The Best AI Notetaker for Lawyers: Why On-Device Transcription Preserves Attorney-Client Privilege

If you have spent an hour comparing AI notetakers for your practice, you have probably noticed that every vendor promises "enterprise-grade security," SOC 2 reports, and encryption in transit and at rest. None of that answers the question a litigator or transactional attorney actually needs to answer: when my client describes the facts of a dispute on a Zoom call, who else hears it, who stores it, and who can be subpoenaed for it?

For most popular notetakers — the ones that join your call as a bot, upload audio to a vendor cloud, and run transcription on shared GPU infrastructure — the honest answer is: your vendor, their sub-processors, and potentially their model providers. That is a meaningful change in the architecture of a privileged conversation. This guide explains why that matters under the current rules and recent guidance, what to look for in an AI notetaker that is appropriate for legal work, and why on-device transcription is becoming the default recommendation for client calls, deposition prep, and witness interviews.

Why "best AI notetaker for lawyers" is really a confidentiality question

Attorneys have duties that general knowledge workers do not. ABA Model Rule 1.6 requires a lawyer to make reasonable efforts to prevent the inadvertent or unauthorized disclosure of information relating to the representation. Rule 1.1 Comment 8 extends competence to the benefits and risks of relevant technology. In July 2024, the ABA issued Formal Opinion 512 on generative AI, which tells lawyers to understand how a tool collects, stores, uses, and discloses client information before putting privileged content into it. In December 2025, the NYC Bar Formal Opinion 2025-6 reinforced the same analysis for New York practitioners.

So when you type "best AI notetaker for lawyers confidentiality" into a search engine, you are really asking three overlapping questions:

• Does the tool introduce a new third party — a vendor, sub-processor, or model host — into my privileged conversation?
• If it does, is that disclosure consistent with my client's informed consent and with my duty of confidentiality?
• If the vendor is ever served with a subpoena, preservation demand, or discovery request, what will they have, and whose it is?

On-device transcription changes the answer to the first question. That, in turn, changes the answer to the other two.

The third-party problem with cloud notetakers

The vast majority of AI notetakers on the market today use one of two architectures:

1. Bot-joins-the-call. A named participant (often "Otter.ai Notetaker" or "Fireflies Notetaker") appears in the meeting, records audio and video, and uploads the recording to the vendor cloud for transcription and summarization.
2. Cloud capture without a bot. A desktop client captures system audio locally, then streams it to the vendor cloud for processing. No visible bot, but the data path is the same.

In both cases, the audio and transcript sit on infrastructure the vendor controls. The vendor's privacy policy, sub-processor list, and data retention schedule govern what happens next. In Brewer v. Otter.ai (2025), a putative class action filed in the Northern District of California, plaintiffs alleged that Otter's notetaker recorded confidential meetings without the consent of all participants and that transcripts were used to train the vendor's models. The case is a reminder that "the bot in the corner" is a party with its own interests, record-keeping, and legal exposure.

The analytic risk for privilege sits on top of that. 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: disclosing content to an outside service provider can defeat the confidentiality that privilege requires. Heppner was about a party, not a lawyer's notetaker, but the reasoning travels. If a cloud notetaker is architecturally a third-party service that receives and stores the substance of a privileged conversation, a court can be asked to decide whether that disclosure was consistent with maintaining privilege.

You do not have to believe that every cloud notetaker will produce a waiver to take the point seriously. The West Technology Group v. Sundstrom line of cases shows that courts will look closely at the handling of allegedly privileged ESI when an opposing party puts it in issue. The question you want to be able to answer, under oath if necessary, is: "The vendor never received the conversation."

What on-device transcription actually means

"On-device" is a term vendors use loosely. For legal work, it should mean something specific: the audio is captured, transcribed, and summarized on the hardware in front of the attorney, using local models, without the audio or the resulting text being transmitted to the vendor at any stage of normal operation.

On modern Apple Silicon Macs, the Neural Engine can run speech-to-text and summarization models at usable speed without a network round-trip. That is the hardware foundation that makes an on-device notetaker practical for a full day of client calls, not just a demo.

When evaluating a tool that claims on-device processing, ask:

• Where does the audio go between capture and transcript? Is there a network call to the vendor or to a model provider?
• Where is the transcript stored? On the device, in the user's own cloud account, or in the vendor's cloud?
• Who is in the sub-processor list? A vendor with no sub-processors cannot be compelled to produce what it does not have.
• Is sync optional, and whose account is it? Sync to the user's own iCloud or OneDrive is a different legal posture than sync to the vendor's bucket.

Comparison: how AI notetaker architectures handle privileged content

Use cases where the architecture matters most

Client intake and ongoing client calls

Intake is where the facts are rawest and the client is most candid. A bot named after a vendor sitting in the Zoom is a disclosure event your client did not sign up for. On-device capture lets you take the same notes you would take by hand, faster, without changing the participant list.

