GuidesAI automation

Tutorial: automatic meeting minutes with AI

By the Banzai Labs teamUpdated on 3 min read

TranscriptDecisionsActionsQuestionsAn hour of meeting, one page to keep
The recording becomes a transcript, then minutes in three parts: decisions, actions, questions.

1. The flow

  1. 1RecordingParticipants informed.
  2. 2TranscriptionWith each speaker's name if possible.
  3. 3ExtractionDecisions, actions, questions.
  4. 4DeliveryNotion, Slack or email, with a link to the transcript.

2. The minutes schema

const tool = {
  name: 'meeting_minutes',
  description: 'Factual minutes of the meeting. Nothing that is not said in the transcript.',
  input_schema: {
    type: 'object',
    properties: {
      decisions: { type: 'array', items: { type: 'string' } },
      actions: {
        type: 'array',
        items: {
          type: 'object',
          properties: { who: { type: 'string' }, what: { type: 'string' }, due: { type: ['string', 'null'] } },
          required: ['who', 'what', 'due'],
        },
      },
      open_questions: { type: 'array', items: { type: 'string' } },
    },
    required: ['decisions', 'actions', 'open_questions'],
  },
};

3. The call

const res = await client.messages.create({
  model: process.env.CLAUDE_MODEL!,
  max_tokens: 2000,
  system: 'You write meeting minutes. A due date that was not stated stays null.',
  tools: [tool],
  tool_choice: { type: 'tool', name: 'meeting_minutes' },
  messages: [{ role: 'user', content: transcript }],
});
const minutes = res.content.find((b) => b.type === 'tool_use')?.input;

4. Good practice

Rule Why
A missing due date stays empty The model doesn’t make up dates
Every action has a named owner Otherwise nobody does it
The transcript stays available To check a sentence when in doubt
A set retention period Recordings are personal data

The basics: calling the Claude API from Node.js. The frame: AI, GDPR and the AI Act.

Frequently asked questions

Do participants need to be told?
Yes. Recording a meeting processes personal data: announce it at the start and say what it is for and how long it is kept.
Who does the transcription?
The video-call tool often offers it. Otherwise, a speech recognition model such as open-source Whisper, or a transcription service.
Does an hour-long meeting fit in the context window?
Easily: an hour of speech is roughly ten thousand words, well below the window of today's major models.

Read next

All guides