The short version. Otter, Fathom and Fireflies all record and store your meeting by default, all store it in the United States, and all hold a SOC 2 Type II report. They differ most on training. Fireflies says it does not train on your content. Fathom trains its own models on de-identified data unless you opt out. Otter trains its own models on de-identified audio and on transcripts. All three put the legal job of getting consent on you.
None of that makes a tool unsafe. It tells you what you are agreeing to. This is general information as at October 2026, not legal advice, and vendors change their terms, so read the linked pages yourself before you decide.
AI note taker privacy concerns: the five questions that matter
Most worry about AI note takers comes down to the same five questions. Each one has a plain answer in the vendor's documents if you know where to look.
- Is the audio recorded and kept? A stored recording is one more copy of the client's voice and details that you have to secure, find and eventually delete.
- Where is it stored? If you are in Australia and the data is held overseas, Australian Privacy Principle 8 in the Privacy Act 1988 (Cth) generally expects you to take reasonable steps to make sure the overseas recipient handles it properly, and you can remain accountable for what they do with it.
- Is it used to train AI models? Look for three separate answers: the vendor's own models, the outside AI companies it sends text to, and whether there is a setting to switch training off.
- How long is it kept, and can you delete it? Australian Privacy Principle 11 expects you to protect personal information and to destroy or de-identify it once you no longer need it. You need to know the tool can do that.
- Who tells the other people on the call? Every vendor we read makes the user responsible for consent. Some send a notice for you. That notice may or may not meet the law where you are.
Otter, Fathom and Fireflies compared
Everything in this table comes from the vendor's own published pages, read in October 2026. Where a vendor does not say something, we say so instead of guessing.
| Otter | Fathom | Fireflies | |
|---|---|---|---|
| Audio recorded and kept | Yes. Audio recordings and transcripts are stored. Its notetaker may also take screenshots in virtual meetings. | Yes with its bot, which records audio and video. Its help centre says recordings are kept indefinitely, or until you delete them. Its newer bot-free version, on Mac only as at October 2026, adds a Transcript Only mode that saves no recording and an Audio Only mode with no video. | Yes when the notetaker bot joins. Its desktop no-bot mode keeps a transcript and notes with no audio or video file. |
| Where data is stored | Amazon Web Services, based in the United States | United States ("All Fathom data is stored in the United States") | United States by default (AWS and Google Cloud). Private storage in a location you choose is an Enterprise option, with processing still in the US. |
| Trains its own models on your data | Yes. On de-identified audio recordings, and on transcripts that its policy says may contain personal information. | Yes by default, on de-identified data. You can opt out. | No. Its policy says it does not use personal information for AI model training. |
| How to turn training off | No general opt-out is described on its privacy policy or security page. Ask Otter in writing. | In My Settings. A Team Edition admin can opt out every user in Organization Settings. | Nothing to switch off, per its policy. |
| Outside AI providers | Says its AI service providers do not train on or store customer data | Says it does not authorise third parties to train on your content | Says vendors are contractually barred from training and hold content for zero days |
| Retention and deletion | Kept as long as needed for the policy's purposes. Deleted conversations sit in Trash for 30 days. Custom retention is an Enterprise workspace setting. | Kept until you delete it. The account stays until you delete it yourself. Its policy aims to delete within 30 days of a request. | Kept while the account is active, deleted within 30 days of closing it. An auto-delete setting, on its Enterprise plan, runs from one week to one year. |
| Telling other participants | The user is responsible. The notetaker joins the meeting as a visible participant. It can post chat messages with a link to the live transcript, and those can be switched off. | The user is responsible. Fathom says there is no way to record silently: a visible bot and notice, or a chat message or consent prompt in bot-free mode. | The user is responsible. Optional email an hour before with a link to decline, plus an in-meeting chat message. |
| Independent audit | SOC 2 Type 2 | SOC 2 Type II | SOC 2 Type II |
See the kind of file note a transcript-only tool produces.
Pick your line of work and we'll send a sample note written from a mocked call. No account, no card, and nothing from your own client files.
One sample note, plus the occasional CallNote update. Unsubscribe any time by replying. We never ask for your call recordings or client files.