Deposition preparation

Prep sessions with a witness are classic work product and, when the witness is the client, privileged. Pushing that audio to a vendor cloud for transcription creates a record of exactly how you prepared the witness that lives outside your file. On-device transcription keeps the record in the matter file where it belongs.

Witness interviews and internal investigations

Internal investigation interviews under Upjohn warnings depend on the interview being for the purpose of providing legal advice to the organization. Introducing a third-party notetaker complicates both the warning and the chain of custody. On-device capture avoids the complication.

Co-counsel and expert calls

Communications with consulting experts and co-counsel are typically covered by work product or common-interest doctrines, but those doctrines assume a controlled circle of participants. Keeping the notetaker off the vendor's servers keeps the circle controlled.

What to look for in an AI notetaker for legal work

Independent of any specific product, the features that matter for a legal buyer are:

• On-device transcription by default, not as a paid add-on or a toggle buried in settings.
• No vendor sub-processors for the content of the conversation. If the vendor uses a third-party ASR or LLM provider, the audio or text goes to that provider.
• Transparent sync model. If audio or transcripts sync anywhere, it should be to the user's own cloud account, and the vendor should say so plainly.
• Willingness to sign a DPA and NDA. Even if the vendor never receives content, a DPA documents the architecture for your file.
• Matter-level organization and privilege labeling. "Privileged & Confidential" headers on exports, consent logs, and matter folders are small features that matter when a file is produced.
• Honest marketing. If a vendor claims "nothing ever leaves your device" while also offering cloud sync, read the architecture, not the headline.

For a deeper walk-through of the ethics-opinion framework, see our companion piece on Basil for Law and the duty-of-competence analysis under ABA Opinion 512.

Consent, notice, and the rules that still apply

On-device capture reduces the third-party disclosure problem. It does not eliminate the recording-consent problem. State wiretap laws still apply. In two-party-consent jurisdictions (California, Florida, Illinois, Massachusetts, Pennsylvania, Washington, and others), you need the consent of all parties to record a confidential communication, regardless of where the recording is processed. The Reporters Committee recording guide is a useful starting point for state-by-state rules, though attorneys should confirm current law for their jurisdiction.

A practical protocol:

1. Tell the client, witness, or opposing participant at the start of the call that you are taking AI-assisted notes on your own device.
2. Get affirmative consent on the record. Log it in your matter file.
3. Keep a short written policy explaining your notetaking architecture that you can share on request. Clients increasingly ask.

Common objections, briefly answered

"My vendor is SOC 2 Type II and signs a BAA — that is enough." SOC 2 is a controls attestation, not a privilege analysis. A vendor can be perfectly compliant with its own security program and still be a third party that received your client's conversation.

"Encryption in transit and at rest solves it." Encryption protects against unauthorized access. It does not change who the authorized recipient is. The vendor is still the recipient.

"On-device models are not as good." For courtroom-quality verbatim transcription, that is often still true. For meeting notes, summaries, and action items from a client call, Apple Neural Engine models are good enough for daily use, and they improve with each OS release.

"If I turn on sync, is it the same as a cloud notetaker?" No. Sync to your own iCloud account means Apple holds an encrypted copy under your Apple ID. That is a different legal and factual posture than a vendor holding the content under its own control. You can also turn sync off.

A short decision framework

When you evaluate a tool for privileged work, run three questions in order:

1. Receipt. Does the vendor ever receive the audio or transcript in normal operation? If yes, you are in a cloud-notetaker risk analysis.
2. Storage. Where does the content live at rest, and whose account is it in?
3. Compulsion. If the vendor is served tomorrow, what can they produce?

A tool that answers "no, the user's device and optionally the user's own cloud, and nothing" is in a different category than a tool that answers "yes, our cloud, and the transcripts." That is the distinction that "best AI notetaker for lawyers confidentiality" is really asking about.

How Basil approaches this

Basil was built by a practicing lawyer for this exact problem. Transcription and summarization run on-device on the Apple Neural Engine. Basil has no server and no sub-processors; Basil never receives your recordings, transcripts, or summaries. iCloud Sync is on by default so your library, audio included, is available across your Macs — it is stored in your own iCloud account under your Apple ID, not in a Basil cloud, and you can turn it off in Settings › iCloud. Basil captures in-person meetings and virtual calls (Zoom, Teams, Meet) through on-device capture on macOS in Computer mode, so no bot joins the call.

The Basil for Law edition adds a privilege attestation, consent log, matter organization, and automatic "Privileged & Confidential" labeling on exports. 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 general Basil app, with a 60-minute free tier, is available today.

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.