AI Meeting Notes for Lawyers: The 2026 Confidentiality & Privilege Checklist
If you are evaluating an AI meeting notetaker for a law practice in 2026, the questions that matter are not the ones on the vendor's landing page. They are questions about where audio actually travels, who can be compelled to produce it, whether a third party sees privileged content, and whether the architecture of the tool is consistent with Model Rule 1.6 and the privilege itself.
This is a buyer-intent checklist. It is organized around the decisions a partner, GC, or solo has to make before a tool touches a client conversation. It is grounded in the ABA's July 2024 Formal Opinion 512 on generative AI, the New York City Bar's Formal Opinion 2025-6, and recent decisions that have begun to test what happens when privileged content flows through an AI vendor.
1. The threshold question: is the AI a "third party" that breaks privilege?
Attorney-client privilege generally protects confidential communications between lawyer and client made for the purpose of obtaining legal advice. The classic rule: if an unnecessary third party is present or receives the communication, privilege is waived. Courts have long recognized agent exceptions (interpreters, paralegals, experts retained to assist counsel), but the exception is narrow and fact-specific.
An AI notetaker that sends audio to a vendor's servers is, in most readings, a third party. In US 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. That case involved a consumer chatbot, not a notetaker, but the reasoning travels: if a non-agent third party receives the content, the protection weakens.
The NYC Bar's Formal Opinion 2025-6 (December 2025) is the clearest recent statement for practitioners. It confirms that lawyers may use generative AI consistent with their duties, but only when they understand the tool's data flow and take reasonable steps to prevent disclosure of confidential information.
2. Audio, transcripts, summaries: three different artifacts, three different risks
Many vendor pages collapse "your meeting" into one thing. For a privilege analysis, you have to separate them:
- Audio. The raw recording. Richest evidence; often the subject of subpoenas and discovery disputes.
- Transcript. The verbatim text. Frequently retained longer than audio and easier to search.
- Summary. Model-generated. May omit, paraphrase, or hallucinate. Still discoverable.
Ask, for each artifact: where is it created, where is it stored, who can access it, and what is the retention. A tool that keeps audio on device but ships the transcript to a cloud LLM for summarization has still exposed the privileged content.
3. The checklist
| # | Question to ask the vendor | Why it matters |
|---|---|---|
| 1 | Does any audio, transcript, or summary leave the device? To whom, and when? | Each hop is a potential third-party disclosure and a potential subpoena target. |
| 2 | Is processing on-device, in your cloud, or via a third-party model provider (OpenAI, Anthropic, Google)? | Sub-processors multiply the compliance surface and the DPA chain. |
| 3 | Do you train models on customer recordings, transcripts, or summaries, now or ever? | Training on client content is incompatible with Rule 1.6 confidentiality. |
| 4 | Will you sign a DPA and an NDA? What is your sub-processor list? | Required by most engagement letters and by Opinion 512's competence duty. |
| 5 | What is the retention default? Can I set it to zero? | Shorter retention reduces discovery and breach exposure. |
| 6 | Does a bot join the call? If so, is it disclosed to participants? | Bots create consent-notice obligations and discoverable metadata. |
| 7 | If the service is subpoenaed for a matter's audio, what can you produce? | If the answer is "everything," you have a server-side privilege problem. |
| 8 | Is there a way to flag content as privileged in the stored record? | Helps downstream review and clawback under FRE 502. |
| 9 | How is consent captured and logged? | Required under one-party and two-party recording regimes. |
| 10 | Does the architecture meaningfully separate storage from processing? | A vendor can process on-device and still store in the user's own cloud; those are different risks. |
4. "No training on your data" is necessary, not sufficient
Every serious vendor now claims they do not train on customer content. That is table stakes. It does not answer the more important questions: who holds the data at rest, who can be compelled to produce it, and which sub-processors see the content in transit.
