Key takeaways
- For most meetings, an AI summary without speaker attribution is enough. Choose the method that intrudes least on participants' rights and still serves the purpose. Audio and video recordings remain the exception.
- You can usually base transcripts and summaries on legitimate interest if you inform people in advance and offer a simple opt-out. For recordings, you usually need consent under data protection law. In Germany, Section 201 of the Criminal Code (StGB) also requires the agreement of all participants when a conversation is recorded.
- Consent from employees is rarely reliable because of the dependent relationship. You should not record job interviews.
- As soon as artificial intelligence (AI) evaluates people, additional limits apply: emotion recognition for employees and applicants is prohibited, AI assessments of applicants and employees are usually high-risk AI, and AI must not make decisions about people on its own.
- Use the data processing agreement (DPA) and the settings of the tool you use to rule out that the provider may use meeting content for AI training or other purposes of its own.
AI meeting notes, transcripts, or recordings: AI notetakers and the built-in features of Teams, Zoom, or Google Meet now deliver all of this at the push of a button, plus task lists. Some bots even join every calendar meeting automatically. As a privacy expert, I get the same questions about this almost every week: May we record and transcribe meetings? Do employees, customers, or applicants have to consent, or can we do without it? How long may we keep recordings and transcripts? And what difference does it make when AI is involved?
The answers depend mainly on three things: what is technically created, what you use it for, and whose data is affected. Our GDPR overview explains the general principles, from the six legal bases to data subject rights.
Which method is enough: summary, transcript, or recording?
Summaries, transcripts, and recordings intrude on participants' rights to different degrees. How far a method intrudes depends mainly on these factors: whether anything is stored at all, how complete it is (verbatim or outcome only), whether contributions are attributed to a person, whether voice, and for video also image, surroundings, and screen content, are preserved, and how easily the material can be searched, shared, and analyzed. The more of these apply, the higher the legal bar. The following table compares the common methods by intrusion, suitability, and risks.
| Method | Intrusion | Benefits and suitability | Drawbacks and risks |
|---|---|---|---|
| Live transcription without storage | Low | Captions for people with hearing impairments and participants who speak another language. Nothing is stored permanently. | No record. Unsuitable for documentation. |
| Summary without speaker attribution | Low to medium | Minutes with topics, decisions, and action items. Sufficient for most meetings. | AI can condense incorrectly. Depending on the tool and its settings, a full transcript may also be stored in the background. |
| Transcript or summary with speaker attribution | Medium to high | Exact wording, traceable record of who said or promised what. Useful for negotiations, interviews, or requirements workshops, for example. | Every statement is attributed to a person and searchable. This allows conclusions about behavior and opinions and can therefore also be used to monitor performance. |
| Audio recording | High | Complete playback of what was said, for absent colleagues, talks, and training sessions where the audio itself is needed. | Voice, tone, and every word are preserved. Risk of misuse, for voice cloning for example, and high effort for access requests and deletion. |
| Video recording | Very high | Only needed when the image itself is needed, for demonstrations, product demos, or training sessions, for example. | Also captures appearance, facial expressions, and gestures, private rooms when working from home, and everything on shared screens, such as customer data or notifications. High risk of misuse, for further analysis and deepfakes, for example. |
In practice, this means: for most meetings, a summary without speaker attribution is enough. A transcript with speaker attribution is only necessary when the exact wording matters and who said what. A recording should be the exception, for content that is genuinely needed as audio or video.
Check what the tool stores in the background. Depending on the tool and its settings, the summary is generated from a full transcript with speaker attribution, which then remains stored. The lower intrusion of a summary only counts if no transcript is actually kept.
Which legal basis applies to what?
In practice, two legal bases are relevant for transcribing and recording meetings:
- legitimate interest (Art. 6(1)(f) GDPR) for transcription and summaries
- consent (Art. 6(1)(a) GDPR) for audio and video recordings
Other legal bases usually only apply in special cases. You can find the wording of every article in the text of the Regulation on EUR-Lex. The following overview shows when each legal basis applies and where its limits lie.
