How to take AI meeting notes without the cloud

Most AI meeting note tools upload your audio — straight to a server, in the background, before you've finished the call. If you've ever used Otter.ai, Fireflies, or Zoom's AI Companion, that's the deal you implicitly agreed to. Here's how to get the same output without it.

Why cloud meeting AI is a privacy trade-off

Tools like Otter.ai, Fireflies, Zoom AI Companion, and Microsoft Copilot for Teams all have the same architecture: your audio (or the video stream) goes to their servers, where it's transcribed and processed. Enterprise plans come with data processing agreements, but the audio still travels. Some services store transcripts indefinitely by default; the free tier of Otter.ai includes 300 minutes of storage per month at no charge but your recordings live on their infrastructure.

That's a reasonable trade for general meetings. It's a less comfortable one when the call involves client strategy, pricing, HR conversations, legal discussions, or anything you'd hesitate to cc a stranger on. Meeting recordings are dense with the kind of detail that's hard to unring once it's left the building.

The alternative isn't to go back to handwritten notes. It's to run the transcription locally, then use a local AI to do the analysis. The result is identical output with zero upload.

The two-step local workflow: Whisper + Outlier

The workflow has two distinct phases. Whisper handles speech-to-text on your Mac — it's an open-source model from OpenAI that runs entirely locally, with no API call. Outlier handles everything after: summarization, action items, follow-up drafts. Neither step sends anything anywhere.

The whole pipeline runs on your chip. Your audio file sits in a local folder. The transcript never leaves your machine. This is what 100% offline looks like in practice.

How to transcribe with Whisper on Mac

Whisper is available as a command-line tool and as several GUI wrappers. The easiest path on macOS uses Homebrew to install whisper-cpp, a C++ port that runs efficiently on Apple Silicon.

1. Install via Homebrew: brew install whisper-cpp
2. Download a model. Whisper Large v3 gives near-human accuracy on English and fits in 4–6 GB of RAM. For shorter, simpler calls, the medium model is faster and lighter: whisper-download-ggml-model large-v3
3. Point it at your audio file: whisper-cli -m /path/to/model.bin -f meeting.mp3 -otxt
4. Wait. On an M-series chip, a 60-minute meeting typically transcribes in a few minutes. The output is a .txt file alongside your audio.

If you prefer a GUI, apps like Whisper Transcription (available on the Mac App Store) wrap the same model in a drag-and-drop interface. Either way, no audio leaves your machine.

Whisper accuracy on standard English is strong. It handles multiple speakers adequately, though it doesn't always label who said what — you may want to make a quick note of attendees before you run the transcript through Outlier.

Prompts that work for meeting notes in Outlier

Once you have the transcript, paste it into Outlier and ask directly. These prompts work well for typical meeting output:

Summary + action items (the workhorse)

Summarize this meeting transcript in bullet points, then list all action items with owners and deadlines.

Outlier Core 27B handles long context well — you can paste 10,000+ word transcripts and it stays coherent. If you're on Nano or Lite, very long transcripts may get cut off; break them into sections if needed.

Decisions only

From this transcript, list only the decisions that were made. Don't include discussion or action items — just decisions and who made them.

Follow-up email draft

Write a follow-up email based on this meeting transcript. Include: what was decided, action items with owners, and next meeting date if mentioned. Keep it under 200 words, professional tone.

Specific attendee's commitments

From this transcript, extract every commitment or action item attributed to [Name]. List them with any deadlines mentioned.

Outlier can also reformat notes into a specific structure if your team uses a template — just describe the format in the prompt.

What to do with the notes

Outlier's output is plain text. Copy it and drop it wherever your notes live. Common destinations:

• Notion or Obsidian: Paste directly; both handle markdown formatting Outlier produces naturally.
• Email: For follow-up drafts, copy the output into your mail client and send. No export step needed.
• Plain text file: For archiving, save the transcript and the Outlier summary as .txt or .md in a local folder alongside the audio.
• Shared doc: If the meeting has multiple attendees who need the notes, paste into a Google Doc or Notion page and share — just the text, not the audio.

There's no export button to click because there's nothing to export from — the content is already in your clipboard, already on your machine, already yours.

Limitations worth knowing

Whisper and crosstalk. When two people speak at the same time, accuracy drops. On calls with clear turn-taking (video calls with mute discipline), transcripts are clean. On chaotic in-room meetings with multiple voices overlapping, expect some garbled segments. A quick skim before pasting into Outlier catches the worst of it.

Heavy accents and acronyms. Whisper Large v3 handles most accents on English well, but domain-specific jargon and acronyms — especially technical abbreviations — sometimes get phonetically mangled. A post-edit pass for company-specific terminology (product names, team names, abbreviations) takes a minute and saves confusion later.

Speaker labeling. Whisper doesn't reliably label who said what. The transcript is sequential text, not attributed dialogue. If you need attribution, add a short note before pasting into Outlier: "The speakers are Alice (PM), Bob (Engineering lead), and Carol (Client)." Outlier can use that context when extracting commitments.

Long transcripts and model tier. Outlier Core 27B handles multi-thousand-word transcripts in one pass. Nano and Lite have shorter context windows and will struggle with a 90-minute meeting transcript as a single paste. If you're on a smaller model tier, chunk the transcript into sections (first 30 minutes, second 30 minutes) and run summaries separately, then ask Outlier to combine them.