FRCP 37(e), Spoliation, and AI Meeting Notes Retention

Every AI notetaker your firm uses is quietly generating a new category of electronically stored information (ESI): audio buffers, transcripts, summaries, action-item lists, and — in some products — vector embeddings and server-side logs. Under Federal Rule of Civil Procedure 37(e), the moment those artifacts are reasonably foreseeable evidence in litigation, your duty to preserve them attaches. The moment you delete them without a defensible policy, you have a spoliation problem.

The paradox is that keeping everything creates its own risk. Transcripts of client meetings can capture privileged strategy, half-formed theories, and offhand remarks that a hallucinating summarizer may render inaccurately. The right answer is not "save forever" or "delete immediately." It is a written, consistently applied retention schedule that satisfies Rule 37(e)'s two safe harbors: reasonable steps to preserve, and — where loss occurs — no intent to deprive.

This article walks through what Rule 37(e) actually requires, why AI meeting notes are ESI, how the case law is starting to treat AI-generated artifacts, and how to design a retention framework that survives a preservation-letter fight.

Why AI Meeting Notes Are ESI Under Rule 34

Rule 34 defines discoverable material broadly: "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." A transcript is a writing. An audio buffer is a sound recording. A vector embedding is a data compilation. All of it falls within Rule 34, and therefore within the preservation duty Rule 37(e) enforces.

That breadth matters because attorneys and clients sometimes assume ephemeral AI outputs — a summary shown on screen, an action-item list emailed once — are not "records." They are. If they exist on any device or server at the moment the duty to preserve attaches, they are subject to a litigation hold. The Sedona Conference's Sedona Principles, Third Edition treat this expansively and remain the leading practitioner reference.

What FRCP 37(e) Actually Says (and Does Not Say)

Rule 37(e), as amended in 2015, applies when: (1) ESI that should have been preserved is lost; (2) because a party failed to take reasonable steps to preserve it; and (3) the ESI cannot be restored or replaced through additional discovery. Only then does a court consider sanctions.

The rule creates a tiered remedy structure:

• Rule 37(e)(1) — prejudice-based measures. If loss prejudices another party, the court may order measures "no greater than necessary" to cure the prejudice.
• Rule 37(e)(2) — intent-based sanctions. Only on a finding that a party "acted with the intent to deprive another party of the information's use in the litigation" may a court impose an adverse-inference instruction, dismiss the action, or enter default judgment.

The Advisory Committee Notes are explicit that negligence — even gross negligence — is not enough for the severe sanctions in 37(e)(2). This is the doctrinal anchor for any retention policy: routine, good-faith deletion under a written schedule is not spoliation, even if the deleted material later turns out to be relevant, provided the duty to preserve had not yet attached or a hold had not been triggered.

When the Duty to Preserve Attaches

The trigger is not the filing of a complaint. It is when litigation is "reasonably anticipated." That standard, articulated in cases like Zubulake v. UBS Warburg and reinforced across the circuits, is fact-specific: a demand letter, a serious client complaint, a regulator's inquiry, or an internal investigation can all trigger the duty.

For AI meeting notes, three trigger patterns recur:

1. Transactional matters that go sideways. Deal negotiations captured by AI notes become evidence when a fraud or breach claim later emerges.
2. Employment matters. One-on-one meetings, performance conversations, and HR intake sessions are high-risk for retention disputes.
3. Internal investigations. The moment counsel opens an investigation, every AI-captured interview transcript is potentially privileged work product and potentially discoverable.

The New Wrinkle: AI Artifacts, Third-Party Platforms, and Waiver

Retention risk is not just about deletion. It is also about where the ESI lives. If your AI notetaker stores transcripts on a vendor's servers, three problems compound:

Third-party doctrine exposure. In United States v. Heppner (S.D.N.Y. Feb. 2026), Judge Rakoff held that a litigant's chats with a public AI platform were not privileged, drawing an analogy to the third-party doctrine. The reasoning transfers directly to AI meeting notes uploaded to a vendor cloud: disclosure to a third party — even a contractual one — can defeat the confidentiality prong of the privilege analysis unless the vendor relationship is carefully structured as an agent of the lawyer.

Vendor-side retention you do not control. A cloud-based notetaker may retain data on backup tapes, in analytics pipelines, in model-training corpora, or in disaster-recovery snapshots long after you "delete" it in the UI. When a preservation letter arrives, you now have to subpoena your own vendor. The ABA Formal Opinion 512 (July 2024) on generative AI tools makes clear that lawyers must understand where client data travels and how long it persists.

Independent litigation against the vendor. Brewer v. Otter.ai (N.D. Cal. 2025) is a live putative class action alleging that a leading AI notetaker recorded conversations without adequate consent. Firms whose transcripts sit on a defendant vendor's servers may find their client data pulled into third-party discovery they never contemplated.

The NYC Bar Formal Opinion 2025-6 (December 2025) reinforces the point for New York practitioners: consent, confidentiality, and vendor architecture are not optional considerations.

What Sanctions Actually Look Like

The severity spectrum under 37(e) matters because the record shows courts are willing to use the full range when the facts warrant it.

Cases like West Technology Group v. Sundstrom (D. Conn. 2024) illustrate how courts scrutinize the sequence of events between a preservation trigger and the loss of ESI. The question is rarely "did the deletion happen?" It is "what did the party know, and when did they know it?"

