AI Notetakers Move Beyond Meetings, Forcing Companies to Rethink Recording Policies

AI meeting records no longer depend on a bot appearing in a video call. A watch, pendant, or pair of smart glasses can capture a conversation without joining the meeting platform, which changes the practical limits of company recording policies. My read is that organizations now need to govern the presence of recording devices as well as the software they’ve approved.

There’s a real business case for AI notes. Teams that spend the day in customer meetings can use summaries to follow through more consistently. At AI-first startups, a new account manager may inherit recorded customer conversations and use an AI system to identify what matters to each account. That can improve continuity and service. The challenge is ensuring that a useful record doesn’t become an undisclosed recording, a misleading summary, or a source of information that travels farther than intended.

Earlier warnings about AI notetakers focused on bots joining calls. Those tools raise familiar concerns: unauthorized adoption across teams, weak vendor security, consent, and possible litigation. They also influence what counts as the meeting’s official record by selecting key points and assigning action items. But recording no longer has to be visible in the meeting software. The microphone can travel with an attendee.

Recording has moved onto personal devices

On July 28, Granola, an AI notetaker frequently used in the authors’ circles, launched an Apple Watch app. A user can start recording with a tap. There’s no meeting bot, on-screen notice, or need to keep a laptop open. That makes the product convenient, but it also removes cues that would tell other people a recording is underway.

The shift showed up in other settings that same week. DEF CON barred Meta-style smart glasses from its conference floor because of recording concerns. Meta had acquired Limitless, a company making AI pendants, bringing always-on personal recording into a major consumer hardware operation. And one public company the authors have worked with now checks attendees for smart glasses before all-hands meetings where unreleased plans are discussed.

The authors also describe testing the watch themselves: worn beneath a jacket sleeve, it captured an entire evening walk and produced a clean transcript. No one around them could tell. That example points to the central policy problem. In the United States, rules on recording without permission vary by state. Yet disclosure practices are unsettled, while wearable hardware is getting less conspicuous.

Organizations can no longer assume that a meeting is unrecorded because no bot joined the call or no recording indicator appeared on screen. A device may be recording a conversation in a conference room, on a walk, or during an informal lunch. Policies built around meeting software alone will miss those situations.

The transcript can be shaped or wrong

AI notes don’t simply preserve what happened. People have described ways to influence them: prompt injection during a live meeting, joining early and narrating preferred takeaways while alone with the notetaker, or triggering a silent failure that leaves no record. The authors say academic research also supports the effectiveness of these techniques. Whether participants view such tactics as manipulation or strategic communication, the result is the same: the transcript can reflect someone’s effort to steer it.

Errors can arise without anyone trying to interfere. The authors have seen a system mishear a word and then interpret the rest of the discussion to fit that mistake. People whose accents are harder for a model to process may face a higher risk of inaccurate notes. That matters when a transcript is used to assign work, settle a disagreement, or document a decision. A confident-looking summary can still be wrong.

Some products also draw on memory from earlier meetings and background information about participants. Granola, for example, can use prior meeting context in notes for a current conversation. That may help users avoid repeating themselves, but it can also expose internal reasoning when a summary goes outside the company. A recipient could see information drawn from discussions they never attended.

A case described by the authors shows why companies should look beyond notetaker software. Last November, a Campbell’s security executive was terminated after an employee secretly recorded a lunch conversation the executive believed was private. The recording surfaced months later through an employment lawyer, and the executive was gone within days. The authors’ point is about the recording mechanism: there was no meeting bot or platform feature involved. One person used one device, and the resulting audio became consequential later.

Set rules for the record, not just the app

Boards and legal teams are asking CISOs to establish policies for AI notetakers. The authors’ recommendations start with a clear disclosure rule: everyone present should know when a recording tool is active. One simple option for virtual meetings is to add “+ AI notetaker” to the user’s display name. The same standard should apply when employees record one another. The authors argue that undisclosed employee recording should be grounds for dismissal, regardless of what the recording contains.

Companies should also decide in advance which conversations cannot be recorded. One organization described by the authors formally excludes M&A discussions, legal and HR meetings, and board executive sessions. Each company will have its own sensitive categories, but a written list gives employees a rule to follow before a recording tool makes the choice for them.

Rules need to cover devices and tools brought in by outside organizations, too. If an auditor’s laptop must meet company standards, the same scrutiny should apply to recording. When another organization records employees, the company should know about it and receive a copy of the resulting record.

Retention needs an explicit choice. A full transcript carries different information from a short summary, so the two shouldn’t be treated as interchangeable corporate records. Companies should decide whether transcripts are deleted by default, subject to legal holds, or kept indefinitely. Leaving that decision to a vendor’s default setting means a product choice is quietly setting company policy.

Vendor review should reflect the sensitivity of the material these companies process. AI notetaker providers may hold confidential customer discussions, employee matters, and internal plans in their cloud systems. Security checks should examine data reuse and contract terms, including indemnification. The authors cite Brewer v. Otter.ai, a class action alleging that Otter recorded meetings without everyone’s consent and used recordings to train models, as well as Novacy’s sudden shutdown.

Finally, every transcript that becomes an official record needs a named human owner. That person should check whether the notes accurately reflect the conversation, regardless of whose tool generated them. Some employees may be tempted to steer the system toward a preferred account of events, much as a lawyer pays close attention to a deposition record. Reviewing for accuracy is sensible; quietly shaping the record is a warning sign that the organization has a deeper trust problem.

The policy question used to be whether to let a notetaker into a meeting. Wearable recorders make that question too narrow. Companies need rules that apply wherever a microphone can go, with clear disclosure, defined limits, deliberate retention, and a person accountable for the record. The practical assumption for sensitive conversations should be that a recording device may already be present.