The Brewer v. Otter.ai litigation (2025) brought public attention to the gap between marketing and architecture around meeting transcription. Whatever the ultimate resolution, the suit is a reminder that plaintiffs will test these claims and that courts will look at the actual data flow, not the privacy page.
ABA Formal Opinion 512 frames the lawyer's duty as technological competence: understand the tool well enough to make an informed judgment. "They told me they don't train" is not understanding.
5. Vendor contracts: DPA, BAA, sub-processor transparency
Under ABA Model Rule 1.6, a lawyer must make reasonable efforts to prevent unauthorized disclosure of client information. For vendor relationships, that usually means:
- A signed Data Processing Addendum that names sub-processors and commits to notice on change.
- An NDA scoped to the engagement.
- For healthcare matters, a Business Associate Agreement; for EU data, GDPR Article 28 terms.
- Written confirmation of retention, deletion on request, and the vendor's response to compulsory process.
If a vendor will not sign a DPA, or cannot name its sub-processors, that is the answer. Move on.
6. Consent, recording law, and the bot problem
Recording law varies by jurisdiction. Roughly a dozen US states require all-party consent, and many foreign jurisdictions do as well. A notetaker that silently joins a call as a bot can create both a consent problem and an evidentiary record that you joined with recording enabled.
Two practical points:
- If a bot joins the meeting with a name like "Firm AI Notetaker," that is itself notice, but it is not consent. You still need affirmative agreement on the record for all-party jurisdictions.
- Bot-free capture (recording the system and microphone audio locally on your Mac) avoids the bot-as-participant problem, but it does not avoid the consent problem. The duty to obtain consent attaches to the lawyer, not the vendor.
See the Reporters Committee's state-by-state recording guide as a starting point, then confirm with current authority in your jurisdiction.
7. What the recent case law suggests
A few data points worth weighing:
- US v. Heppner (S.D.N.Y. Feb 2026). Chats with a public AI platform were held not privileged under a third-party doctrine analogy. The holding is narrow but the reasoning is portable to any AI service that receives client content.
- West Technology Group v. Sundstrom (D. Conn. 2024). Addressed obligations around confidential information in the vendor context and reinforced that failing to protect it has consequences beyond ethics rules.
- Brewer v. Otter.ai (2025). Pending consumer litigation focused on recording and transcription practices; a reminder that discovery into a notetaker's data flow is live.
None of these cases hold that AI notetakers are per se incompatible with privilege. They hold, cumulatively, that the architecture matters and that lawyers who treat these tools as invisible will have a hard time later.
8. On-device processing: why it is the cleanest path
If audio, transcript, and summary are all generated on the lawyer's own device, the third-party disclosure problem shrinks dramatically. There is no vendor server holding the content, no vendor employee with access, and no vendor subpoena target for the underlying audio.
Two caveats worth keeping in mind:
- Processing is not storage. A tool can process on-device and still store the resulting library in a cloud the user controls (for example, iCloud). That is a legitimate design, but it is a different risk profile from "nothing exists anywhere." If iCloud sync is on, the library sits in the user's own Apple account and is subject to Apple's legal process, not the vendor's. Users who want to limit that can turn sync off.
- Speech recognition modes matter. Some speech frameworks on Apple platforms can route audio to Apple servers unless on-device-only recognition is selected and supported for the language. Confirm the setting.
The practical takeaway: on-device processing is the architectural choice that gives you the strongest story about privilege and the shortest list of third parties to describe in an engagement letter. It does not, by itself, remove every obligation.
9. A short internal process for adoption
Before rolling an AI notetaker to a practice group, firms we work with typically do the following:
- Write a one-page data-flow diagram based on vendor answers to the checklist above.
- Decide a default retention and a privileged-matter retention (often zero).
- Update the engagement letter to disclose the use of AI assistance and recording.
- Train attorneys on consent scripts for all-party jurisdictions.
- Decide whether summaries are shared with clients and under what label (for example, "Privileged & Confidential").
- Document the decision. Opinion 512 rewards a reasoned, written adoption record.