Designing a Defensible Retention Policy for AI Meeting Notes

A defensible policy answers six questions in writing:

1. What is captured? Audio, transcript, summary, embeddings, metadata. Enumerate each artifact class.
2. Where does it live? Device, firm server, vendor cloud, backup. Data-flow diagrams protect you in a Rule 26(f) conference.
3. How long is it kept by default? A time-based schedule keyed to matter type. Litigation matters often warrant longer; routine intake shorter.
4. Who can delete, and when? Role-based deletion authority with an audit log.
5. How is a hold implemented? The mechanism to suspend deletion immediately upon a trigger event.
6. How is the policy audited? Periodic review, ideally by someone outside the practice group that generated the notes.

The Advisory Committee Notes to Rule 37(e) explicitly reward this kind of documented, good-faith approach. What courts punish is inconsistency: a firm that keeps transcripts "sometimes" and deletes them "sometimes" has no policy to defend.

Retention Schedules by Matter Type

There is no single correct number. But a sensible starting framework looks like this:

These are starting points, not prescriptions. State bar rules, industry-specific regulations (HIPAA, FINRA, SEC 17a-4), and client-side retention obligations will move the numbers. The point is to have the numbers in writing.

The Litigation Hold Mechanic

When a hold triggers, three things must happen within hours, not days:

• Suspend automated deletion for all custodians and matter tags within scope.
• Notify the AI notetaker vendor if data resides on their infrastructure, and confirm in writing that vendor-side retention has been extended.
• Document the trigger, scope, and timing in a hold memo. This is your Rule 37(e)(1) "reasonable steps" record.

This is where cloud-based notetakers create real operational friction. If your vendor's standard SLA is a 30-day deletion cycle and you cannot reach a preservation contact within that window, you have a problem the moment a court asks about it.

Privilege, Work Product, and the Deletion Decision

Retention policy interacts with privilege in a way that surprises practitioners. Deleting an AI transcript is not automatically a preservation-of-privilege move, and keeping one is not automatically a waiver. What matters is the workflow around the artifact.

An AI-generated summary of a privileged client meeting is generally protected under ABA Model Rule 1.6 and the attorney-client privilege, provided the confidentiality chain is unbroken. Once the transcript is transmitted to a third-party vendor without an appropriate agency relationship, the Heppner reasoning suggests the confidentiality prong may be compromised. Reducing the number of copies — and the number of custodians — is a privilege-preservation tactic in itself.

See our internal discussion of Basil for Law for how architectural choices interact with the privilege analysis.

Common Retention Mistakes That Become 37(e) Problems

• Ad hoc deletion after a preservation trigger. Even one manual delete after a hold letter can support an intent finding under 37(e)(2).
• Assuming vendor deletion is real deletion. Vendor backups, replication, and training-data pipelines often outlast UI-level deletes.
• Retention schedules that exist only in policy, not in the tool. If the software cannot enforce the schedule, the schedule is aspirational.
• Treating AI summaries as "not really records." They are ESI. Treat them accordingly.
• Failing to log consent. Consent-log gaps compound spoliation questions with two-party recording statute questions.

How Basil Approaches This

Basil is architected so that the retention question is genuinely in the lawyer's hands. Audio, transcription, and summarization run entirely on-device via the Apple Neural Engine. Nothing is uploaded, there is no server, and there are no subprocessors. Basil never receives your data — which means there is no vendor-side backup, no replication cluster, and no training pipeline to subpoena, and no cloud-retention SLA to fight when a hold letter arrives.

When you delete a matter in Basil, the artifacts are gone from the only place they lived. That makes both sides of the 37(e) analysis easier: preservation is a matter of your own file management, and deletion under a written schedule is defensible because there is no shadow copy elsewhere. The general Basil app (free tier, 60 minutes per month) is available today. The Basil for Law edition — with privilege attestation, consent log, matter organization, and Privileged & Confidential labeling — launches August 2026 (Solo $19.99/mo or $199.99/yr, with a 3-day monthly trial and 7-day annual trial). Learn more at Basil for Law.

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

Frequently asked questions

When does the duty to preserve AI meeting notes attach?

When litigation is reasonably anticipated — not when a complaint is filed. Demand letters, regulator inquiries, serious client complaints, and internal investigations can all trigger the duty. Once triggered, any automated deletion of AI transcripts, audio, or summaries within scope must be suspended immediately.

Can I delete AI meeting notes on a routine schedule without risking spoliation?

Yes, if the deletion happens under a written, consistently applied retention policy and no preservation duty has attached. Rule 37(e) and its Advisory Committee Notes protect good-faith, routine deletion. The risk arises when deletion continues after a hold trigger, or when the policy is applied selectively.

Are AI-generated summaries considered ESI under Rule 34?

Yes. Rule 34 defines ESI broadly to include any data or data compilations stored in any medium. Transcripts, summaries, action-item lists, audio buffers, and embeddings all qualify and are subject to preservation obligations.

What sanctions can a court impose for spoliation of AI meeting notes?

Under Rule 37(e)(1), courts can order curative measures like fee-shifting or additional discovery when loss causes prejudice. Under 37(e)(2), on a finding of intent to deprive, courts can impose adverse-inference instructions, dismiss claims, or enter default judgment.

Does storing transcripts with a cloud AI vendor waive privilege?

Not automatically, but it raises the risk. The reasoning in United States v. Heppner (S.D.N.Y. 2026) — analogizing AI platform disclosures to the third-party doctrine — suggests confidentiality can be compromised absent a well-structured agency relationship with the vendor. On-device architectures avoid the issue by design.

How long should we retain AI meeting notes for a closed litigation matter?

Practices vary, but many firms retain litigation-related transcripts for the duration of the matter plus the applicable statute of limitations or malpractice exposure period — often seven years. The correct answer depends on jurisdiction, matter type, and any client-side retention obligations.

Keep client conversations on your device

Basil transcribes and summarizes entirely on-device — no cloud, no bot, no server to subpoena. See Basil for Law → · Legal-tool reviews →

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