Short answer first: most of them do. The majority of AI note takers join your call as a bot, record the audio (often the video too), and store that recording on a server before they write anything. A smaller group captures audio on your own device. A very small group never touches audio at all. If you handle sensitive client information, the difference between those groups is the whole ballgame.
This guide walks through exactly what each type does with your call, why it matters when you are a client-facing professional in Australia, and how to check what your tool is actually doing rather than what the marketing says. None of this is legal advice, just a clear explanation of how the technology works.
The four levels of what an AI note taker does with your call
"Does it record?" sounds like a yes or no question. It is really a spectrum. There are four meaningfully different things a note-taking tool can do, and each one leaves a different amount of your client's data sitting on a server somewhere.
Level 1: Bot joins, records audio and video, stores the recording
This is the most common setup. A bot appears in your Zoom, Teams or Meet call as a visible participant. It records the audio, usually the video, transcribes it, then writes the summary. The recording is stored, sometimes by default, sometimes until you delete it. Tools that work this way include Otter, Fireflies, tl;dv and Avoma. They are popular for a reason: the transcription is often very accurate, the free tiers are generous, and the integrations into CRMs are broad. The trade-off is that a full recording of your client conversation now lives on the vendor's servers.
Level 2: Device capture, records audio and video, stores the recording
Same recording outcome, friendlier delivery. Instead of a bot joining the meeting, a local app on your machine captures the call. There is no extra participant in the room, which clients tend to find less awkward. Fathom is the well-known example here (bot-free by default, with an optional bot on higher plans). It records audio and video and stores recordings, in Fathom's case on AWS in the US. The data footprint is the same as Level 1, you have just removed the visible bot.
Level 3: Device capture, transcribes, then deletes the audio
A meaningful step down in data retention. The tool captures your device audio, transcribes it in real time, and then deletes the audio, keeping only the text. Granola is the clearest example. No bot joins the call, no audio recording is kept, and only the transcript goes up to the cloud for note generation. This is genuinely better data minimisation than Levels 1 and 2, and Granola is well regarded for it. Two things to keep in mind: the audio still passes through the tool before deletion, and Granola is macOS-only and US-hosted, with no Windows app documented as of mid-2026.
Level 4: Never touches the audio at all
The tool never records, never sends a bot, and never transcribes. It receives a transcript that your phone system or meeting platform already produced, and turns that text into a structured note. This is how CallNote works. Your Dialpad call, your Zoom meeting, your voice memo, your forwarded transcript: the text already exists, and CallNote works from the text. The audio never enters the tool, because there is nothing to record. A couple of other tools sit near this idea too. Jamie says it captures audio locally on your device, and Zocks says it captures data without recording audio, so both are leaning toward less retained data, though they still handle the audio side themselves rather than starting from a transcript that already exists.
Why this matters more for client-facing professionals
If you are a mortgage broker, financial adviser, buyer's agent, lawyer or accountant, your calls are not casual chats. They contain income figures, asset positions, health details, family circumstances, and identity information. That data carries obligations regardless of which note tool you use.
Under the Privacy Act and the Australian Privacy Principles, you are responsible for personal information you collect, and APP 11 requires you to take reasonable steps to protect it. When a tool stores a recording of your client call, that recording is personal information you are responsible for, and it now sits wherever the vendor put it. The question stops being "is the note accurate" and becomes "who else holds a copy of my client's voice, and where."
There is also a plain practical angle. A stored recording is a stored liability. It can be subpoenaed, breached, or requested under an access request. A note that was generated from a transcript and then sealed leaves you with the record you actually need and nothing extra to defend. This is general guidance rather than legal advice, but the principle is straightforward: you cannot lose control of data you never collected in the first place.
The data residency question most people miss
Recording is only half of it. The other half is where the data lives. Most of the popular note takers are hosted offshore. Otter, Fireflies, Fathom and Avoma are US-hosted. tl;dv is EU-hosted by default. Jamie stores data in Europe. None of those are bad companies, and several have strong security postures and explicit no-AI-training guarantees. But for an Australian professional, offshore hosting means your client's data crosses a border and sits under another country's laws.
