AI Scribe Discovery Checklist for Medical Malpractice Cases
Roughly 42% of US medical groups now run an ambient AI scribe (DAX, Abridge, Suki, Augmedix, etc.) during patient encounters. The AI generates a draft note that the physician edits before signing — leaving a "dual record" most plaintiff attorneys aren't yet asking for in discovery. This checklist walks you through finding it.
Before you ask whether an AI scribe was used, scope who and when.
Ambient AI scribes started rolling out at scale in 2023 and crossed the 40% adoption threshold in early 2026 (Elion Health / KLAS data). Care delivered before mid-2023 is unlikely to involve an AI scribe; care delivered after January 2024 at any mid-to-large health system should be treated as presumptively scribed until disclaimed.
Tasks
Note: Telemedicine and tele-consult encounters are disproportionately scribed because the ambient audio capture is already part of the videoconference stack. Don't skip those notes.
Step 2 — Determine if the practice or health system uses an AI scribe
Five vendors dominate the US ambient-scribe market. Check each one before discovery to scope your asks.
Before serving discovery, do a free public-records pass to narrow which vendor is in play. This makes interrogatory drafting precise and signals to defense counsel that you know what you're asking for.
Top 5 AI Scribe Vendors (as of April 2026)
Vendor
Parent
EHR Integrations
Audio Retained?
Draft Retrievable?
Subpoena Target
Abridge
Abridge AI, Inc. (Pittsburgh, PA)
Epic (deep), Oracle Health, athenahealth
Per-tenant configurable; audio is processed and typically discarded post-generation, but enterprise contracts can elect retention
Yes — versioned drafts surface in Epic's note history when the integration is enabled
Abridge AI, Inc. — Registered Agent on file with PA Dept. of State; [email protected]
Nuance / Microsoft (Dragon Medical One + DAX classic)
Microsoft / Nuance Communications
Universal (sits on top of any EHR via dictation)
Dragon Medical One does not record ambient audio; the older DAX (pre-Copilot) used a human "scribe-in-the-loop" model and retained audio for QA — check contract date
Pre-Copilot DAX produced a written transcript reviewed by a human scribe; transcripts are retained by Nuance per the customer's BAA
Same as Microsoft DAX Copilot above
Augmedix
Augmedix, Inc. (San Francisco, CA) — acquired by Commure (2024)
Epic, Oracle Health, Meditech, athenahealth
Yes — Augmedix's hybrid model uses remote human scribes who hear live audio; audio is streamed and policy varies by customer (often not retained, but transcript is)
Yes — Augmedix delivers a structured note to the EHR and retains its own copy in the Augmedix platform for audit
Caveat: Vendor retention policies are contractual — the BAA between the covered entity and the vendor governs. Two hospitals using the same vendor can have opposite retention answers. Step 3 interrogatories ask for the BAA itself.
Public-records pass
If the public pass shows a vendor: draft Step 3 narrowly. If it shows nothing: draft Step 3 broadly to compel disclosure of any vendor used.
Step 3 — Send specific discovery requests
Paste-ready interrogatories and requests for production. Edit caption + plaintiff name before serving.
Important: These are drafts for attorney review, not legal advice. Read every word against your jurisdiction's rules and your case theory before serving. Have your discovery-rules attorney sign off.
Interrogatory 1 — Identify the AI software
Interrogatory
INTERROGATORY NO. ___: Identify by manufacturer, brand, product name, model, and version (including any subscription tier or feature flag) any artificial intelligence, ambient documentation, voice-recognition, generative-AI, large language model, or "AI scribe" software, application, or service that was used, accessed, or permitted to assist in creating, drafting, transcribing, summarizing, or proposing any portion of any medical record, progress note, History & Physical, consultation note, operative report, discharge summary, after-visit summary, telephone or telemedicine encounter, or other documentation relating to the care of [PLAINTIFF NAME] from [START DATE] through [END DATE], at any facility owned, operated, or controlled by [DEFENDANT].
Interrogatory 2 — Identify the people
Interrogatory
INTERROGATORY NO. ___: For each AI scribe, ambient documentation, or generative-AI tool identified in response to Interrogatory No. ___, identify (a) the vendor's full legal name and registered agent for service of process; (b) the date the tool was first deployed at the facility(ies) where Plaintiff received care; (c) every clinician, staff member, or third party (including human-in-the-loop scribes) who used the tool to document any portion of Plaintiff's care; and (d) the custodian within Defendant's organization who is responsible for the tool's contracts, configuration, and audit logs.
