AI Meeting Assistants for Law Firms After Heppner

The short answer: United States v. Heppner (S.D.N.Y., February 2026) held that a defendant’s chats with a consumer AI tool were not privileged. A law firm should therefore record privileged meetings only with a tool that has contractual no-training terms, isolated or on-device processing, and a written counsel-directed workflow. For the strictest matters, Hedy can process meetings on your own device on supported hardware, so the conversation does not have to reach a third party’s server. This guide covers the privilege question; for a ranked shortlist, see our guide for lawyers.

What Heppner held, and what it did not

In United States v. Heppner, No. 25-cr-00503-JSR (S.D.N.Y.), Judge Jed S. Rakoff ruled from the bench on 10 February 2026 and issued a written memorandum on 17 February. The court held that documents the defendant generated with the consumer version of Anthropic’s Claude were protected by neither the attorney-client privilege nor the work-product doctrine. As the reasoning is reported, three things went wrong:

1. The AI is not an attorney. A user prompting an AI tool is not consulting a lawyer.
2. Confidentiality was not preserved. The tool’s privacy policy contemplated data collection and disclosure to third parties, so the communications were not confidential.
3. It was not at counsel’s direction. The defendant used the tool on his own initiative, so the outputs were not part of seeking or providing legal advice.

Two cautions. First, Heppner was about a person’s own chats with a chatbot. It was not about meeting assistants and not about audio recordings, so no court has yet applied it to meeting transcription. Second, the judge left room for a different result if counsel had directed the use of the tool. Our conclusion that enterprise AI, used at counsel’s direction on contractually confidential terms, can survive the same analysis is our own inference, not a holding. We found no appeal reported as of 27 October 2026.

What carries over to meeting tools is the logic: whether a conversation stays confidential depends on what the vendor’s terms let it do with the content. That is the same question at the center of the Otter.ai lawsuit, where a court let claims proceed on the theory that a vendor that keeps and uses recordings for its own purposes can be a third-party eavesdropper. See the lawsuit tracker for its status.

The evaluation checklist

For a meeting assistant to be defensibly deployed inside a law firm or compliance-bound organization, four conditions matter:

1. No third-party retention or training. The vendor must commit in a signed DPA or BAA that meeting audio and transcripts will not be used to train models, will not be disclosed to subprocessors beyond what service delivery needs, and will be deleted on request. “We don’t train on customer data” in marketing copy is not sufficient. It should be in the contract.

2. Demonstrable data isolation. Either the audio never leaves the device, or it lives in a tenant-isolated environment with documented controls.

3. Deployed at counsel’s direction. This is a workflow concern, not a vendor concern. Firms need policies that specify which matters can be recorded, when consent must be explicit, and which AI features are approved for privileged contexts.

4. Verifiable security posture. For a cloud vendor that holds your conversations, SOC 2 Type II is the usual minimum, with HIPAA terms if health information is involved and pen test reports under NDA. For a tool that processes on your device, less data reaches the vendor, so the architecture carries more of the weight than the audit.

How the major tools compare

Facts below are from the vendors’ published terms as we re-checked them on 14 October 2026.

On-device tier (strongest privilege posture)

Hedy’s speech recognition always runs on your device. Since its Local AI release in May 2026, the rest of the pipeline (summaries, detailed notes, chat and live coaching) can run there too on Apple Silicon Macs, capable Windows GPUs, iPhone 15 Pro and newer, recent iPads and, since September 2026, Android phones with a supported GPU. In that mode the conversation does not have to reach a Hedy server, so there is no vendor-held copy to subpoena or to argue was used for the vendor’s own purposes. Cloud Sync is optional. If you choose Hedy’s cloud AI instead, its AI providers are contractually barred from training on your data, and you can pick EU or US residency at signup.

Two honest limits. Hedy completed a SOC 2 Type I examination in April 2026, not Type II, and its BAA is available on Enterprise only, with a 200-seat minimum. A small firm that needs a signed BAA today may need another vendor, or may decide that on-device processing makes the BAA question moot for its matters. Hedy’s bot-free capture also does not replace consent. You still tell the room you are recording.

MacWhisper is the strictest fully local option: a one-time-purchase Mac app that transcribes with local Whisper models and needs no cloud component by default. Its AI summaries run either on a local model you set up yourself or on a cloud provider using your own API key. It has no live coaching, but for a transcription-first use case it is a low-risk tool with no vendor-side transcript store to subpoena.

Cloud with strong controls

Fellow states that it never trains AI on customer data and offers a BAA on Enterprise. Fireflies.ai states it does not train on meeting content and holds SOC 2 Type II, with HIPAA support, private storage and retention controls on the Enterprise plan. Default workspace settings retain content longer than most firms want, so the configuration matters. Fireflies also faces Illinois biometric-privacy litigation over voiceprints, which a firm’s vendor review should ask about.

Zoom AI Companion inherits Zoom’s mature compliance posture, and Zoom says it does not use audio, video or chat content to train AI. It is still a cloud service tied to Zoom’s pipeline. For firms standardized on Zoom with the right contract terms it can be defensible.

Otter.ai holds SOC 2 Type II and offers HIPAA as an Enterprise add-on, but it trains on de-identified data by default. A federal class action, In re Otter.AI Privacy Litigation, survived a motion to dismiss in part on 13 August 2026, and Otter filed its answer on 17 September. The litigation does not establish liability, but its theory (a vendor that keeps and uses recordings for itself) maps onto the workflow concern Heppner highlights. Granola also trains on de-identified data by default, signs a BAA on Enterprise only, and is the defendant in a proposed class action over silent capture.

Where more scrutiny is needed

Most other cloud-based assistants in our rankings (tl;dv, Read.ai, Notta, Tactiq, Sembly, MeetGeek, Grain) offer some combination of SOC 2 and GDPR compliance, but their default postures are built for business productivity, not regulated work. Some train by default or do not state a position. Deploying any of them inside a law firm requires more diligence than their marketing pages disclose.

TwinMind suits mobile-first practices: it captures on iOS and Android with no bot, and its raw audio is deleted after transcription by default. Its main Ear-3 transcription and all AI chat run in the cloud, though; the on-device Ear-2 model only takes over when you are offline. For privileged interviews it is a convenience pick, not a replacement for MacWhisper or Hedy.

Practical guidance for firms

Three actions every legal or compliance-bound team should take this quarter:

Audit which tools are actually being used. Associate-installed Otter accounts, free Fireflies tiers used by junior staff, browser extensions that transcribe Google Meet calls: these are the real exposure surface, not the tool you formally sanctioned.

Update written policies. Specify which matters can be recorded, who must consent, which tools are approved for which contexts, and who at the firm can expand that list. The policy provides the “at counsel’s direction” element.

Move privileged-context recording to on-device or contractually isolated tools. For privileged communications, sealed matters and confidential client interviews, the architecture that most cleanly removes the third-party question is one where the conversation does not leave the device. See our privacy guide and, for health-related matters, the healthcare guide and our HIPAA-compliant notetaker comparison.

Heppner did not invent new compliance requirements. It made explicit a standard that careful firms were already trying to meet, and the meeting assistant market has several products that can meet it. The work is matching the right tool to the right matter, and writing the policies that hold both in place.

Sources: the Heppner bench ruling (10 February 2026) and memorandum (17 February 2026), as reported by law-firm summaries; vendor terms as re-checked on 14 October 2026; Reporters Committee for Freedom of the Press recording guide for state consent rules. This is analysis, not legal advice.