, and can be reachable by a valid subpoena or records request.

What retention period should I accept for scribe audio?

Zero is best; if audio must be stored, days rather than months is a reasonable ask, and the number should be configurable per practice. Also ask about the backup horizon, because a 7-day deletion policy sitting on 90-day backups is really 90-day retention. A vendor who answers with "as long as necessary to provide the service" has not answered.

Can my patients' audio be used to train the vendor's AI models?

Possibly — and the permission may live in the terms of service rather than in the BAA. Ask whether audio or transcripts are used to train or fine-tune any model, including a subprocessor's, and require the no-training commitment in the signed BAA or data processing addendum. An emailed reassurance from a sales rep is not enforceable.

How is ambient scribing different from dictation for privacy purposes?

Ambient scribing captures the whole operatory conversation, including the patient, family members, staff, and anything audible nearby, which raises consent questions in all-party consent states. Dictation captures only the provider speaking to the machine. The capture surfaces differ, so a single retention answer covering both paths is worth pushing on.

Does Your AI Dental Scribe Keep the Recording of the Visit?

Q: Does Intake.Dental store dictation audio?

No. On the dictation path, audio is held in memory only and is never written to disk, S3, or a database. Only the transcript text is retained, for 24 hours, so an interrupted recording can be resumed or a finished dictation continued within that window. Separate opt-in, consent-gated recording features do upload audio to S3, which is why the claim is scoped to dictation.

Somewhere in the first ten minutes of every AI scribe demo, a dentist asks the only question that really matters: does this thing keep a recording of what happens in my operatory?

If the answer comes back as some version of “we’re fully HIPAA compliant, everything is encrypted in transit and at rest,” that is not an answer. It’s a response shaped like one. “Encrypted at rest” only means something if the audio is at rest — which is to say, stored.

There are three architectures in this category. Once you can name them, the vendor conversation gets much shorter.

The three architectures

1. Audio uploaded and retained indefinitely. The recording goes to the vendor’s object storage and stays there. The stated reasons are usually “model improvement,” “quality review,” or “so you can go back and listen.” Retention is described in adjectives, not days. Ask the hardest questions here — particularly about whether your patients’ audio trains anyone’s model.

2. Audio uploaded, transcribed, then deleted on a schedule. Better. The recording lands in storage, gets processed, and a retention job removes it after some window — 7 days, 30 days, 90 days. Two follow-ups decide whether this is real: is the deletion window configurable per practice, and does it apply to backups? A 7-day deletion policy sitting on top of 90-day backup snapshots is a 90-day retention policy wearing a costume.

3. Audio never stored. The stream goes to transcription and is discarded; only text survives. Nothing is written to disk, S3, or a database, so there is no recording to retain, subpoena, breach, or delete. This is technically harder — you can’t re-run a transcription you no longer have — but it removes an entire class of problem.

Whichever architecture you’re looking at, ask about transcript text separately. “We don’t store audio” and “we don’t store anything” are different claims, and the second one is rarely what’s on offer.

Why the recording is a heavier object than the note

The chart note is a curated clinical record. The recording is not. It captures the side conversation about the patient’s divorce, the treatment coordinator’s financing pitch through the wall, your assistant’s opinion of the last patient, and the offhand remark you’d never put in a chart.

Four consequences follow.

It’s PHI. Voice plus clinical content plus identifiers. It belongs in the BAA and counts in breach math.

It’s discoverable. If a case is filed, the note is the record — but the recording is a second, uncurated record that can be requested and can contradict the note. Your documentation quality is then judged against an unedited transcript of the room.

It expands breach surface. Every retained hour of audio is inventory someone else is holding on your behalf. Breach notification obligations scale with records, not with how sorry the vendor is.

It usually isn’t needed. Once the note is written, reviewed, and filed to the chart, the audio’s clinical job is done. Record retention rules are written around the chart; read your own state’s language rather than assuming it reaches raw capture. Storing the audio means carrying obligation with no matching benefit.

The questions to ask, and what a straight answer sounds like

Ask these verbatim. Write down the answers.

