Cyberpediaby Brain

AI meeting note-takers: recording and automatic sharing

AI bots can join a Teams meeting, record it and store its content outside the company's compliance framework.

On this page
  1. 1In 30 seconds
  2. 2What it is
  3. 3Straight answers
  4. 4How it works
  5. 5What you do
  6. 6Pitfalls
  7. 7Real cases
  8. 8Sources
  9. 9Cite this article

In 30 seconds

Where it happens
In a Teams meeting, when an external participant detected as a bot is waiting in the lobby[1].
The risk
A bot may record or transcribe without the participants knowing, and store the meeting data in third-party systems, outside the organization's compliance boundaries[1].
The reflex
A participant detected as a bot stays in the lobby until someone explicitly admits it[1]. Do not admit it if you do not know which tool it is or who launched it, and ask which tool is approved in your company.

What it is

External assistant tools for Teams, such as transcription or note-taking bots, are increasingly used in meetings[1]. Microsoft acknowledges that they may improve individual productivity, but adds that they may also introduce security and compliance risks when they access a meeting without the organizer's awareness or consent[1].

This article covers that category of tools without naming any particular one. It starts from what the employee sees: an unexpected participant in the lobby, a meeting that is recorded or transcribed, a summary produced by an online service.

Straight answers

What is an external AI note-taker?

Microsoft uses this term for meeting assistant tools that sit outside the organization, for example transcription or note-taking bots[1]. According to Microsoft, these bots may record or transcribe meetings without participants being aware of it, store meeting data in third-party systems outside the organization's compliance boundaries, and introduce compliance, privacy and data leakage risks[1].

Do participants have to agree?

Teams offers an explicit agreement rule, which the administrator turns on through a policy. It is off by default: participants are then not prompted, and all of them are included in the recordings and transcriptions of the meetings of the organizers concerned[2].

This setting covers recording and transcription by Teams. External bots fall under a different meeting policy setting, bot detection[1].

When the rule is on and someone starts recording or transcription in a meeting that is not recorded automatically, the other participants are muted, and their camera and content sharing are turned off. The person who starts the recording is considered to have agreed[2]. To speak again, turn the camera on or share content, each participant is asked to answer Yes or No. A participant who answers No has a read-only experience[2]. Each person's choice appears in the attendance report[2].

What does ANSSI say about sensitive data?

In its security recommendations for a generative AI system, ANSSI writes that sensitive data must absolutely not be sent to third-party generative AI services[3]. It counts among such data Restricted Distribution or classified information, research work covered by the PPST, personal data, a company's contractual, legal and financial data, and IT secrets such as passwords or API keys[3]. Its reasoning: the client entity does not control the service and cannot make sure that the confidentiality of the data sent meets its requirements[3].

ANSSI also notes that some third-party generative AI tools offer connections to business applications, and that particular attention must be paid to their access rights to the entity's data: email, document spaces, source code repositories, audio and video conferencing services[3]. It recommends reviewing these rights as soon as the product is activated, then regularly, for example every month[3].

What happens to what you send to an online AI service?

According to ANSSI, sending information to a publicly accessible generative AI service amounts to placing it in a storage space owned by the provider, and with most services, the data sent is collected and used by the provider to optimize its models[3].

The NCSC (UK) writes, in a post published on 14 March 2023, that a large language model does not automatically add queries to its model for others to query. But the query is visible to the organization providing the model, it is stored, and it will almost certainly be used at some point to develop the service or the model[4]. The NCSC recommends not including sensitive information in queries to public models, and not submitting queries that would cause problems if they became public[4].

Has the CNIL set rules for these tools?

The CNIL's 2026 work programme, presented on 7 April 2026, announces upcoming work, not rules in force. The CNIL states that it will launch work to clarify the conditions under which automatic transcription and analysis tools, notably in call centres or videoconferencing software, may be deployed, based on use cases that match actors' practices[5]. It specifies that the programme is indicative and may change[5].

Who is responsible if the tool is misused?

The CNIL states that, in the general case, it is the user organization that bears legal responsibility if its staff misuse AI[6]. It recommends framing the use of generative AI with internal policies or charters that define authorized and prohibited uses[6]. End users should submit only information they are allowed to share, for example never sharing personal data or company data with a consumer service[6].

How it works

Teams detects external bots from signals collected when the participant joins the meeting, of an infrastructural and behavioral nature[1]. A participant detected as a bot is placed in the lobby, whatever the lobby configuration of the meeting[1]. The organizer, and depending on the meeting settings the presenters, must explicitly admit it, and Teams reminds them of the risk of admitting a bot[1].

