An AI note-taker promises that nobody has to take notes during the meeting. It records the conversation, turns it into text, and at the end sends out a summary and a list of action items. That is why employees often switch one on themselves, with a free account, in a few clicks. From then on an outside service sits in on the meetings, and the audio recording, the transcript, and the summary stay on the provider's servers.
The general issues of AI tools used without approval are covered in the article on shadow AI.
How an AI note-taker gets into a meeting
The most common route is calendar integration. At sign-up, the user gives the service access to their calendar, and from there the service sees every event that contains a meeting link. With many tools the default is for the note-taker to join every such meeting automatically, even when the user is running late or does not attend at all. On the meeting platform the note-taker shows up as a separate participant, typically with the user's name and a suffix such as “note-taker” or “notetaker”.
Calendar access covers every event the user is invited to, including meetings someone else organized. The organizer may be a partner, a client, or an authority, and the invitee who switched on the note-taker does not necessarily ask them. Some services also send the summary to the other participants after the meeting, along with a link inviting them to sign up. Anyone who signs up through that link and connects their own calendar brings the tool into their own meetings, possibly in another organization.
A note-taker can also get in other ways:
- Built-in feature: the meeting platform's own recording, transcription, and AI summary. The platform administrator manages it centrally, and participants can usually see that recording is running.
- Desktop or mobile app: records the audio of the computer or phone and does not join as a participant. The others see nothing of it, and it works the same way in an in-person meeting in a conference room.
- Browser extension: reads the audio and captions of a meeting running in the browser, and often has access to the rest of the page content as well.
If the employee granted calendar access with their organizational account, the application that requested read access to the calendar appears on the identity platform as a user consent. The same list shows how many employees have connected their account to that service. A note-taker used with a personal calendar, installed as an extension inside the meeting platform, or run as a desktop app does not necessarily leave a trace there.
What it records: audio, transcript, summary, screen
The material a note-taker produces has several layers, and each carries a different risk:
| Layer | What goes into it | What to watch for |
|---|---|---|
| Audio recording, sometimes video | The whole meeting, with the participants' voices and images | The most personal data in one place, and the voice sample can also be used to identify the speaker |
| Transcript | Verbatim text, by speaker and with timestamps | Searchable and copyable, and side remarks stay in it too |
| Summary, action items | An extract written by a large language model (LLM) | It can misread a statement, leave it out, or attribute it to the wrong person |
| Screenshots, shared content | Images of the projected slides, spreadsheets, and documents | Data gets captured that was never said out loud |
Speaker separation works from the audio recording: the service determines which voice is speaking when and assigns a name to it from the participant list. Some tools also build a voice profile so they can recognize the speaker in later meetings. If the purpose is to uniquely identify a person from the voice sample, that is biometric data under Article 4(14) of the GDPR, which Article 9 allows to be processed only under narrow conditions.
The model writing the summary can hallucinate: it may turn an unfinished discussion into a firm decision, or return a question as a statement. If the summary goes out to the participants automatically, the wrong version becomes the official record of the meeting. Some tools also score the participants' mood or engagement. If such a feature infers employees' emotions from tone of voice or facial expressions, Article 5(1)(f) of Regulation (EU) 2024/1689 (AI Act) has prohibited it in the workplace since February 2, 2025, except for medical or safety reasons.
Where it is stored and who can access it
Standalone note-taker services store the material in their own cloud, outside the organization's systems. Where it is stored, how long it is kept, and whether the provider uses the recordings to develop its models are set out in the terms of use and the privacy notice. On a free or personal plan, the employee accepts these in their own name. As a result, the organization has no data processing agreement with the provider, although Article 28 of the GDPR would require one if the provider processes personal data on the organization's behalf. If the data is stored outside the European Economic Area, the transfer conditions in Chapter V of the GDPR must also be met.
Access is typically tied to the user who switched on the note-taker. That user sees all the recordings and decides who to share them with. Common ways of sharing:
- an automatic email with the summary to every invitee, external participants included;
- a link to the full recording and transcript, which with many services anyone holding the link can open;
- a shared workspace where team members can also search each other's meetings;
- an integration that forwards the summary or action items to other systems, for example a CRM, a task manager, or a chat channel.