Interrogatory 3 — Audio recording
Interrogatory
INTERROGATORY NO. ___: State whether any audio recording, voice capture, or acoustic data of any encounter (in-person, telephonic, or video) between Plaintiff and any provider was made, processed, transmitted, stored, or retained — by Defendant, by any vendor, or by any subcontractor — at any time. If yes, identify (a) the date and duration of each recording; (b) the location where the recording was made; (c) every entity that received or stored a copy; (d) the current retention status of each recording (retained, deleted, overwritten, encrypted, archived, etc.) and the date of any deletion; and (e) the policy or contract governing retention.
Interrogatory 4 — Drafts and revisions
Interrogatory
INTERROGATORY NO. ___: For each provider note, summary, or other documentation that any AI tool drafted, proposed, or contributed to, state whether the AI-generated draft was preserved separately from the final, signed version of the documentation. If preserved, describe (a) where the draft is stored (EHR, vendor platform, third-party system); (b) the format; (c) the retention period; and (d) the steps required to retrieve it. If not preserved, state when, how, and by whom the draft was deleted, overwritten, or otherwise rendered unavailable.
Request for Production 1 — The drafts themselves
Request for Production
REQUEST FOR PRODUCTION NO. ___: Produce all draft, intermediate, pre-edit, pre-signature, and proposed versions of any provider note, History & Physical, progress note, consultation note, operative report, discharge summary, after-visit summary, or other documentation generated, drafted, suggested, or proposed by any AI scribe, ambient documentation tool, generative-AI tool, large language model, or voice-recognition system in connection with Plaintiff's care, in the form maintained in the ordinary course of business, including (without limitation) any version saved, autosaved, cached, logged, or transmitted between Defendant and any vendor, regardless of whether the draft was ultimately signed, edited, deleted, or replaced.
Request for Production 2 — Audit logs and metadata
Request for Production
REQUEST FOR PRODUCTION NO. ___: Produce all audit logs, access logs, change logs, version histories, "Note Activity" reports, "Author Activity" reports, and metadata (including but not limited to Epic Note History, Epic Audit Trail, Cerner / Oracle Health AuditView reports, or equivalent) reflecting the creation, drafting, editing, signing, amendment, addendum, or deletion of any documentation relating to Plaintiff's care from [START DATE] through [END DATE], including all entries identifying any AI tool, ambient documentation tool, voice-recognition system, or third-party scribe service as a contributor, author, source, or drafting agent.
Request for Production 3 — Vendor contracts
Request for Production
REQUEST FOR PRODUCTION NO. ___: Produce all contracts, master service agreements, business associate agreements (BAAs), data processing agreements, statements of work, order forms, and amendments between Defendant (or any of its parent, affiliate, or sister entities) and any vendor of an AI scribe, ambient documentation, generative-AI, or voice-recognition product that was active at any time during Plaintiff's care, together with all written policies, training materials, and standard operating procedures governing clinician use, configuration (including retention settings for audio and drafts), audit, supervision, and quality assurance of such products.
Tip: Serve these along with — not instead of — your standard medical-records and EHR audit-trail requests. The point is to add a parallel discovery track that captures what the EHR audit trail won't.
Step 4 — Subpoena the vendor directly if needed
If the defendant's responses to Step 3 are evasive, the vendor itself is a third-party with relevant records.
Most AI scribe vendors are non-parties. They will respond to a properly served subpoena duces tecum. Some vendors also offer a "law enforcement / legal request portal" that handles civil subpoenas (Microsoft and Augmedix/Commure both publish such portals).
Vendor records you typically cannot get from the hospital but can get from the vendor:
The original audio recording (if retained) — typically held by the vendor in cloud storage even when the hospital's local copy was purged
The full uncondensed transcript (most scribes generate a verbose internal transcript and a summarized note; only the note ends up in the EHR)
The structured "AI confidence scores" or hallucination-risk flags some vendors generate per sentence
The vendor's audit log of the draft → final-edit diff (separate from the EHR's audit log)
Internal ticketing / incident records if the physician reported an inaccurate draft
Vendor Subpoena — Paste-Ready Schedule A
Subpoena Duces Tecum — Schedule A
SCHEDULE A — DOCUMENTS AND THINGS TO BE PRODUCED
For the period [START DATE] through [END DATE], produce the following items relating to encounters between [PLAINTIFF NAME] (DOB [____]) and any clinician affiliated with [DEFENDANT HEALTH SYSTEM] (Customer Tenant / Account ID [____ if known]):
1. All audio recordings, voice captures, or acoustic data captured, processed, or stored by your platform in connection with the above encounters, in their native format, together with the encryption keys necessary to play them.