If you are earlier in the evaluation, our Basil for Law pillar page outlines the architectural choices in more detail, and our write-up on attorney-client privilege and AI notetakers goes deeper on the doctrine.
10. Red flags in vendor marketing
Phrases that should trigger a follow-up question, not a purchase:
- "Bank-grade encryption" without a statement of who holds the keys.
- "SOC 2 certified" without scope; SOC 2 is not a privacy standard and the report controls.
- "Court-ready" or "privilege-protected" as marketing adjectives.
- "No human ever sees your data" without an explanation of exception handling, QA sampling, and support access.
- Silence on sub-processors or model providers.
For a broader framework on competence and AI, the Georgetown Journal of Legal Ethics and your state bar's recent opinions are better starting points than vendor blogs.
How Basil approaches this
Basil runs transcription and summarization on the device using the Apple Neural Engine. Basil, the company, never receives your recordings, transcripts, or summaries. iCloud Sync is on by default, which means the library, audio included, is stored in your own iCloud account rather than on a Basil server; you can turn sync off in Settings › iCloud. On macOS, Computer mode captures in-person and virtual meeting audio locally without a bot joining the call; on iPhone and iPad, the microphone is used. Recording consent remains the lawyer's responsibility.
The Basil for Law edition adds a privilege attestation flow, a consent log, matter organization, and Privileged & Confidential labeling. Solo pricing is $19.99/month or $199.99/year, with a 3-day free trial on monthly and 7-day on annual (one trial per Apple Account). Basil signs DPAs and NDAs on request. The architecture is designed to reduce the third-party-disclosure surface described above; it is not a legal guarantee, and no vendor can give you one.
This article is for information only and is not legal advice.
Frequently asked questions
Can lawyers use AI meeting notetakers without waiving attorney-client privilege?
Potentially, yes, but it depends on the architecture. If audio, transcripts, or summaries travel to a vendor acting as a non-agent third party, privilege is at risk under the third-party doctrine and recent decisions like US v. Heppner (S.D.N.Y. Feb 2026). On-device processing shrinks that surface. Lawyers should still obtain informed consent, sign DPAs, and document the decision consistent with ABA Formal Opinion 512 and NYC Bar Opinion 2025-6.
Is "we do not train on your data" enough to protect client confidentiality?
No. A no-training commitment is table stakes. It does not address who holds the data at rest, which sub-processors see it in transit, retention defaults, or the vendor's response to compulsory process. Ask for a DPA, a sub-processor list, and a written description of the data flow for every artifact (audio, transcript, summary).
Does on-device processing mean "nothing is uploaded"?
Not necessarily. Processing and storage are different. A tool can transcribe on the device and still store the resulting library in a cloud the user controls, such as iCloud. That is a legitimate design but a different risk profile. Check whether sync is on by default, whose cloud holds the data, and whether the user can turn sync off.
Do I need all-party consent to use an AI notetaker on a client call?
In roughly a dozen US states and many foreign jurisdictions, yes. Even in one-party-consent states, your engagement letter, local ethics rules, and the privilege analysis may require disclosure. Capture consent on the record at the start of the call and keep a log. The duty attaches to the lawyer regardless of whether the tool uses a bot.
What should an engagement letter say about AI meeting notes?
Disclose that AI-assisted transcription and summarization may be used, describe where processing and storage occur, identify sub-processors if any, state the retention policy, and obtain client consent. ABA Formal Opinion 512 rewards a written, reasoned adoption record and informed-consent discussion for matters where AI meaningfully touches client information.
How is Basil different from cloud-based AI notetakers for confidentiality purposes?
Basil runs transcription and summarization on the device using the Apple Neural Engine, and Basil the company never receives recordings, transcripts, or summaries. iCloud Sync is on by default, which stores the user's library in their own iCloud account; it can be turned off. On macOS, Computer mode captures audio without a bot joining the call. This reduces the third-party-disclosure surface but does not replace the lawyer's duties around consent, competence, and documentation.
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.