Is Otter AI safe?
Otter is a mature product with real security controls. Its security page describes AES-256 encryption at rest on Amazon Web Services and a SOC 2 Type 2 report. Its privacy policy (effective 16 June 2026) names Amazon Web Services, based in the United States, as its storage provider.
The point to weigh is training. Otter's policy says it uses your information for "training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)". Its security page adds that the process is automatic, that no person reviews your audio without your explicit consent, and that the outside AI providers it uses do not train on or store customer data. We could not find a general way to opt out of that training on either page, so a firm that needs one should ask Otter for it in writing.
On consent, Otter's terms say you are "solely responsible for providing any notices to, and obtaining consent from, individuals in connection with any recordings". Its help centre says the notetaker joins your Zoom, Google Meet or Microsoft Teams meeting to record and transcribe it, that it can send chat messages with a link to the live transcript and summary, and that you can turn those messages off.
Our side-by-side is on the CallNote vs Otter page, and there is a wider list in Otter alternatives in Australia.
Is Fathom AI safe?
Fathom is clear about what it does. Its privacy policy (last updated 16 August 2026) says it receives the meeting audio, video and speaker identification, and its help centre says all Fathom data is stored in the United States. It lists SOC 2 Type II and says it will sign a data processing agreement with customers in the EU and UK.
On training, Fathom's terms say it may use de-identified data from your content to train or improve its in-house models, and that you can opt out in your account settings. Its help centre says an organisation on Team Edition can opt out all of its users at once. It also says the AI companies it uses are not permitted to train on your data. If you choose Fathom for client work, switching that setting off on day one is a two-minute job.
Fathom now offers capture with a bot or without one. Its bot records full audio and video. Its newer bot-free version, which its help centre says is on Mac only as at October 2026, has a Transcript Only mode where no recording is saved and an Audio Only mode with no video. Either way, its help centre says there is no way to record silently: the older version shows a visible bot and a recording notice, and the bot-free version posts a notice in the meeting chat or shows consent prompts. Its terms still require you to have "obtained the required consent of every meeting participant to record each call before the call begins". When you delete your account, Fathom says recording data is removed, and backups follow after a further seven days. See CallNote vs Fathom for the full comparison.
Is Fireflies AI safe?
Fireflies has the firmest written position on training of the three. Its privacy policy (last updated 28 September 2026) says: "We do not use personal information for AI model training and we contractually prohibit our vendors from using this information for their own model training." Its terms repeat that it will not use your content to train generative AI models.
By default your data is stored and processed in the United States on AWS and Google Cloud. Enterprise customers can use private storage in a location they choose, although Fireflies says the data is still processed on its servers in the US. It lists SOC 2 Type II, and on its Enterprise plan it has an auto-delete setting that removes meetings after a period you pick, from one week to one year. An admin can set that for the whole workspace.
Fireflies gives you more consent tools than most. You can have it email participants an hour before the meeting with a link to decline, and post a message in the meeting chat when it joins. Its help centre also describes transcript-only and summary-only modes, on the Enterprise plan, for organisations that cannot store audio. Its terms still make you responsible for making sure each participant consents.
The full comparison is on the CallNote vs Fireflies page.
Are the Microsoft Teams and Zoom AI note takers secure?
The built-in tools are a different case, because the data stays with a platform your firm already uses.
- Microsoft Teams and Copilot. Microsoft's documentation says prompts, responses and the data Copilot reads through Microsoft Graph "aren't used to train foundation LLMs". Recordings are saved to the organiser's OneDrive or SharePoint, and transcripts are stored with them, so they follow your own Microsoft 365 retention rules. Everyone is notified when recording or transcription starts, and an admin can require explicit consent. Copilot can also be set to run only during the meeting, using speech-to-text data that Microsoft says is not saved afterwards. We cover that in Teams transcription without recording.
- Zoom AI Companion. Zoom's privacy statement and terms say it does not use your audio, video, chat, screen sharing, attachments or similar customer content to train its own or third-party AI models. Its terms make the host responsible for recording laws, and say participants are notified when recording is switched on. We did not verify where AI Companion summaries are stored for Australian accounts, so ask your Zoom admin.