2. All transcripts of those encounters in their longest and most complete form, including any pre-summarization or "raw" transcript that was used as input to your note-generation models.
3. All AI-generated draft notes, summaries, problem lists, assessments, or plans that your platform generated, proposed, or transmitted to any user or to any electronic health record system in connection with the above encounters, including all intermediate and revised versions and any version that was overwritten or replaced.
4. All audit logs, version histories, and metadata reflecting the creation, editing, signing, transmission, deletion, or expiration of the items in paragraphs 1-3, including timestamps, user IDs, and IP addresses where available.
5. All confidence scores, uncertainty flags, hallucination-risk indicators, citation flags, or quality-assurance scores generated by your platform with respect to the items in paragraphs 1-3.
6. All correspondence (including support tickets, customer-success notes, and incident reports) between your company and [DEFENDANT] or any of its clinicians, employees, or agents that references the above encounters, [PLAINTIFF NAME], or any inaccuracy, hallucination, omission, or correction in the AI-generated documentation of those encounters.
7. The Business Associate Agreement, master service agreement, order form(s), and any amendments thereto in effect between your company and [DEFENDANT] (or any parent, affiliate, or sister entity) at any time during the relevant period, together with the data-retention policy applicable to that customer tenant.
8. All policies, procedures, training materials, and product documentation in effect during the relevant period describing (a) your retention of audio and draft notes; (b) the criteria your platform uses to flag low-confidence transcription or hallucination risk; and (c) procedures for customer requests to retrieve or preserve audio, transcripts, or drafts.
Produce in native digital format on encrypted media or via secure transfer. Produce a custodian declaration sufficient to authenticate the records under FRE 803(6) / 902(11) / 902(13).
Service tip: For Microsoft DAX Copilot, route the subpoena through Microsoft's Law Enforcement Compliance team. For Abridge, Suki, and Augmedix/Commure, serve the registered agent listed in Step 2 and email legal@ as a courtesy. Expect 30-60 day response time and meet-and-confer on scope.
Step 5 — Compare scribe-output draft vs. final note in the EHR
This is the "dual record" problem the ABA and at least one filed lawsuit have flagged. It can show what the physician actually heard, observed, or planned — before edits.
Once you have the AI draft from the vendor (Step 4) and the signed final note from the EHR (standard records request), run a side-by-side comparison. Document every material difference. The pattern that wins cases is when something clinically significant appears in the draft but is missing from the signed note.
Tasks
Spoliation theory: If the vendor's BAA shows the AI draft was retrievable for 30 / 60 / 90 days and the defendant didn't preserve it after the litigation hold attached, that's a candidate for an adverse-inference motion. The ABA Standing Committee's commentary on Formal Op. 512 explicitly raised this risk for lawyers; an analogous duty arguably attaches to a healthcare defendant once foreseeability of suit is established.
Critical edge case: Some EHR / scribe configurations auto-purge the draft within 24 hours of the physician signing. If your case involves a defendant that uses such a configuration, the draft is gone before any litigation hold can attach. The remaining evidence is the audit-log entry showing a draft existed plus circumstantial proof of the configuration. This is precisely the spoliation theory plaintiff bar will need to develop in 2026-2027.
Step 6 — Look for "AI artifacts" in the dictated note
Even when the draft itself is gone, the signed note often carries telltale phrasing that signals AI origin. These are flags, not proof.
Ambient AI scribes have characteristic linguistic fingerprints. None of these are dispositive on their own — but a cluster of them in a note from a 2024+ encounter is a strong tell.
Phrase patterns that signal AI origin
Boilerplate openers like "The patient is a [age]-year-old [gender] with a past medical history significant for..." appearing identically across notes for different patients by the same provider.
"The patient denies any..." followed by a comprehensive, alphabetical list of symptoms — humans don't dictate ROS in alphabetical order; LLMs do.
"As we discussed..." or "We had a detailed conversation about..." in the assessment, as a connective filler.
Identical assessment-and-plan templates across unrelated patients (cluster the provider's notes — AI scribes regress toward template).
Plural attribution where the speaker is singular ("We feel that..." when the dictating physician was alone in the room).
Hedged objective findings ("appears comfortable, in no apparent distress") on a patient the chart elsewhere describes as in 8/10 pain.
