AI Medical Scribes and HIPAA: A 2026 Buyer's Framework

The AI medical scribe category went from three serious vendors in early 2024 to twelve by mid-2026. Every clinician demo turns into "this is incredible, when can we start using it?" Every compliance officer evaluation turns into "wait, what about HIPAA?" Both reactions are reasonable. Neither produces a buying decision.

This guide is the framework for an actual buying decision. The 10 questions below distinguish vendor-side compliance from practice-side configuration, and they identify the four failure modes that produce most OCR findings in early scribe deployments.

The Vendors in the Category

The major AI medical scribes deployed in U.S. healthcare practices in 2026:

Abridge — dominant deployment in academic medical centers (Mayo Clinic confirmed enterprise-wide; UNC Health, Emory, KUMC). BAA available. SOC 2 Type II. Mobile and EHR-embedded. See Is Abridge HIPAA Compliant? for the dedicated review.

Microsoft DAX Copilot (formerly Dragon Ambient eXperience) — strong in Epic-integrated workflows because of the Microsoft-Epic partnership depth. BAA via Microsoft. Available standalone or embedded.

Suki AI — clinician-friendly UX, growing in independent practice market. BAA available. Distinct embedded scribe + conversation summarizer split.

Heidi Health — rising vendor with strong mobile-first workflow. BAA available. Common in Australian healthcare adoption that's now spilling into the U.S. independent practice market.

Nuance Dragon Medical One — the incumbent dictation platform (now Microsoft after the 2022 acquisition). Includes ambient capability. BAA via Microsoft.

DeepScribe — independent vendor focused on outpatient ambulatory care. BAA available.

Augmedix — clinical-documentation-as-a-service model with human-in-the-loop. BAA available. Different model than pure-AI scribes — humans review/finalize the notes.

AWS HealthScribe — Amazon's medical scribe API, eligible since 2023 under AWS BAA. Building block more than turnkey product, but increasingly used by EHR vendors and clinical platforms.

The category is consolidating. By 2027, expect 5-7 of these to survive as standalone products. The rest will be acquired or pivot to embedded white-label deployments inside EHRs.

The 10-Question Framework

Apply these to any AI scribe under consideration. The vendor's marketing answer to each may differ from the operational reality; the goal is to determine the reality.

Question 1: Does the vendor sign a BAA for the specific deployment tier you're using?

Most AI scribe vendors offer BAAs. Verification is per-deployment, not per-product. A vendor's enterprise tier may include a BAA; their free trial or individual clinician tier may not. Confirm that the BAA covers the exact configuration your practice is deploying.

This question catches the failure mode where a clinician demos the scribe on personal patients before the practice has executed the BAA — creating a documented HIPAA exposure before the formal deployment even starts.

Question 2: Where does the audio actually go?

Each vendor processes audio differently. Some stream to a real-time inference endpoint with no persistent storage. Some store recordings for model retraining (with customer opt-out). Some retain audio for X days for quality assurance.

For HIPAA scope, the question is whether the audio is retained, in what region, under what controls, and for how long. The BAA covers the vendor's handling — but you should know what that handling is in practice. The vendor's data-handling addendum is often a separate document from the BAA.

Question 3: Does the platform support state two-party recording consent?

HIPAA does not address recording consent. State law does. Roughly a dozen states (commonly cited: CA, CT, DE, FL, IL, MD, MA, MT, NH, OR, PA, WA — with several others having nuanced statutes) require all-party consent for recording. In all-party-consent states, your patient must affirmatively consent to the AI scribe being on — at every encounter, not once on the intake form.

Some scribe vendors build the consent prompt into the workflow. Some leave it to the practice to script. The vendors that build it in have a meaningful workflow advantage in two-party states.

Question 4: What happens with bystander PHI?

When a patient brings a family member, translator, or chaperone, the microphone captures everyone in the room. If a bystander discloses health information about themselves — their medication, a family-history diagnosis — that information now lives in the patient's chart with the wrong consent.