| Legal basis | When it applies | Limits |
|---|---|---|
| Consent Art. 6(1)(a) GDPR | Audio and video recording | Must be freely given, informed, specific, and given before the start (Art. 4(11), Art. 7 GDPR). Can be withdrawn at any time; the company bears the burden of proof (Art. 7(1), (3) GDPR). Rarely reliable for employees. |
| Performance of a contract Art. 6(1)(b) GDPR | For holding the meeting itself. For recording, only conceivable in exceptional cases, for example when the recording is part of the agreed service. | Usually does not cover transcription or recording, because meetings can take place without them. |
| Legal obligation Art. 6(1)(c) GDPR | A statutory recording duty, e.g. for investment services in Germany (Section 83(3) WpHG). | Rarely relevant in ordinary meetings. |
| Legitimate interest Art. 6(1)(f) GDPR | Transcripts and summaries, for employees and external participants alike | Only after a documented balancing test and with a right to object (Art. 21 GDPR). |
| Explicit consent Art. 9(2)(a) GDPR | When special categories such as health data are discussed. | Required in addition to Art. 6. Better not to record or transcribe such meetings. |
Purpose and data determine the assessment. The same recording can be lawful or unlawful, depending on what you use it for. It makes a difference whether you record a product demo for customers who cannot attend, or use the same recording to assess the sales rep's behavior and performance.
When does legitimate interest apply?
Legitimate interest only applies when three conditions are met: a legitimate interest, the necessity of the processing for that interest, and a balancing test in which the interests and fundamental rights of the data subjects do not prevail.
1. Legitimate interest. Efficient and reliable documentation, traceable agreements, and accessibility are legitimate interests. This step is rarely the problem.
2. Necessity. Is there a less intrusive means that is equally effective? For minutes or a meeting summary, necessity can usually be shown. A verbatim transcript, on the other hand, is only necessary when the exact wording matters. Switched on across the board for every meeting, transcription is generally not necessary, so assess necessity case by case. For an audio or video recording, the case usually fails at this step already, because a transcript or even a summary without speaker attribution serves the usual purposes just as well.
3. Balancing. Do the interests and fundamental rights of the data subjects prevail? What matters is what the data subjects can reasonably expect in the situation (Recital 47 GDPR), how sensitive the content discussed is, and the relationship between the data subjects and the controller. Safeguards can reduce the intrusion and ultimately tip the balance in the controller's favor.
Effective safeguards weigh in the company's favor, in particular:
- transparent advance information, for example in the invitation and when joining the meeting,
- a visible indicator that transcription is running, for example an icon or a clearly named bot among the participants,
- a simple opt-out without having to give reasons,
- short retention periods,
- a narrow circle of people with access,
- no speaker attribution, and
- a binding commitment not to use transcripts to monitor performance or behavior.
The following weigh in the data subjects' favor, in particular:
- confidential or sensitive content, for example in HR, salary, or conflict discussions,
- a relationship of dependence, for example with employees or applicants,
- transcription without a specific reason, for example because the bot joins every meeting automatically,
- further analysis, for example of speaking time or conversational behavior,
- use of the transcripts for other purposes, for example to train the provider's AI models, and
- disadvantages that data subjects may later suffer because of what they said, for example in employment disputes.
Document the three-step balancing test, ideally directly in your record of processing activities, next to the processing it covers.
When is consent valid?
Consent only works as a legal basis if it was validly given. The conditions follow from Art. 4(11) and Art. 7 GDPR, and all of them must be met.
1. Freely given. Data subjects must have a genuine choice and be able to say no without fearing disadvantages (Recital 42 GDPR). Taking part in the meeting must therefore not depend on consenting to the recording if the meeting can also take place without one (prohibition of tying, Art. 7(4) GDPR). For employees and applicants, this condition is the main problem (more on this below).
2. Specific and informed. Consent must relate to a specific purpose. Blanket consent for all future meetings, for example in terms and conditions or in the employment contract, is not enough. If you pursue several purposes, such as follow-up for absent colleagues and later use as training material, data subjects must be able to agree to each purpose separately. For consent to be informed, data subjects should know at least the following from the first layer of information: who the controller is, which categories of data are processed for which purposes, and that they can withdraw their consent at any time with effect for the future.
The remaining mandatory information under Art. 13 GDPR, in particular on recipients, transfers to third countries, retention periods, and data subject rights, can usually be provided on a second layer, for example in a linked privacy notice that is easy to access before consent is given.