When the employee leaves, their personal account and every recording in it leave with them. The organization then cannot delete the material and cannot even say what is left of it. A data subject request raises a similar problem, when someone asks under Article 15 of the GDPR what data is processed about them. If the recording is part of the organization's processing, the answer should cover it too, but the organization does not know about it. Whether a recording in an employee's personal account counts as the organization's processing is a contested question.
Informing the other participants
Recording a meeting means processing the other participants' personal data. Under Article 6 of the GDPR this needs a legal basis, and under Article 13 the data subject must be informed before the recording about its purpose, the storage, and their rights. Section 2:48 of the Hungarian Civil Code (Ptk.) separately requires the data subject's consent to make an audio recording. If the employer introduces the tool or uses the recordings to monitor employees, Section 11/A of the Hungarian Labor Code also applies: the employer may monitor only conduct related to the employment relationship, and must inform employees in advance about the tool it uses.
In practice this comes down to a few habits:
- the meeting invitation says if the meeting is recorded or transcribed by an AI tool, and who handles the material;
- at the start of the meeting the organizer also says it out loud and asks whether anyone objects;
- anyone who does not want their voice recorded has a real way to avoid it, for example by pausing the recording or contributing in writing;
- with external participants, the organizer asks even if a colleague, and not the organizer, switched on the note-taker.
A bot in the participant list is not notice in itself. Many people do not notice it, or do not know what the suffix next to the name means. For a recording made with a desktop or mobile app, spoken or written notice is the only option left.
Approving, blocking, and default settings
The organization's first decision is which tool to allow. In many places the built-in platform feature is the obvious choice, for the storage reasons described in the note above. If a standalone service is needed, it is best rolled out with an enterprise plan, a data processing agreement, and single sign-on (SSO), so that the accounts belong to the organization.
Keeping unapproved tools out
- Application consent: on the identity platform you can require that only an administrator may grant consent to external applications requesting read access to calendars or mail. Consents already granted can be revoked, which also cuts the note-taker off from the calendar.
- Meeting platform: external or unauthenticated participants can be sent to a waiting room, where the organizer decides whether to admit them. Most note-takers join as a bot, as a participant from outside the organization, so they get stuck in the waiting room. The organizer can also remove a bot that is already in the meeting.
- Web filtering: the domains of unapproved note-taker services can be blocked on the network and in the managed browser, and browser extensions can be restricted with an allowlist. The technical details are covered in the article on controlling AI services with technical tools.
Blocking protects the organization's own meetings. If a partner sends the invitation and a note-taker runs on their side, your own settings do not reach it. In that case all you can do is ask and object.
Default settings for the approved tool
- the note-taker should not join every meeting automatically, only when the user turns it on for a given event;
- the summary should not go out automatically; the organizer reviews and corrects it first;
- sharing should default to named access within the organization, with no public link;
- the retention period for audio recordings should be short if the goal is only the transcript and the summary;
- voice profile building and sentiment analysis should be turned off;
- if the provider allows it, submitted material should not be used to further train the model.
Rules in the organization's AI usage policy
Note-takers deserve their own section in the AI usage policy, because here the data that goes into an AI tool belongs to the other participants rather than to the employee. A rule is easy to follow when it is concrete and short. What it typically covers:
- Approved tools: listed by name, with a corporate account made mandatory. Every other note-taker is prohibited, including those used with a personal account.
- Meetings not to be recorded: for example HR, disciplinary, and health-related discussions, legal advice, acquisition or tender negotiations, and security incident handling. What kind of data should not go into an AI tool in general is summarized in the article on sensitive data.
- Notice and objections: who is responsible for informing the participants, and what happens if someone does not consent.
- Handling the summary: an AI-written extract may go into minutes or a decision document only after review, and must be marked as a machine-generated summary.
- Retention and deletion: how long recordings, transcripts, and summaries are kept, and who deletes them once the purpose has been met.
- External meetings: employees may bring their own note-taker into a partner's meeting only with the organizer's explicit consent, and if a note-taker runs on the partner's side, they should speak up before a confidential topic comes up.
Compliance with the rule can be checked from technical traces: alongside the consent list described in the first section, the web filter log shows which blocked note-taker services people still try to reach. If a blocked tool turns up for many users, it often means the approved solution is cumbersome or not well known, and the rollout or the training needs another look.
Related service: Shadow AI assessment with an AI usage policy →