No vendor solves this entirely at the technology layer. The mitigation is procedural — clinicians need to pause the recording when bystander PHI is being discussed, and the consent script needs to explicitly cover everyone in the room. The vendor's contribution is making the pause and resume controls easily accessible.

Question 5: How does the generated note enter the EHR?

The most common downstream-routing failure: the scribe generates a note, the clinician reviews it, then copies it into the EHR. During the copy step, the note may pass through OneNote, a personal email draft, or a clipboard manager that doesn't have a BAA. The scribe vendor has zero visibility into this.

Vendors with deep EHR integration (DAX with Epic, Suki with athenahealth, Abridge with multiple EHRs) close this gap by routing the note directly. Vendors that rely on copy-paste do not.

Question 6: Who can access the recordings later?

Recordings — when retained — are PHI. The platform's access control determines who in the practice can play them back. Practices have been surprised in audit to discover that all workforce members with platform access could play back any recording, not just their own.

Verify the access model: per-clinician, per-encounter, with audit logs. Default to most-restrictive.

Question 7: How long are recordings and notes retained — and where?

HIPAA requires the practice to retain documentation for six years; some state retention laws require longer. The vendor may retain recordings for a shorter period (90 days is common for quality assurance) or indefinitely (for ongoing model training, with opt-out).

The two retention periods don't have to match — the vendor's QA retention can be 90 days while the EHR's clinical-record retention is the long-haul one. But the practice should know both and document the difference.

Question 8: Is the model trained on your data?

Most vendors offer opt-out from model training on customer data. Some default to opt-in. The choice has both HIPAA implications (data use for purposes beyond treatment, payment, operations may require additional authorization) and competitive implications (your specialty-specific workflows training a vendor's model that's then sold to your competition).

Verify the default and confirm your selection in writing.

Question 9: What happens during a vendor security incident?

The BAA requires the vendor to notify you of breaches affecting your PHI. The notification timeline, the format, and the supporting documentation differ by vendor. Some vendors are bureaucratic about this; some are responsive.

If the vendor has had a public security incident in the past 24 months, study how they handled it. The handling of past incidents predicts the handling of future ones.

Question 10: What's the off-ramp if the vendor fails or pivots?

The category will consolidate. Some vendors you select today will not be operating in their current form in 2027. The off-ramp question — can you export your accumulated data, in what format, with what notice — is part of the buying decision even if the vendor seems stable today.

Practices that ignored this question with the previous generation of healthcare SaaS vendors (telehealth platforms that pivoted, patient engagement tools that shut down) spent significant time and money extracting data under pressure. The AI scribe wave is going to produce the same pattern in 2027-2028.

The Four Failure Modes

Across early AI scribe deployments, four patterns produce most of the OCR findings:

Failure 1: BAA gap during evaluation. Clinician demos the scribe on real patients before the practice has executed the BAA. The exposure is logged before formal deployment.

Failure 2: State two-party consent ignored. Practice deploys in California or Illinois without updating the consent script. Every encounter is a state-law violation, regardless of HIPAA status.

Failure 3: Bystander PHI in patient records. A family member's health detail ends up in the patient's chart. The bystander never consented to that.

Failure 4: Downstream routing through unsecured channels. Generated notes pass through clipboard or non-BAA tools before reaching the EHR.

A practice that scores well on the 10 questions above and addresses these four failure modes in policy and training has the strongest defensible position the category supports. A practice that signs a BAA and assumes the rest is handled is exposed.

How Patient Protect Helps

Patient Protect tracks AI scribe deployments as a vendor category in your BAA inventory, integrates the consent workflow with your state-law profile, surfaces the integration BAA chain (EHR, clipboard tools, document storage destinations), and monitors the configuration drift that turns a compliant deployment into a non-compliant one over time.

For independent practices adopting an AI scribe — and the category is moving fast enough that "adopting" is the expected state for most practices by year-end — the platform turns a one-time vendor decision into an ongoing compliance program.

The vendor side of AI scribes is finally solid. The practice side is where the next wave of OCR enforcement is going to land.