Your AI notetaker might be a legal liability (here's how to fix that)

A sales rep reads out a client's account number on a call. The AI notetaker sitting in the background captures every word, transcribes it, stores it, and quietly makes it searchable to anyone with access to the workspace. Six months later, a regulator asks for records. Nobody remembers that recording exists — until someone finds it in a place it was never supposed to be.

This isn't a hypothetical. It's the exact scenario playing out in courtrooms right now. Over the past year, AI meeting assistants have gone from "nice productivity add-on" to a genuine legal exposure category, and most companies adopted them without ever asking the questions a compliance team would ask.

Why are companies suddenly worried about AI notetakers?

Because the lawsuits are real, and they're piling up. Otter.ai has become a focal point for lawsuits that allege unlawful recording, storage, and use of participants' conversations, often without the clear, informed consent that state and federal wiretap statutes require. In one high-profile case, plaintiffs allege that Otter's notetaker automatically joins meetings across Zoom, Teams, and Meet, records conversations—including those of non-users—and uses these discussions to train its machine learning models, all without proper consent or disclosure.

The design flaw at the center of that case is worth understanding, because it's common across the category. Otter's notetaker seeks permission only from the meeting host, and even then, only if the host is not themselves an Otter user. Other participants cannot disable the tool. If the host has integrated their calendar with Otter, the bot joins and begins transcribing without any affirmative consent from anyone in the room.

Otter isn't the only tool named. In Cruz v. Fireflies.AI Corp., filed in December 2025, the plaintiff alleges that Fireflies.AI collected voiceprint biometrics from meeting participants without the informed written consent required by the Illinois Biometric Information Privacy Act, and the complaint asserts the company's AI notetaker joined meetings, captured voice data, and processed it to create speaker-identifying voiceprints without satisfying BIPA's notice and consent requirements. Another suit targets a competitor's marketing claims directly — Granola's own marketing copy told prospective customers that other people on a call "won't know it's there," and plaintiffs cite that line in a federal court filing as evidence.

Even law firms are getting formal guidance on this now. In December 2025, the New York City Bar Association's Committee on Professional Ethics issued Formal Opinion 2025-6 on the ethical issues affecting AI recording of client conversations, addressing scenarios where a client shares sensitive financial details on a call that an attorney is quietly transcribing. If bar associations are writing opinions about this, it's not a fringe concern anymore.

What actually creates the legal risk?

It comes down to a handful of overlapping issues, and most teams are only aware of one or two of them.

Consent law isn't one rule. Some states permit recording with consent from only one participant, while others require consent from all participants, and the distinction is critical. Remote meetings make this messier — remote work amplifies this risk, since virtual meetings often include participants located in multiple states, and businesses may not know where every attendee is physically located at the time of a call.

Visibility doesn't equal consent. A bot showing up as a visible participant in the call feels like disclosure, but it isn't legally the same thing. Whether you use a bot that joins visibly or desktop recording that captures audio locally, you still need affirmative disclosure — the bot's visibility may make the conversation about consent more natural, but the visibility itself is not a substitute for actually obtaining it.

Retention turns a conversation into a permanent record. Recordings in the Otter case were allegedly retained indefinitely and used to train AI models, including the voices of individuals who were unaware they were being recorded. That's the pattern regulators and plaintiffs' attorneys keep pointing to — data that should have had a short shelf life instead sits around forever.

Biometric data adds a whole separate layer of exposure. Speaker-identification features in AI notetakers may trigger separate biometric privacy laws, such as Illinois' BIPA, beyond standard recording consent, creating distinct compliance exposure. If your notetaker fingerprints voices to tell speakers apart, that's not just a recording question anymore.

How do you actually vet an AI notetaker for compliance?

Before rolling out any tool, get clear answers to these:

1. Who obtains consent, and from whom? A tool that only asks the host, and only sometimes, is a red flag.
2. Does it train on your data by default? Some competitors explicitly train their AI models on customer data, and many lack HIPAA compliance entirely.
3. Can you delete the raw recording while keeping the summary? This "decoupled retention" model matters more than almost anything else on this list.
4. Does it capture biometric voiceprints, and can that be turned off?
5. What does the audit trail look like if a regulator asks for records six months from now?

What does a compliance-first notetaker actually look like in practice?

Fellow is one of the few tools built around these questions rather than bolting them on after a security review flagged them. The company's approach is a useful blueprint for what to look for, regardless of which vendor you end up choosing.

Zero-Day Retention. Instead of keeping raw audio forever, Fellow's optional Zero-Day Retention feature prevents recordings and transcripts from being retained after a meeting ends, and it's built for MNPI-sensitive environments. The summary and action items survive; the raw recording that could leak an account number doesn't.

Redaction that's visible, not silent. Every review action, comment, dismissal, and redaction is logged, and redactions are always visible rather than silent, creating a documented record for audit and review. That distinction matters — a redaction nobody can trace is just as risky as no redaction at all.

Pause & Resume that actually deletes, not just hides. Pause & Resume lets you stop recording at any point during a meeting — any content captured while paused is permanently excluded from the transcript and meeting notes, giving teams full control over what gets recorded. That's a meaningfully different posture from hoping someone remembers to hit a mute button.

Consent capture built into the workflow. Fellow's most recent update added this directly: consent capture gives attendees the option before recording begins, addressing the exact gap that's driving the Otter and Fireflies lawsuits.

Certifications that back up the marketing. Fellow is SOC 2 Type II certified and supports encryption controls including AES-256, and also complies with GDPR and HIPAA requirements, making it suitable for deployment in regulated industries including financial services, legal, and healthcare. Security documentation confirming this is available for due-diligence review — meaning your compliance team doesn't have to take it on faith.

An audit trail built for regulators, not just IT. The Super Admin API produces audit-ready deletion logs documenting what was deleted, when, and by which policy trigger — formatted for production in regulatory examinations without additional transformation.

For a deeper look at how these pieces fit together, Fellow's feature documentation breaks down each control individually, and trust.fellow.ai hosts the actual compliance paperwork rather than just marketing claims.

How do you roll this out without breaking your workflow?

Getting the governance right doesn't mean losing the productivity benefit that made AI notetakers appealing in the first place. Here's a sequence that works:

1. Map your meeting types first. Client calls, board meetings, and personnel discussions need different retention settings — don't apply one blanket policy everywhere.
2. Turn on Zero-Day Retention for anything touching regulated or sensitive data, and leave standard retention for internal team syncs where the summary is more valuable than the raw transcript.
3. Build consent into your meeting invite templates, not just into the tool itself. A calendar note that says "this meeting will be recorded and transcribed by AI" closes the gap that plaintiffs' attorneys keep exploiting.
4. Set redaction rules per meeting type for account numbers, names, and anything that counts as non-public information in your industry.
5. Ask for the SOC 2 report and BAA language in writing before deployment — request the SOC 2 Type II audit report, current DPA, and BAA language for your specific plan before deployment, and involve your security and legal teams in the review.

Book a Fellow demo if you want to see how these controls get configured for your specific compliance requirements, or start with the free plan to test the redaction and Pause & Resume features on a few low-stakes meetings first.

The bots aren't going away. The question was never whether to use one — it's whether the one you picked was built by people who thought about the regulator's phone call before they thought about the demo. Most weren't. Choose accordingly.