How to transcribe a sensitive meeting without a bot joining

Not every sensitive meeting needs a transcript. Some conversations — legal advice, disciplinary matters, or candid one-to-ones — may be better served by typed notes you control from the first keystroke. Recording can also affect privilege and expectations of confidentiality in ways that vary by jurisdiction. This page is a practical procedure, not legal advice; if privilege is in play, ask the lawyer involved before hitting record.

Whatever tool you choose, tell the other side before you record and get their agreement. A simple script works: "I take notes with a transcription app so I can focus on the conversation. It records locally on my computer — nothing is uploaded — and I'll delete the transcript after I send you the summary. Is that okay?" Consent rules differ by state and country, so check our guide to recording consent laws before you rely on a verbal nod. Note that desktop recorders give no automatic notice to participants; Fireflies Desktop's consent audio, which announces recording by default, is the exception.

This is the decision that matters most for privacy. There are three workable routes, and the right one depends on how sensitive the meeting is.

For the most sensitive meetings — HR, legal, board, M&A — the on-device route is the strongest choice, and Notta Desktop's Privacy Mode is the most flexible option in it: unlike MacWhisper, it works on both Mac and Windows, and local transcription in Privacy Mode does not consume your cloud transcription minutes. It is still in beta, so test it before relying on it for a critical meeting. If you need cloud AI summaries, Granola is the privacy-conscious cloud option because it deletes audio after real-time transcription.

For the local route, switch Notta Desktop to Privacy Mode (or enable MacWhisper's meeting recording) before the call starts — see our guide to transcribing confidential meetings without uploading for the full offline set-up, including model choice. For the cloud route in Granola, do two things before the meeting: turn off model training on your transcripts, and set transcript auto-deletion to the shortest window that fits your workflow. Doing this in advance means you are not hunting through settings while the meeting is starting.

Run a one-minute test call before the real thing. Confirm that both your microphone and the other side's audio are captured — desktop recorders capture system audio, and a wrong audio input device is the most common reason a transcript comes back half-empty. Check that the transcript appears and, if you are using the local route, that it completes without prompting you to go online.

Record the call and take part normally. If you are screen-sharing during the meeting, do not share the window of the recorder app if that would disclose the transcript or its content to the other side mid-call. With Notta Privacy Mode or MacWhisper, no bot joins the meeting, so participants see nothing unless you tell them — which is exactly why step 2 matters.

Read the transcript before anyone else does. Correct names and technical terms the model got wrong, and delete passages that are more sensitive than they need to be — a summary of the decision is often all that should survive. Note that redaction is manual in Granola, so budget time for this step rather than assuming the tool cleaned it up.

Export the transcript (or your edited summary) to wherever your organisation approves — a case file, a client folder, an HR system. Then delete the transcript and recording from the app itself, and check the local folder where the tool stores recordings so nothing lingers on disk. If you set an auto-deletion window in step 4, verify it actually fired.