The administrator sets this behavior through a meeting policy. The default option requires approval when a bot is detected. The other option does not detect bots, which then appear like any other external participant[1]. Microsoft recommends keeping the default option and avoiding turning detection off unless a specific business need requires it[1].

Recording and transcription fall under other settings. In the global meetings and events policy, they are on by default[7]. If the administrator allows automatic recording, the organizer's setting is off by default for meetings and on for events[7]. The administrator can also allow organizers to restrict copying or forwarding of the chat, captions, transcripts and AI-generated items in recaps[7].

What you do

  • If a participant is marked as a bot, it stays in the lobby until someone admits it[1]. If you do not know which tool it is or who invited it, do not admit it.
  • If a colleague is mistaken for a bot, admit them from the lobby and choose the "This is not a bot" option: they then appear as an ordinary participant for that meeting[1].
  • If a bot was not detected, Microsoft invites you to report it through the Teams feedback option[1].
  • If you organize, ask your administrator to configure meetings so that only organizers and co-organizers admit participants from the lobby. Microsoft gives this setting as a tip: a presenter then cannot admit a participant without the required scrutiny[1].
  • Warn participants before starting a recording or transcription: the risk Microsoft describes is precisely recording without participants knowing[1].
  • For a meeting that deals with sensitive data, do not hand it to an external bot: ANSSI writes that such data must not be sent to third-party generative AI services[3]. With Teams Premium, sensitivity labels or a meeting template can control who may record and transcribe[7].
  • Ask which tool is approved and what the internal charter allows, since the CNIL recommends clearly defining authorized and prohibited uses[6].

Pitfalls

Detection is imperfect. Microsoft states that some external bots may not be detected, and that detection may occasionally classify a human participant as a bot[1].

Recording settings do not carry over from one policy to another. Microsoft gives the example of a finance team from which the administrator removes recording and transcription in the calling policy. With no meeting or events policy assigned, the team stays on the global policy, where they are on: it can record and transcribe meetings, group calls and events, but not 1:1 calls[7].

In a meeting that requires explicit agreement, a participant who joins from an unsupported device has a read-only experience[2].

A summary produced by AI may be wrong while looking plausible. The CNIL states that generative AI systems may produce inaccurate results that look plausible, often called hallucinations[6]. It describes never reproducing the outputs of these systems as they are as good practice[6].

Real cases

Samsung, late March 2023: a meeting recording sent to ChatGPT

According to ESET, the South Korean outlet The Economist Korea reported three independent incidents at Samsung Electronics in late March[8]. The company had asked its employees to be careful about the information they entered in their queries, but some accidentally disclosed internal data while using ChatGPT[8]. One employee entered source code tied to a measurement database for a semiconductor facility, another a program for identifying defective equipment, and a third sent recordings of a meeting to get the minutes[8]. Samsung announced that it planned to develop its own internal AI service to help its employees with their tasks[8].

Sources

  1. Microsoft, “Manage external bots and their access to meetings hosted in your organization”. https://learn.microsoft.com/en-us/microsoftteams/manage-external-bots (accessed 14 November 2026)
  2. Microsoft, “Manage participant agreement for recording and transcription”. https://learn.microsoft.com/en-us/microsoftteams/participant-agreement-recording-transcription (accessed 14 November 2026)
  3. ANSSI, “Security recommendations for a generative AI system”, 29 April 2024. https://messervices.cyber.gouv.fr/documents-guides/security_recommandations_for_a_generative_ai_system.pdf (accessed 14 November 2026)
  4. NCSC (UK), “ChatGPT and large language models: what's the risk?”, 14 March 2023. https://www.ncsc.gov.uk/blog-post/chatgpt-and-large-language-models-whats-the-risk (accessed 14 November 2026)
  5. CNIL, “Accompagnement des professionnels : le programme de travail de la CNIL pour 2026”, 7 April 2026. https://www.cnil.fr/fr/accompagnement-des-professionnels-le-programme-de-travail-de-la-cnil-pour-2026 (accessed 14 November 2026)
  6. CNIL, “Les questions-réponses de la CNIL sur l’utilisation d’un système d’IA générative”, 18 July 2024. https://www.cnil.fr/fr/les-questions-reponses-de-la-cnil-sur-lutilisation-dun-systeme-dia-generative (accessed 14 November 2026)
  7. Microsoft, “Overview- Recording and transcription for Teams meetings, events, and calls”. https://learn.microsoft.com/en-us/microsoftteams/recording-transcription-overview (accessed 14 November 2026)
  8. ESET, “Meet “AI”, your new colleague: Could it expose your company's secrets?”, 17 May 2023. https://www.welivesecurity.com/2023/05/17/meet-ai-new-colleague-work-company-data/ (accessed 14 November 2026)