Structural artifacts
SOAP / APSO formatting that's perfectly consistent across the provider's notes — physicians' freehand structure varies; AI structure doesn't.
Bulleted plans with even capitalization and parallel grammar.
Patient quotations rendered in formal grammar ("The patient stated, 'I have been experiencing...'") — humans don't speak in those clauses.
Procedure or imaging recommendations phrased as suggestions ("We may consider...") rather than the imperative dictation style most physicians use.
Internal contradiction between the HPI and the plan that a physician proofreading would catch — AI scribes don't proofread, and rushed physicians often sign anyway.
Cross-checks
Use this for cross, not for filing: AI artifacts establish a question, not an answer. Deploy them in deposition to ask the provider directly: "Doctor, did an AI scribe draft this note?" The answer — and the body language — are the evidence.
Caselaw, opinions, and commentary
What's been filed, what regulators have said, what's still unsettled.
Sharp Rees-Stealy AI Scribe Litigation (S.D. Cal., filed Jan. 2026)
Putative class action — patient plaintiffs alleging that the health system's deployment of an ambient AI scribe produced systematically inaccurate notes diverging from what was discussed in the encounter, with discoverable drafts allegedly purged or overwritten.
The case is in early pleadings as of April 2026 and has not produced merits rulings. It is the highest-profile filed pleading raising the dual-record / spoliation theory in the AI-scribe context. Verify caption, docket, and current status before citing.
ABA Formal Opinion 512 (Generative AI Tools)
American Bar Association Standing Committee on Ethics and Professional Responsibility (2024)
Op. 512 addresses lawyers' duties when using generative AI, but the Standing Committee's commentary explicitly anticipates that the same dual-record / spoliation problem will arise on the healthcare side and that opposing counsel should plan to discover AI-tool inputs and outputs separately from the final work product. While the opinion does not bind healthcare defendants, plaintiff briefs cite it for the persuasive proposition that AI drafts are discoverable artifacts in their own right.
HHS / OCR HIPAA guidance on AI documentation (ongoing)
U.S. Department of Health and Human Services, Office for Civil Rights — guidance documents 2023-2025
OCR has confirmed that AI scribe vendors are HIPAA business associates and that the audio, transcripts, and drafts they generate are PHI subject to the BAA's retention and access provisions. This is the foundation for asserting that those records exist within the covered entity's HIPAA "designated record set" and are accessible via a HIPAA right-of-access request as well as via discovery.
State medical board guidance (varies)
California, Texas, Florida, New York medical board opinions — 2024-2026
Several state medical boards have issued informal opinions or position statements requiring physicians to review and attest to AI-generated documentation before signing. These opinions provide a standard-of-care benchmark for cross-examining a physician who signed an AI draft without meaningful review. Verify your jurisdiction's current guidance directly with the relevant board.
Independent verification required: All caselaw and regulatory references on this page are pointers for further research, not legal authority. Confirm dockets, opinions, and status with primary sources before relying on them in a brief. Healthcare-AI law is moving fast.
Background — why this checklist exists
How big is the AI scribe market today?
As of early 2026, industry surveys (Elion Health, KLAS, AMA member surveys) consistently put US ambient-AI scribe adoption among medical groups at roughly 42%, up from under 10% at the end of 2023. Adoption is concentrated in primary care, emergency medicine, and high-volume specialty clinics; it is rising fastest in mid-sized health systems that signed enterprise contracts with Abridge, DAX Copilot, or Suki in 2024-2025.
What is the "dual record" problem?
An ambient AI scribe captures the patient encounter (audio, transcript, structured note draft), produces a draft note, and presents that draft to the physician for review. The physician edits and signs. The signed version goes into the EHR; the draft frequently does not. The result is two records of the same encounter — a contemporaneous AI-generated record and an edited human-signed record — that may diverge on clinically material facts.
Why does the plaintiff bar care?
In a malpractice case, the signed note is what the defendant will produce. But the AI draft (when it exists and is retained) is a contemporaneous, less-edited record of the same encounter. It can corroborate or contradict the patient's account, expose what the physician chose to omit on edit, and surface failures in the standard of care that the final note papered over.
Aren't there privacy issues with subpoenaing audio of patient encounters?
The patient is the plaintiff and has typically waived privacy as to records of their own care for the purposes of the litigation. A protective order is standard. The vendor's privacy policy generally permits disclosure under valid legal process. The harder question is incidental privacy of other patients whose voices may have been picked up — that's typically handled with redaction.