Ambient AI Scribes Explained: How They Work Without Recording Everything
Ambient AI scribes don't need to function like permanent surveillance tools, and most aren't designed to. A 2026 policy analysis in NEJM AI found that leading vendors handle audio wildly differently: Freed deletes it within 60 seconds of generating the note, while Microsoft Dragon Copilot can retain it up to 90 days. Nabla processes audio in chunks and never stores it at all. So "ambient" doesn't mean "always recording forever," it means listening in the background and converting speech into a note, with what happens to the raw audio afterward left almost entirely up to vendor design choices. But a 2026 lawsuit shows that even brief capture still raises real legal questions. Here's the actual pipeline, what gets kept, and why it matters.
- Ambient AI scribes create four distinct data types per visit: raw audio, an interim transcript, an unedited draft note, and the physician-approved final note. Each can be retained, shared, or used for training under different terms.
- Audio retention varies enormously by vendor: from immediate deletion (Freed, 60 seconds) to 90 days (Microsoft Dragon Copilot), with no industry-wide standard.
- HIPAA sets no minimum or maximum retention period for audio or transcripts. Only the final approved note is subject to Medicare's 5-year medical record retention requirement.
- A 2026 federal class action, Washington et al. v. Sutter Health, alleges that recording and transmitting patient conversations without meaningful informed consent violates California wiretap and privacy law, even for a HIPAA-compliant tool.
- Deleting audio quickly reduces long-term exposure but doesn't resolve the underlying consent question, since the capture itself, not just the storage, is what triggers wiretap and consent statutes.
The Real Pipeline: What Actually Happens to a Conversation
An ambient scribe doesn't produce one artifact from a visit, it produces four, and each is handled differently:
The key distinction: HIPAA sets no retention limit, minimum or maximum, for the audio or the interim transcript. Only the final signed note is treated as part of the medical record. That gap is exactly where vendor policy, not law, ends up deciding how long a sensitive recording actually survives.
How Long Vendors Actually Keep the Audio
A 2026 NEJM AI Policy Corner compared published retention practices across major vendors. The spread is wide:
| Vendor | Audio Retention |
|---|---|
| Freed | ~60 seconds |
| Nabla | Never retained (chunked processing) |
| Tali AI | Up to 48 hours |
| Sunoh.ai | 7 days |
| Microsoft Dragon Copilot | Up to 90 days |
| Abridge | Disclosed only via private contract |
Transcript and note retention tends to run longer and vary even more, from 14-day defaults at some vendors to "never delete" as the default at others. Most vendors that train models on this data restrict training to deidentified data, though deidentifying audio is technically much harder than deidentifying text, since the voice itself functions as a biometric identifier.
Retention is a choice, not a default you can assume. A health system that hasn't explicitly specified its retention preference is inheriting whatever the vendor defaults to, which in some cases is "never delete." Specifying collection, storage, access, retention, and deletion for each of the four data types, before signing a BAA, is the only way to actually control this.
Why "It Gets Deleted Fast" Doesn't End the Legal Question
In April 2026, three California patients filed a federal class action, Washington et al. v. Sutter Health et al. (Case No. 4:26-cv-03012, N.D. Cal.), against Sutter Health and MemorialCare over their use of Abridge's ambient AI scribe. The complaint alleges the health systems recorded and transmitted patient-clinician conversations to third-party systems for processing without meaningful, informed consent, violating California's Invasion of Privacy Act, Confidentiality of Medical Information Act, Unfair Competition Law, the federal Wiretap Act, and common-law invasion of privacy.
Notably, the vendor in question does sign Business Associate Agreements with its HIPAA-covered clients, so this isn't a HIPAA compliance failure. The lawsuit argues that HIPAA compliance and lawful recording consent are separate questions, and a tool can satisfy one while failing the other. A separate, earlier suit, Saucedo v. Sharp Healthcare, raised a nearly identical theory months before.
The core legal argument: capturing and transmitting a conversation is what triggers wiretap and two-party-consent statutes, independent of how long the resulting audio is later retained. A recording deleted in 60 seconds was still, for that 60 seconds, captured and processed, which several state laws treat as the legally significant act, not the storage duration that follows it.
What Health Systems Should Actually Verify
Based on the compliance guidance and the emerging case law, four things are worth confirming before adopting any ambient scribe. This matters more than usual here: AI-built healthcare apps carry 30-40% more code vulnerabilities than conventionally developed software, and the average healthcare data breach now costs $7.42 million.
- A signed Business Associate Agreement that explicitly states how long audio and transcripts are retained, whether the data trains the vendor's models, and where it's stored.
- Documented patient notice and consent built into intake workflows and your Notice of Privacy Practices, not left to individual clinicians to explain ad hoc.
- A retention and deletion policy you can actually audit, not just a vendor's marketing claim. The safest posture for most organizations is deleting raw audio once the note is finalized.
- Mandatory clinician review before sign-off, since draft notes can hallucinate findings or drop context, and an unverified AI draft should never auto-file into the record.
For the broader picture on how AI scribes work and what to check before adopting one, see our companion guide, What Is an AI Medical Scribe and How Does It Work? For the evidence on time savings and accuracy tradeoffs, see AI Scribe vs Manual Documentation: Time and Accuracy Compared. For the encryption side of this specifically, including the proposed 2026 HIPAA Security Rule update, see our guide to end-to-end encryption in clinical communication platforms.
Documentation That Doesn't Outsource the Privacy Decision
HosTalky's AI Scribe keeps clinicians in control of every note, with transparent, auditable data handling instead of a vendor default you have to go looking for.
See How AI Scribe WorksThe Bottom Line
"Ambient" describes where the listening happens, in the background, without dictation, not what happens to the recording afterward. That's a governance decision each vendor makes differently, and right now there's no industry norm forcing consistency. The most privacy-protective implementations discard audio within seconds and never build a persistent archive. Others treat 90 days of retention as reasonable, often to support model training or quality audits. Neither approach is inherently illegal, but as the Sutter Health case shows, the decision to capture a conversation at all still needs the patient's informed consent, no matter how quickly the recording is deleted afterward.
FAQs
Do ambient AI scribes record everything said during a visit?
What are the four data types an ambient AI scribe creates?
Is it legal for an ambient AI scribe to record a patient visit without consent?
Does deleting the audio recording eliminate the privacy risk?
Sources and References
- NEJM AI. (2026). Policy Corner: AI Scribe Data Retention and Governance. DOI: 10.1056/AIpc2501401
- Medcurity. (2026). Ambient AI Documentation and HIPAA: A 2026 Compliance Guide for Healthcare. medcurity.com
- HIPAA Journal. (2026). Lawsuit Alleges AI Platform Illegally Recorded Patient-Clinician Conversations. hipaajournal.com
- Paubox. (2026). How to Evaluate AI Clinical Scribes for HIPAA and Consent Risks. paubox.com
