Who Approved the Bot in Your Meeting?

Something joined your last meeting that nobody voted on. It took notes, built a summary, and filed a transcript somewhere. AI notetakers have become the default guest on video calls, and most of them are genuinely useful. The trouble is that almost no one has decided how they should behave: who turns them on, which meetings they stay out of, where the recording lives, and how long it survives. Those are decisions, and leaving them unmade is its own choice, usually the wrong one.

An unmanaged notetaker is a legal and confidentiality exposure wearing the badge of a productivity tool. Here is the part worth getting right.

A bot that records is still recording

The law does not care that the thing capturing the conversation is software. Federal law lets you record a call if one person consents (18 U.S.C. § 2511), but about a dozen states require every participant to agree. The law firm Mayer Brown names California, Connecticut, Florida, Illinois, Maryland, Massachusetts, Michigan, Montana, Nevada, New Hampshire, Pennsylvania, and Washington. One person dialing in from a two-party state can pull the entire meeting under that stricter rule, even if you are sitting somewhere that only needs one. California's privacy statute can reach past you to the vendor itself.

A banner the bot flashes once it has already started listening is not consent. The notice has to come before anything gets captured. A recording made the wrong way is not just awkward. In the strict states it carries real liability.

The transcript outlives the meeting

Every summary and transcript is a document, and documents get subpoenaed. Mayer Brown puts it directly: AI-generated transcripts "can expand the volume of data subject to litigation hold obligations." A performance conversation or a termination that used to live in memory now exists as a verbatim, searchable record that an opposing lawyer can ask for. Transcripts also get things wrong, dropping tone and context in ways that read badly a year later in front of a jury.

For a law or accounting firm the edge is sharper. A federal court held this year, in United States v. Heppner, that material a person prepared with a consumer AI tool was not privileged. A notetaker sitting quietly in a client call can erode the exact protection your client is paying you to preserve, and no one in the room would notice it happening.

So govern it, do not ban it

The answer is not to rip the tool out. It is to make the decisions the tool assumes you already made. A workable rule fits on two pages.

Default to off. The notetaker runs when someone turns it on for a reason, not automatically on every invite that lands. Mayer Brown's own guidance is to "default to not using an AI notetaker unless legal, privacy, and information security safeguards have been confirmed."

Announce it and get consent, every time. When you do not know where everyone is sitting, treat the meeting under the strictest rule that could apply. Say the notetaker is on, and let people object before it hears a word.

Draw a hard line around sensitive rooms. Legal advice, HR investigations, board discussions, anything privileged or regulated. The bot does not come in. That is one sentence in a policy, and it prevents most of the damage.

Approve the tools centrally. One or two vetted notetakers, not whatever each person found in an app store. Set a retention limit, build a deletion step, and decide which version of the notes is the official record, so you are not holding three conflicting ones when it matters.

Vet the vendor like anyone else touching your data. Does it train its models on your conversations. Is it independently audited. Is the data encrypted. Can you control how long it is kept. A vendor that cannot answer those cleanly has just answered them.

The one we use

I have been running Circleback on my own meetings for a while now, and I am happy with it. It earned that by clearing the same checks we would put in front of any client's tool. Circleback's security page states that the company does not use customer data to train models, that it is SOC 2 Type II and EU-U.S. Data Privacy Framework certified, and that it encrypts data at rest and in transit. Workspace admins control how long meeting data is kept, and the company posts an external penetration-test report rather than asking you to take its word.

We do not take a referral fee for saying that. We are vendor neutral and we do not work on commission, so when we name a tool it is because it held up, not because someone bought the slot. What Circleback handles is the vendor side of the risk, which is the side a policy alone cannot fix. The rest, consent and which meetings stay private, is still yours to run.

The real point

The notetaker is not the problem. Not deciding is. Pick the tool on purpose, write the short rule for when it records and when it stays out, and tell your people what it is. The firms that do this get all of the convenience and almost none of the exposure.

If you want help drawing that line for your own meetings, that is a conversation we have often.