“Is the audio written to disk or object storage at any point, or is it held in memory only?” Straight: “It’s streamed to the transcription service in memory; nothing is written to S3 or the database.” Dodge: “It’s encrypted at rest.” That sentence concedes storage while sounding like a denial.

“If it’s stored, what is the retention window in days — and does that number apply to backups?” Straight: a number, plus a backup horizon. Dodge: “as long as necessary to provide the service.”

“Does any human ever listen? QA sampling, low-confidence review queues, support escalation?” Straight: “No,” or “Yes, under these conditions, and it’s logged in an audit trail you can pull.” Dodge: silence about the review queue.

“Is audio or transcript text used to train or fine-tune any model, yours or a subprocessor’s? Is that on by default?” Get the answer in the signed BAA or DPA. Not in an email from sales.

“Who are your subprocessors, and what does each one retain?” The scribe vendor may not be the party holding the audio — an upstream speech-recognition provider can be, under its own terms.

“Does that answer cover every audio feature in the product, or only the scribe path?” This one is worth asking because products accumulate audio features: dictation, ambient capture, recorded consults, call recording. Those paths can have different storage behavior, and a single answer may describe only one of them.

Ambient and dictation are not the same question

Ambient scribing listens to the whole visit. That means the patient, the parent in the corner, the hygienist next door, and whatever else is audible — which is why consent, and the all-party consent statutes in states such as California, Florida, and Pennsylvania, come up quickly. Dictation captures only the provider talking to the machine. Much narrower capture, much narrower exposure.

Different paths deserve different retention answers. A single blanket claim covering both is worth a follow-up question.

Accuracy is a separate question from retention

Worth pairing with the retention conversation: a JMIR Medical Informatics study reviewed 7,545 AI-generated clinical notes and evaluated 356 of them in detail, finding omissions in 18%, hallucinations in 11.5%, and accidental inclusions in 9.3%. That is the argument for provider review before signing, whatever the vendor’s audio policy is.

What ours does, stated with the precision we’re asking for

On the dictation path, our clinical notes product never stores audio. It’s held in memory and never written to disk, S3, or a database. Only transcript text persists, and only for 24 hours — long enough that an interrupted recording can be resumed, or a finished dictation continued, within that window. After that, the note in the chart is the record.

Now the part that matters more than the claim: this is scoped to dictation. Separate opt-in, consent-gated recording features do upload audio to S3. Writing “we never store audio” full stop would be the exact sloppiness this page exists to help you catch.

PHI is encrypted at the field level at rest, and we sign a BAA. The security page is there if you want to put us through the same questions. (Disclosure lives in the footer: same owners as this site.)

Two related specifics, since they get vague treatment elsewhere. The approved note writes into the patient’s chart through the PMS API, dated to the visit — on Dentrix Ascend it files against a real provider field; on Open Dental, a commlog has no provider field at all, so the provider’s name is written onto the note itself. And signed notes are never rewritten; corrections file as dated addenda.

Trust signals that don’t come from the vendor

Two are worth checking before the demo.

Open Dental publishes a third-party vendor list with explicit trust tiers. At last count: Authorized (87), Authorized In Transition (11), Dangerous/Unknown (22), and Not Currently Active (11), alongside 14 Native Integrations. Dangerous/Unknown covers direct database access or unwhitelisted plugins. Listing is free — vendors request it through vendor.relations@opendental.com — so ask a vendor who isn’t on the list why not. A vendor willing to reach into your production database directly is telling you how it thinks about data handling generally.

Open Dental isn’t the only one doing this. Henry Schein One’s API Exchange also publishes a public vendor list, and authorization there requires SOC 2 Type II and OAuth 2.0. That’s a floor, not a guarantee — but it’s a floor somebody else enforces. Check whichever list matches the system you run.

The ten-minute version

Ask whether audio hits storage. Ask for a retention number in days, including backups. Ask whether humans listen. Ask whether your data trains models, and require the answer in the BAA. Ask whether the answer covers every audio path in the product. If most of those come back in adjectives, you have learned what you needed to learn.