AI Notetaker Transcripts: Discoverability, Litigation Hold, and Privilege

The question is no longer whether lawyers use AI notetakers. They do. The question is what happens when opposing counsel serves a discovery request that reads, in substance, produce all AI-generated transcripts, summaries, and recordings of meetings relating to this matter. If you have been running Otter, Fireflies, Zoom AI Companion, or a Teams Copilot summary across your client calls, the answer is uncomfortable: those artifacts are electronically stored information, they are almost certainly within your possession, custody, or control, and the vendor that generated them is a third party whose presence may already have broken privilege.

This article walks through the three doctrines that collide here — the scope of discoverable ESI, the duty to preserve on reasonable anticipation of litigation, and the attorney-client privilege and work-product protection — and explains what a defensible AI notetaker policy looks like in 2026.

AI Transcripts Are ESI. Full Stop.

Federal Rule of Civil Procedure 34 defines the scope of document requests to include any designated documents or electronically stored information — including writings, drawings, graphs, charts, photographs, sound recordings, images, and other data or data compilations — stored in any medium from which information can be obtained. A transcript generated by an AI notetaker is text. A summary is text. The underlying audio, if retained, is a sound recording. All three fit comfortably within Rule 34.

The state analogs follow the same pattern. New York CPLR 3101(a), California Code of Civil Procedure 2031.010, and the uniform rules adopted in most commercial jurisdictions all pull ESI into the discovery universe. There is no meaningful argument that an AI transcript is categorically outside the scope of discovery because it was machine-generated; the Federal Rules of Evidence have long treated computer output as admissible if authenticated, and the 2015 and 2017 amendments to Rule 37(e) presuppose that ESI generated by third-party tools is preservable and producible.

The practical consequence is that if a transcript exists on a vendor server, on your laptop, or in a shared cloud folder, it is fair game. The further consequence — the one firms are only now beginning to internalize — is that the transcript exists in places you did not necessarily authorize and may not be able to delete.

Possession, Custody, or Control Reaches the Vendor

Rule 34 requires production of ESI within a party's possession, custody, or control. Courts have interpreted control broadly to include the practical ability to obtain documents on demand, which is exactly the relationship most firms have with SaaS notetakers. If you can log into the Fireflies or Otter admin console and download the transcript, you control it. If the vendor's DPA gives you a right to export, you control it.

This matters for two reasons. First, you cannot credibly claim in a discovery response that transcripts are unavailable because they sit on a vendor server. Second, your duty to preserve attaches to those vendor-held copies the moment litigation is reasonably anticipated, which means your litigation hold has to reach beyond your own file system.

Litigation Hold Duties Attach Early, and They Attach to Transcripts

The duty to preserve is triggered by reasonable anticipation of litigation, not by the filing of a complaint. Zubulake v. UBS Warburg and its progeny established that counsel must issue a written hold, identify custodians and sources, and take affirmative steps to suspend routine deletion. For AI notetakers, this creates three concrete obligations:

• Suspend auto-delete. Most consumer-grade notetakers default to retention windows (Fireflies and others document auto-delete features that purge meetings on a schedule). Those rules have to be paused for custodians and matters within the hold.
• Identify vendor-held copies. The hold notice should list every notetaker in use, because a hold that covers Outlook but not Otter will not survive a Rule 37(e) motion if transcripts go missing.
• Preserve derivative artifacts. Summaries, chapter markers, action items, and shared workspace comments are all ESI. They are also often auto-generated after the fact and stored separately from the raw transcript.

Rule 37(e) sanctions for failure to preserve ESI run from curative measures to adverse inference instructions and, in cases of intent to deprive, case-terminating sanctions. In West Technology Group v. Sundstrom (D. Conn. 2024), the court analyzed spoliation in the context of cloud-based business communications and reinforced that modern custodians cannot ignore SaaS-held records when a hold is in place. The reasoning applies directly to AI notetaker artifacts.

Privilege: The Third-Party Problem

Attorney-client privilege protects confidential communications between lawyer and client made for the purpose of obtaining or giving legal advice. The confidentiality element is doctrinally fragile: the presence of a non-essential third party during the communication generally destroys the privilege, and voluntary disclosure to a third party after the fact waives it.

An AI notetaker that transmits audio to a vendor's servers for transcription is, structurally, a third party in the room. Whether that presence destroys privilege depends on two questions courts are now actively working through:

1. Is the vendor a privileged agent? Under United States v. Kovel, third parties whose presence is necessary to facilitate the legal advice (accountants, interpreters) can fall inside the privileged circle. The argument that a transcription vendor is a Kovel agent is untested and weak where the vendor independently uses the data for model training, analytics, or product improvement.
2. Did the client knowingly consent to third-party processing? Even where consent exists, courts analyzing cloud providers have drawn a line between pure storage (often protected) and active processing that generates new work product (more exposed).

The United States v. Heppner decision (S.D.N.Y. Feb. 2026, Judge Rakoff) sharpened this. The court held that a litigant's chats with a public AI platform were not privileged, reasoning by analogy to the third-party doctrine: voluntarily handing content to a service provider defeats the reasonable expectation of confidentiality that privilege demands. Heppner involved a party's own use of a chatbot, not a lawyer's notetaker, but the analytical move — treating the AI vendor as a third party outside the privileged relationship — maps directly onto SaaS transcription.

