AI Meeting Assistants: Transcription, Notes and Action Items

What an AI meeting assistant actually does

An AI meeting record is produced in several steps. First, automatic speech recognition (ASR) turns spoken words into text. Speaker diarization attributes statements to individual people. A language model (LLM) then condenses the transcript into a summary and extracts decisions plus action items with owners and deadlines.

Most tools plug directly into Microsoft Teams, Zoom or Google Meet, or record on-site through a microphone. The benefit is real: less manual note-taking, searchable archives and faster follow-up. The key is to treat the output as a draft, not a finished, legally binding record.

Transcription: accuracy and its limits

Transcription quality is often measured by word error rate. With clear standard German or English, a good microphone and little overlap, modern systems are very strong. Error rates rise noticeably once real meeting conditions appear. Never blindly trust figures, names or quotes taken from a raw transcript.

• Swiss German: dialect is often recognised worse than standard German - many systems only approximate Mundart or translate it inaccurately into High German.
• Jargon, product names and acronyms cause mix-ups unless a custom vocabulary is configured.
• Overlapping voices, poor connections and background noise clearly reduce accuracy.
• Speaker attribution is error-prone when several people share one microphone or sound alike.

From transcript to notes and action items

The summary is the most valuable but also the trickiest step. A language model decides what matters, rephrases decisions and derives tasks. This is where hallucinations appear: the model invents plausible-sounding details, attributes statements to the wrong person or sets a deadline that was never mentioned. Such errors are dangerous because the notes look professional and correct.

So treat the AI notes like an intern's draft: useful but in need of review. Before sending, deliberately check decisions, ownership, numbers and dates against the transcript or your own memory. A short human review step is what turns the output into something trustworthy.

Recording and consent: applying revDSG correctly

In Switzerland, secretly recording a non-public conversation is a criminal matter: the Criminal Code penalises recording other people's private conversations without consent. An AI meeting assistant essentially always records - so you need the informed consent of every participant before recording starts. A clear notice at the very beginning plus a genuine way to object are mandatory, not optional.

• Transparency: under revDSG, people must know that and why their data is processed, and who the provider is.
• Processing on behalf: the tool vendor is usually a processor - a data processing agreement under revDSG (Art. 9) is required.
• Cross-border transfer: US cloud providers require adequate safeguards (e.g. standard contractual clauses or the Swiss-U.S. Data Privacy Framework).
• Sensitive topics: high-risk processing or sensitive personal data may require a data protection impact assessment; the EDOEB is the supervisory authority.

Choosing and rolling out a tool

Before a team adopts a meeting assistant broadly, lightweight governance pays off. Clarify data location and retention, define which meetings are off-limits (e.g. HR talks, legal advice, works-council matters) and decide who may delete recordings. A short internal policy plus a standard consent line at the start of each meeting creates clarity for everyone.

• Data residency: where are audio, transcript and summary stored, and for how long?
• Training use: is content used to train models? Disable it where possible.
• Access rights: who can see notes, and can individuals or meetings be excluded?
• Languages and dialect: test the tool with your real languages, including Swiss German, before committing.