3. Unambiguous and before the start. Consent requires a clear affirmative act (Recital 32 GDPR). Silence or simply carrying on is not enough. Anyone who stays in the meeting after hearing "We're recording" has therefore not consented under data protection law. Suitable forms are a click on "Agree" in the tool, a yes in the chat, or an explicit spoken yes that is documented. Consent must be in place before the recording starts. This also applies to people who join later.
4. Demonstrable and revocable. You must be able to show who consented to what and when (Art. 7(1) GDPR), for example through the tool's automatic consent log or the chat history. Consent can be withdrawn at any time, and withdrawing must be as easy as giving it (Art. 7(3) GDPR).
Decide on the legal basis before processing starts. According to the European Data Protection Board (EDPB, Guidelines 05/2020), if you base a recording on consent, you cannot retroactively switch to legitimate interest when consent is withdrawn or turns out to be invalid.
What applies to health data and voice?
If special categories under Art. 9(1) GDPR come up, such as health data, you also need an exception under Art. 9(2) GDPR, in practice usually explicit consent. My recommendation: exclude such meetings from transcription and recording.
A question that comes up often is whether a voice alone is biometric data under Art. 9 GDPR. Usually not. A voice recording only becomes biometric data when it is processed with specific technical means to uniquely identify a person or confirm their identity, for example through a voiceprint (Art. 4(14) GDPR). Simply recording and transcribing a meeting does not fall under this. The usual speaker attribution in online meetings also works through participants' user accounts. It is different when a system recognizes people from stored voice profiles, as some meeting room solutions offer. In that case, it processes biometric data for identification, and you need an exception under Art. 9(2) GDPR.
Does it matter who is in the meeting?
Which legal basis applies depends not only on the method you choose but also on who is in the meeting. The relationship between the company and the data subjects matters in two places. It shapes what data subjects can reasonably expect, and with it the balancing test for legitimate interest. It also determines whether consent can be freely given at all. A customer can easily say no. For employees in front of their managers, or applicants in a job interview, that is much harder.
Why is employee consent so rarely reliable?
Consent is only valid if it is freely given (Art. 4(11) GDPR). That is exactly where it often falls down with employees and applicants. People who depend on their employer, or who want to get a job, do not say no lightly. Whether consent is freely given must therefore be assessed strictly (Recital 43 GDPR). In meetings this is particularly obvious: when the team lead asks the group at the start of a call whether everyone is fine with recording, hardly anyone objects in front of managers and colleagues. A yes given under that pressure is a questionable basis.
Freely given consent is still not ruled out. The German Federal Data Protection Act (BDSG) names as indicators a legal or economic advantage for the employee or aligned interests of employer and employee (Section 26(2) sentence 2 BDSG). The EDPB also considers consent in the employment relationship possible when saying no has no consequences at all (Guidelines 05/2020).
Beyond the voluntariness test, Section 26(2) BDSG goes further than the GDPR in two respects: In Germany, consent in the employment relationship must be given in writing or electronically. Spoken consent is only valid in special circumstances. The employer must also inform the employee in advance, in text form, about the purpose of the processing and the right to withdraw consent under Art. 7(3) GDPR.
A free decision in a meeting is more likely when
- the initiative comes from the employees or the recording mainly benefits them,
- there is an equivalent alternative, for example taking part without camera and microphone or written minutes instead of a recording,
- consent is asked for in advance and individually rather than live in front of the group,
- the recording clearly does not serve performance assessment, and
- saying no has no consequences.
Even so, consent often does not work for standard processes. Anyone who declines still has to attend, because the team meeting is part of the job. A single no therefore prevents the recording for everyone. Then there is the effort: you must obtain everyone's consent and be able to prove it (Art. 7(1) GDPR). If someone withdraws consent (Art. 7(3) GDPR), you may no longer store their contributions. You then have to cut them out of the recording or delete the whole recording. In day-to-day work, that is hardly feasible. For recurring meetings, the better route is therefore a transcript or summary based on legitimate interest.
Exactly where the line runs has not been conclusively settled. Supervisory authorities tend to be strict, while parts of the legal literature see more room. A useful guide is the question of whom the recording serves. The more it benefits the employees themselves and the easier it is for them to say no, the more likely consent is to hold. The more it is part of the mandatory routine and serves the company, the less likely.
