AI meeting notes without recording in 2026
By Kyle Nelson, Founder, Fazit
AI meeting notes without recording are possible if the app transcribes live audio in memory, writes the note, and never creates an audio file. For confidential 1:1 client calls, the clean version is local capture, on-device transcription, local note generation, and a plain Markdown note in your own vault.
By Fazit's founder. I'm not a lawyer, and this is not legal advice. Consent and professional-conduct duties still apply to the person running the call.
Can AI meeting notes work without recording?
Yes. The recording isn't the note. A recording is a durable audio artifact of the call. A note is a written output created from the conversation.
Most people search for "ai meeting notes without recording" because they don't want a bot in the room, they don't want a replayable audio file, and they don't want a vendor storing client conversations. Those are three different questions. A product can remove the bot and still stream audio to a server. A product can delete audio later and still have recorded it first. A product can avoid video recording and still store a transcript.
Fazit's position is narrower and easier to audit: a native macOS menubar app captures a 1:1 call locally, transcribes with Parakeet via CoreML, generates the note using a local model on localhost, and writes one Markdown file into an Obsidian vault or Apple Notes. The call audio is never written to disk.
What does "without recording" mean?
"Without recording" should mean no audio file exists, not "we deleted the recording after processing." Those are different risk profiles for consultants, coaches, lawyers, financial advisers, and therapists.
In Fazit, audio exists in a fixed-size RAM ring buffer. The buffer has no write API. It is destroyed on every exit path, including errors. Every note records audio_retained: false in its frontmatter. The durable artifact is the note you chose to keep, not a hidden recording behind it.
That distinction matters because deleting later still creates a period where the audio exists outside your control. A cloud notetaker may need the audio on someone else's infrastructure before it can transcribe. A local notetaker can treat audio as a temporary signal and discard it as soon as text exists.
For a deeper engineering view, see why the no-recording claim is an invariant. The short version is the product promise maps to one implementation constraint: Fazit has no code path that writes call audio to disk.
How do AI notes without a bot differ from AI notes without recording?
Bot-free and no-recording are separate claims. Bot-free describes what other participants see in Zoom, Teams, Meet, WhatsApp Desktop, or Slack. No-recording describes what happens to the audio.
A bot-free cloud tool may still send audio to a vendor server for transcription. A bot-based tool joins as a participant and ships meeting audio to the vendor cloud. A local no-recording tool avoids both: nothing joins the meeting, and the call content doesn't leave the Mac for note generation.
| Approach | What participants see | Where call content goes | What survives after the call |
|---|---|---|---|
| Meeting bot with cloud processing | A bot participant joins the call | Audio goes to a vendor cloud | A recording, transcript, note, or account record can remain depending on the tool |
| Bot-free cloud notetaker | No bot appears in the meeting | Audio still streams to vendor clouds or third-party speech services | Transcript or notes can remain in the vendor system |
| Built-in meeting summary | The meeting app may show its own notice or banner | The meeting platform processes and stores the transcript or summary according to its settings | The transcript or summary is stored on the host's account in the platform's system |
| Fazit | No bot joins the call | Call content stays on the Mac for transcription and note generation | One Markdown note in your Obsidian vault or Apple Notes, with audio_retained: false |
The bot question is worth caring about. Fazit's own research notes a 6-week test with 47 meetings and 22 client calls where 7 of 22 clients, 32%, questioned the visible recording bot when Fireflies or Otter joined, while zero questioned botless system-audio capture. The audio question is the one that decides who holds a copy of the conversation after the call ends.
What should consultants, coaches, lawyers, advisers, and therapists ask?
Ask where the audio goes before you ask how pretty the note is. Summaries have converged. The privacy and confidentiality question is architectural.
There are five plain checks.
First, does anything join the meeting? If yes, you're adding a third-party participant to a confidential client call. Fazit doesn't join the meeting because capture happens at the audio layer on the Mac. The app reads the microphone as "You" and the call app output as "Them."
