Transcribing meetings and conversations with AI: privacy and professional secrecy
Short answer
Transcribing and summarising a recording with AI saves a lot of time, but a recorded conversation often holds the most sensitive data an organisation has. Inform participants in advance, record a legal basis and retention period, and have speech recognition and minutes processed in your own environment. Open-weight speech models make that possible without an external transcription service.
Why conversations in particular are sensitive
In a conversation people say things that would never go on paper: doubts about a case, a patient’s complaints, a client’s financial situation, an assessment of an employee. A recording captures that word for word, including the voices of people who do not always realise they are being recorded.
Cloud transcription and note-taking tools send that recording to their servers, often outside the EU, and keep audio and text under their own terms.
What the rules require
- Inform: participants know in advance that the conversation is recorded and why.
- Legal basis and purpose: recording and transcribing is processing of personal data under the GDPR. Record why it is needed.
- Retention: delete the audio once the minutes are approved, unless there is a reason to keep it.
- Special categories: health, criminal or trade union data require extra safeguards. In healthcare, a recording of a consultation falls under medical secrecy. See AI in healthcare.
- DPIA: when conversations with clients, patients or employees are recorded routinely, a DPIA is often required. See A DPIA for generative AI.
Transcription in your own environment
Speech recognition does not have to happen in the cloud. Open-weight speech models such as Whisper recognise English, Dutch, German and French well and run on the same GPU server as a language model. The process then becomes:
- the recording arrives on your own node;
- the speech model converts audio into text;
- the language model drafts minutes with decisions and action points;
- a participant checks and approves the minutes;
- the audio is deleted after the agreed period.
Transcription with werqly
On a werqly node, speech recognition and minute-taking can run as an automation, connected to where your recordings arrive. Audio and text stay on the node, in werqly AI Control you decide who can create minutes, and you set the log retention period yourself. See What is private AI?
Frequently asked questions
Questions on this topic
May I record a meeting and have AI transcribe it?
Yes, if participants know in advance that the conversation is recorded and why, and you have recorded a legal basis, purpose and retention period. With an external transcription service, a data processing agreement and an assessment of transfers come on top.
Can AI transcribe a consultation or client meeting?
Technically well, but the conversation falls under professional secrecy and often contains special categories of personal data. So process it in your own environment, with consent or a clear legal basis, and have the professional check the minutes.
Does speech recognition work without the cloud?
Yes. Open-weight speech models such as Whisper run on your own GPU server and recognise English, Dutch, German and French, among others. The recording then never leaves your environment.
How long may I keep recordings?
No longer than needed for the purpose. Often that is until the minutes are approved. Record the period in your policy and have deletion happen automatically.
Sources
This article is general information, not legal advice. Rules and guidance change; check the sources listed or ask your lawyer or DPO if in doubt.