This is where being built for the Australian market actually changes the answer. CallNote is hosted in AWS Sydney, encrypted with AES-256, and the data is never used to train AI models. For a mortgage broker explaining to a client where their financial details go, "it stays in Australia" is a far easier sentence than "it goes to a server in Oregon." Claras, the established Australian competitor for financial advisers, makes the same Australian-hosting promise, and that shared commitment to keeping data onshore is one of the clearest dividing lines between the local tools and the offshore ones.
| Level | What it does with audio | Examples | Audio stored? |
|---|---|---|---|
| 1 | Bot records audio and video | Otter, Fireflies, tl;dv, Avoma | Yes |
| 2 | Device records audio and video | Fathom | Yes |
| 3 | Device captures, then deletes audio | Granola | No (transcript kept) |
| 4 | Never touches the audio | CallNote | No audio ever |
How to check what your tool is really doing
Marketing pages can be vague on purpose. "Privacy-first" and "secure" do not tell you whether a recording is being kept. Here is how to get a real answer for any tool, including the one you use now.
- Ask the direct question: does it store a recording of the call, or only the transcript? If only a transcript, ask whether the audio is deleted and when.
- Find out where the data is hosted. Look for a named region or country, not just "the cloud". Australian hosting matters if your clients are Australian.
- Check whether your data trains their AI. Many reputable tools now say no in writing. If a tool will not say, treat that as a no until proven otherwise.
- Look at how you get data out and whether you own it. If leaving the tool means losing your notes, that is lock-in.
- For compliance work, ask whether lodged notes are tamper-evident. An append-only audit log and a sealed, timestamped note are what make a file note defensible later.
If you want to go deeper on the bot question specifically, we wrote a longer piece comparing meeting bots versus transcript-only note tools, and a roundup of the AI note takers that don't record.
Where CallNote sits, honestly
CallNote is a Level 4 tool, and it is built narrowly for one job: the compliant Australian client record. It receives the transcript your phone or meeting system already made, generates a clean file note in about two minutes, and lets you review it before you Lodge and Lock it with a SHA-256 seal and an append-only audit log. It ships with per-state recording-consent scripts for all eight states and territories and an NCCP loan-suitability template built on the s130 "not unsuitable" standard.
It is not the right tool for everyone. If you want a full searchable video library of every meeting, a recorder is a better fit, and tools like Otter and Fireflies do that well. If you live in Zoom all day and want the most accurate live transcript, that is their strength, not ours. CallNote is for the broker or adviser who wants a defensible written record after the call and does not want a recording of their client sitting offshore. Different tools for different jobs.
Today CallNote gets your transcript in four live ways: paste it, upload a voice memo (transcribed via Deepgram or Whisper for phone and in-person calls), forward it to a unique email address, or connect Dialpad with your API key so every call transcript becomes a note. More phone and CRM connectors are rolling out.
Common questions
Do AI note takers record your calls?
Most do. The majority join the call as a bot, record the audio and often the video, and store that recording. A smaller group captures audio on your device, and a few transcript-only tools like CallNote never touch the audio at all because they work from a transcript your system already made.
Are AI meeting notes private?
It depends entirely on the tool. Privacy comes down to three things: whether a recording is kept, where the data is hosted, and whether your data trains the vendor's AI. Many tools store offshore recordings by default. Always ask all three questions before you trust a tool with client calls.
Which AI note takers don't record audio?
Granola captures device audio then deletes it, keeping only the transcript. CallNote goes a step further and never touches audio at all, receiving a transcript your phone or meeting system already produced. Both keep far less data than bot-based recorders.
Where is my client data stored with these tools?
Mostly offshore. Otter, Fireflies and Fathom are US-hosted, tl;dv and Jamie are EU-hosted. CallNote and Claras are the Australian-hosted options, which keeps client data onshore under Australian law.
Is it safe to use an AI note taker for financial client calls?
It can be, if you choose carefully. Prefer a tool that keeps no recording, hosts your data in Australia, does not train AI on your content, and produces a tamper-evident note. For NCCP file notes, a sealed audit trail matters. This is general guidance, not legal advice.