Second, does the app need your calendar? Meeting bots often need calendar access so they know which calls to join. Fazit doesn't need calendar access for capture because you pick the call app in the menubar. More detail is in the calendar-access piece.
Third, does the audio leave the device? In Fazit, transcription runs on-device and note generation runs on localhost. Account, licensing, payment, updates, and optional product analytics are a separate lane from call content.
Fourth, is audio written to disk? In Fazit, no. Audio lives in RAM only and is destroyed when capture ends or fails.
Fifth, where is the note stored? Fazit writes a plain Markdown file into a folder of your Obsidian vault, or to Apple Notes. There is no Fazit database holding a copy.
Does local transcription mean nothing leaves the Mac?
Be precise: in Fazit, "nothing leaves the Mac" is a claim about call content. It doesn't mean the app never contacts the internet for any reason.
Models are downloaded once from public CDNs on first run. Account, licensing, payment, updates, and optional product analytics may use Fazit's servers. Those are disclosed separately from the content path.
The content path is the part professionals should care about for client confidentiality. The microphone stream and the call app stream are captured locally. Parakeet via CoreML transcribes locally. The note-writing model runs on localhost. The audio, transcript, and note content don't leave the device for note generation.
This is why hardware matters. Cloud transcription became common when laptops couldn't run speech models well enough. Apple Silicon changed the local path. Fazit's related post on on-device transcription explains why a server-class speech model can now run on a MacBook fast enough for client calls.
How do Obsidian notes fit into a no-recording workflow?
The output should be a file you control. A note in a vendor database is still a vendor-held copy of the call's substance.
Fazit writes one Markdown file into your Obsidian vault. The note can include a summary, notable moments, action items with owners, a paste-ready follow-up email, and a You/Them transcript for 1:1 calls. Because your microphone and the call app audio are separate streams, Fazit doesn't need diarization guesswork to label a two-person call.
That 1:1 boundary is deliberate. In calls with 3 or more people, remote voices arrive mixed into one system-audio stream, and separating them would require diarization. Fazit's own research found open local diarization error rates of 13.3% to 17.7% in the cited benchmarks, so the product doesn't promise multi-party speaker labels in the MVP.
If you're building a long-term knowledge base, the storage format matters as much as the note text. Fazit's position is covered in the Obsidian files article: a meeting note is yours when it's a file you can grep in your vault.
What legal consent duties remain if no audio file is kept?
No-recording architecture reduces vendor exposure and audio-file exposure. It doesn't remove your consent duties, client notice duties, or professional-conduct duties.
That sentence matters for every confidential profession. A lawyer has privilege duties. A therapist or coach may have client trust and privacy duties. A financial adviser has recordkeeping policies and vendor-risk questions. A consultant may have an NDA that limits disclosure to third parties. The tool can avoid sending call content to a vendor, but the person running the call still controls whether and how participants are told.
Fazit's own documentation tells users to give a plain notice before a call. The honest version is simple: you use a local tool to create notes, the audio isn't kept, and the other person can say no.
For legal context, read primary law and your own counsel. Fazit's related posts, including recording client calls and AI notetakers for lawyers, describe the architecture and link the legal questions to the recording artifact, but they don't replace advice from a lawyer.
The safest product claim is the one the code can keep: Fazit doesn't record the call audio, doesn't send call content to a vendor for note generation, and writes one user-owned note.
Sources
- Fazit site statement of facts and accuracy notes: https://getfazit.com/llms.txt
- Fazit security and architecture page: https://getfazit.com/security/
- Apple Developer documentation for Core Audio taps: https://developer.apple.com
- insidegui AudioCap Core Audio process-tap sample: https://github.com/insidegui/AudioCap
- FluidAudio Swift SDK and Parakeet CoreML information: https://cocoapods.org/pods/FluidAudio
- Obsidian: https://obsidian.md