Regardless of the legal basis, an existing German works council (Betriebsrat) has a right of co-determination when you introduce or use technical equipment that is suitable for monitoring employees' behavior or performance (Section 87(1) no. 6 of the Works Constitution Act, BetrVG). According to the Federal Labor Court, objective suitability for monitoring is enough. The employer's intention does not matter. Transcription and recording features that attribute contributions to individuals usually meet this condition. Involve the works council before you introduce them, typically through a works agreement.
May you record job interviews?
In a job interview, the imbalance of power is large. Someone who wants the job rarely says no. Consent is therefore hardly reliable. I advise against recording job interviews and, at most, creating a summary without verbatim wording. Anyone thinking about also letting AI assess applicants should also keep the obligations and limits of the EU AI Act and Art. 22 GDPR in mind (see below).
What applies to customers, leads, business partners, and other external participants?
You can usually base transcripts and summaries on legitimate interest if you inform people transparently in advance and make it easy to object. For a recording, on the other hand, you obtain consent. For your own employees in the same meeting, assess the legal basis separately, as described above.
As a rough rule of thumb:
| Employees | Applicants | Customers, leads, business partners, and other external participants | |
|---|---|---|---|
| Transcript / summary | Legitimate interest. Inform in advance and allow objection. | Legitimate interest. Only by exception, at most a summary without verbatim wording. | Legitimate interest. Inform in advance and allow objection. |
| Audio / video recording | Consent. Only valid if all conditions are met, with particular attention to whether it is freely given. Otherwise do not record and switch to a less intrusive method. | Do not record. | Consent. Only valid if all conditions are met. |
Additional limits apply to AI analysis of people, for example of sentiment. More on this in the AI section.
What does German criminal law (Section 201 StGB) require when you record a meeting?
Section 201 of the German Criminal Code (StGB) protects the confidentiality of the spoken word: anyone who, without authorization, records another person's privately spoken words faces up to three years in prison or a fine (Section 201(1) no. 1 StGB, see the English translation of the Criminal Code). For a recording, you therefore need the agreement of all participants. For live transcription, it has not been conclusively settled whether the technically necessary buffering of the audio signal already counts as a "recording" within the meaning of Section 201 StGB.
Unlike consent under data protection law, agreement under criminal law can also be given implicitly. Anyone who was clearly informed in advance and takes part without objecting usually agrees. For transcription, the model of transparent advance information, a visible indicator, and a simple way to object usually covers this. For recordings, the stricter consent under data protection law still applies.
What else applies when AI creates the meeting notes?
Almost every transcript and every set of AI meeting notes is now created with AI. That changes little in the data protection assessment above, as long as the AI only transcribes and summarizes. Do take the use of AI into account in the threshold assessment for a data protection impact assessment (DPIA), though. AI usually counts as a new technology within the meaning of Art. 35(1) GDPR and therefore meets one of the nine criteria supervisory authorities use to decide whether a DPIA is needed.
Under the EU AI Act, such a notetaker is generally not a high-risk system. As a deployer, your main duty is to promote the AI literacy of the employees who work with it (Art. 4 AI Act). How to organize this and your other deployer obligations is covered in our guide on how to prepare your business for the EU AI Act.
Additional obligations arise when the provider uses the content for its own purposes, when the AI analyzes or evaluates people ("profiling"), and when decisions about people are made on that basis. You rule out use by the provider in the DPA. What to look for is described in our article on how to review a data processing agreement.
The following overview classifies the common AI features:
| AI feature | Classification | What you need to do |
|---|---|---|
| Transcript and summary | Not a high-risk system. | Promote AI literacy (Art. 4 AI Act), check transcripts and summaries. |
| Use of the content by the provider, for example for AI training | Processing for the provider's own purposes (Art. 28(10) GDPR) | Rule it out in the DPA and in the settings. |
| Emotion recognition for customers, business partners, or other external participants | High-risk AI (Annex III no. 1(c) AI Act) | Inform the people affected (Art. 50(3) AI Act), carry out a DPIA, and keep the deployer obligations under the AI Act as well as profiling and automated decision-making (Art. 22 GDPR) in mind. |
| Assessment of applicants, assessment of employees' performance or behavior | High-risk AI (Annex III no. 4 AI Act) | Inform the people affected under Art. 26(11) AI Act (this obligation applies from December 2, 2027), carry out a DPIA, and keep the deployer obligations under the AI Act as well as profiling and automated decision-making (Art. 22 GDPR) in mind. |
| Emotion recognition from voice, facial expressions, or tone for employees and applicants | Prohibited (Art. 5(1)(f) AI Act) | Switch off the feature or do not use the tool. |
| Decisions about people based solely on the AI analysis | Automated decision-making that is generally not permitted (Art. 22 GDPR) | Ensure a meaningful human review. |
The duty to inform people about emotion recognition is one of the transparency obligations under Art. 50 AI Act, which have applied since August 2, 2026. What deployers have to disclose is explained in our article on Article 50 of the EU AI Act.