The ABA Formal Opinion 512 (July 2024) on generative AI and NYC Bar Formal Opinion 2025-6 (December 2025) both reinforce that lawyers must evaluate the confidentiality terms of AI tools under ABA Model Rule 1.6 before feeding them client information. The core move in both opinions: the lawyer is responsible for the vendor's handling of client data, and generic SaaS terms of service are rarely enough.

Work Product: Better Protected, But Not Automatic

Work-product doctrine under Rule 26(b)(3) protects materials prepared in anticipation of litigation. Transcripts and summaries of witness interviews, strategy sessions, and matter-specific meetings can qualify. Opinion work product — counsel's mental impressions, conclusions, opinions, or legal theories — enjoys near-absolute protection.

Two cautions. First, a raw transcript is often characterized as fact work product at best, which can be overcome by a showing of substantial need and undue hardship to obtain the equivalent by other means. Second, work-product protection, like privilege, can be waived by disclosure to adversaries or in a manner that substantially increases the risk of adversary access. Routing the recording through a vendor that may use it for model training, or storing it in a workspace other firm personnel can browse, implicates both concerns.

A Landscape View: What the Tools Actually Do

The vendor architecture matters because it determines what you have to preserve, what you have to log, and whether the vendor's presence is an evidentiary problem on its own.

Independent coverage of the category reflects the same split. See Mayer Brown's analysis of AI notetakers as emerging legal risk and the AI Policy Desk 2026 privacy comparison of Zoom, Otter, and Fireflies. The Brewer v. Otter.ai (2025) putative class action — alleging unlawful recording by a bot that joined calls — is a reminder that the vendor's conduct can create standalone liability apart from any discovery question.

Building a Defensible AI Notetaker Policy

Firms that want to keep using these tools without inviting a Rule 37(e) hearing should treat AI notetakers as a distinct records category inside the information governance program. The elements:

• Approved tool list. No notetaker that lacks a signed DPA, a no-training commitment for customer data, and documented retention controls should be used on client matters. For a deeper walkthrough of the vendor diligence questions, see our note on on-device alternatives for privileged meetings.
• Matter-level opt-in. Default-off for client communications. Opt-in should be logged — who authorized, when, for which meeting, and with what client consent.
• Consent on the record. Verbal consent captured at the top of the recording, or written consent in the engagement letter, addresses both wiretap-style statutes in two-party-consent jurisdictions and the voluntariness element of any later privilege argument.
• Retention that matches the matter. Short default retention windows, with hold overrides tied to the matter management system. See our detailed treatment of transcripts as discoverable ESI under litigation holds.
• Hold notices that name the tools. The hold template should enumerate every AI notetaker in the environment and instruct custodians to suspend auto-delete on each.
• Privilege labeling. Transcripts and summaries of legal-advice meetings should be marked Privileged and Confidential at generation, not later. Labels do not create privilege, but they support the claim and they make inadvertent-disclosure clawbacks under FRE 502 more defensible.

What a Hold Notice Should Say About AI Notetakers

A one-paragraph addendum, roughly: You are instructed to preserve all audio recordings, transcripts, summaries, action items, and derivative artifacts generated by any AI meeting assistant, including but not limited to [list]. Suspend any auto-delete or retention policy applicable to the matter custodians identified above. Do not delete, archive, or modify any such artifact without written approval of the general counsel's office. This duty extends to artifacts stored on vendor infrastructure that you have the practical ability to access.

That language is not magic, but it closes the gap most firms have today, which is a hold that assumes email and Teams chat and silently misses the Otter workspace.

How Basil approaches this

Basil for Law is built so the hardest problem in this article — the third party in the privileged room — does not arise by architecture. Transcription and summarization run on the device using the Apple Neural Engine. Basil has no server and no subprocessors, and Basil never receives your recordings, transcripts, or summaries. There is no vendor copy to preserve, produce, or subpoena from Basil, because Basil does not have one.

iCloud Sync is on by default so your library, audio included, is available across your devices. It is stored in your own iCloud account under Apple's terms, not Basil's, and you can turn it off in Settings › iCloud at any time. For in-person meetings and for Zoom, Teams, and Meet calls on macOS, Basil captures locally in Computer mode — no bot joins the call, which removes a category of consent and wiretap exposure that bot-based tools create.

The Basil for Law edition adds privilege attestation, a consent log, matter organization, and Privileged and Confidential labeling. Pricing for solos is $19.99 per month or $199.99 per year, with a 3-day free trial on monthly and 7-day on annual, one trial per Apple Account. The general Basil app is available today with a 60-minute monthly free tier. Basil signs DPAs and NDAs on request.

None of this eliminates the duty to think carefully about consent, retention, and hold compliance. It does mean the architectural question — is there a vendor transcript out there I have to find and preserve — has a cleaner answer when the notetaker never sent the audio anywhere in the first place.

This article is for information only and is not legal advice.

Keep client conversations on your device

Basil makes transcripts and summaries on your device. No bot joins the call, and Basil never receives your client conversations. See Basil for Law → · Legal-tool reviews →

This article is for information only and is not legal advice.