Regulation (EU) 2026/1744 postponed the high-risk obligations for Annex III systems to December 2, 2027. That leaves time to identify the systems affected. Which applications fall under it and which exceptions exist is covered in our overview of high-risk AI systems under the EU AI Act.
How long may you keep recordings and transcripts?
"As long as necessary, as short as possible." There is no fixed statutory period. How long recordings, transcripts, and meeting summaries may be kept depends mainly on the method used and the purpose pursued, and has to be assessed case by case.
| What? | Criteria for setting the retention period |
|---|---|
| Transcript with speaker attribution | As soon as the exact wording with speaker attribution is no longer needed, it should be shortened to the outcome or summarized, or deleted. Examples:
|
| Meeting summary without speaker attribution | Because the summary does not attribute contributions to individuals, a longer retention period is easier to justify than for a transcript. It should be deleted as soon as decisions, action items, and agreements are no longer needed and no retention obligation applies.
|
| Recording of a talk or training session | As soon as the recording is no longer used for the defined purpose, or the speaker withdraws consent, the recording must be deleted. |
| Meeting summary of a job interview | As soon as the application process is complete and no further claims are to be expected, the summary should be deleted.
|
Retention periods only work if you know where the data ends up. Many companies underestimate exactly this: the provider's cloud, the personal storage of the person who organized the meeting, the CRM, and everyone's inbox when the tool sends out the summary. Deleting in the tool itself is usually not enough.
How does Kertos help with AI meeting notes and recordings?
At Kertos, certified experts take on the role of your external data protection officer. They work out with you which method is enough for which type of meeting, which legal basis applies, and whether the use of AI requires a DPIA.
You record the three-step balancing test in the record of processing activities on the Kertos platform, directly with the processing it covers, together with the purpose and retention periods for transcripts, summaries, and recordings.
With Vendor Management, you keep track of the providers of your meeting tools and review their DPAs, for example whether the use of meeting content for AI training is ruled out. We are happy to show you how this works in practice in a personal demo.
Frequently asked questions
May we transcribe Teams meetings with AI?
Usually, yes. You can generally base AI transcription or AI summaries of Teams, Zoom, or Google Meet meetings on legitimate interest (Art. 6(1)(f) GDPR) if you document the balancing test, inform people in advance, show visibly that transcription is running, and offer a simple opt-out. For an audio or video recording, on the other hand, you need valid consent under data protection law and, in Germany, agreement under criminal law.
Can employees give consent once for all meetings?
For transcription, this is not necessary, because transcription is usually based on legitimate interest. For recordings, blanket consent is usually not "specific" (Art. 4(11) GDPR).
What applies if someone later asks for their contributions to be deleted?
If you base transcription on legitimate interest, the person can object on grounds relating to their particular situation (Art. 21(1) GDPR). After an objection, you must in principle stop processing that person's contributions. The only exceptions are if you can show compelling legitimate grounds for further processing that override the interests, rights, and freedoms of the person objecting, or if you need the transcript to establish, exercise, or defend legal claims. Otherwise, the contributions must be deleted (Art. 17(1)(c) GDPR). With consent, it is simpler: withdrawal removes the legal basis and the contributions must be deleted (Art. 17(1)(b) GDPR).
Is the automatic notice from Teams or Zoom enough?
For transcription based on legitimate interest, the notice meets the requirement for a visible indicator. It does not replace the information required under Art. 13 GDPR, though. Link to your privacy notice in the notice, which most tools allow. For recordings, the banner is not enough.
Do we have to hand over transcripts in response to an access request?
Yes. Transcripts and recordings contain the personal data of the speakers and fall under the right of access, including a copy (Art. 15(1) and (3) GDPR). You must redact or cut out other participants' contributions first (Art. 15(4) GDPR). How to handle the deadlines and scope of a data subject access request under GDPR is covered separately.