For both, safe depends mostly on how your own tenant is set up: who can start a transcript, where files land, and how long they are kept.
A checklist to assess any AI note taker
Use this on any tool, including ours. Each answer should point to a page or a clause, and a vendor that cannot point to one has given you your answer.
- Audio. Does it record? Is the audio kept, deleted after transcription, or never received? Is video or a screenshot captured too?
- Location. Which country and which cloud provider hold the data? Is processing done somewhere different from storage?
- Training. Does the vendor train its own models on your content? Is it de-identified first? Is there an opt-out, and is it per user or for the whole firm?
- Outside AI providers. Which companies receive the text, and do their contracts stop them keeping it or training on it?
- Retention. What is the default period? Can an admin set a firm-wide rule? What happens to backups?
- Deletion. Can you delete one meeting, one client and the whole account? How long does it take?
- Consent and notice. What does the tool tell other participants, when, and can a staff member switch it off? Does that meet the recording law in your state? See is it legal to record phone calls in Australia.
- Voice data. Does it build voiceprints to recognise speakers? Is that opt-in?
- Access. Who inside your firm can see a colleague's meetings by default? Can vendor staff open your content without your permission?
- Evidence. Is there a current audit report such as SOC 2 Type II, and will the vendor sign a data processing agreement?
- Your own record. Write down what you checked and what you decided. Our AI note taking policy template and law firm AI policy template give you a starting point.
- What we were deciding
Whether advisers can use a meeting note taker on client review meetings held by video.
- What the vendor's documents say
- Audio and video recordings are stored until we delete them.
- Data is stored in the United States.
- Our content is used to train the vendor's own models unless we opt out. An admin can opt out the whole firm.
- The vendor's terms make us responsible for getting each client's consent.
- Decision
Approved for internal meetings only. Not approved for client meetings until the firm-wide training opt-out is switched on, our privacy policy names the overseas storage, and advisers use the agreed consent wording at the start of each meeting.
- Next steps
- Practice manager to switch on the opt-out and confirm in writing.
- Review again in six months, or sooner if the vendor changes its terms.
Where a transcript-only tool differs, and where it does not
CallNote works differently from the tools above. It does not record calls, does not join meetings and does not store call audio. It takes a transcript that already exists, from Aircall, Dialpad or Microsoft Teams, or one you paste, upload or forward, and writes a file note in your firm's format. That removes two of the five questions for the note tool itself: there is no recording for it to keep, and no bot for it to announce.
It does not remove the questions altogether, and we would rather say so. The transcript has to come from somewhere. If Aircall, Dialpad or Teams made it, then that system captured the call, and its recording settings, storage location and consent notice are still yours to check. A transcript is also personal information in its own right. It holds the same income figures and health details the audio did, in text.
Here are our own answers to the checklist. Notes and transcripts are stored in Sydney on Amazon Web Services. The note is written by Anthropic's Claude models through its commercial API, which means the transcript text is processed outside Australia. We do not train AI models on your data. CallNote does not hold a security certification such as SOC 2 or ISO 27001. Otter, Fathom and Fireflies each list a SOC 2 Type II report, so on that point they are ahead of us. The full list of providers and what each one receives is on our security page.
If a searchable library of recordings is what you want, a recorder is the better fit. If you want a written record of a client call and no new recording, read AI note takers that don't record and do AI note takers record your calls.
Common questions
Are AI note takers safe to use for client calls?
They can be, if you check five things first: whether audio is kept, where data is stored, whether your calls train the vendor's models, how deletion works, and who tells the other participants. Otter, Fathom and Fireflies all store recordings in the United States and all make the user responsible for consent. This is general information, not legal advice.
Is Otter AI safe and secure?
Otter lists SOC 2 Type 2 and AES-256 encryption, and stores data on Amazon Web Services in the United States. Its privacy policy says it trains its own AI on de-identified audio recordings and on transcripts that may contain personal information. A proposed class action about recording without consent is before a US court. Those claims are allegations and have not been decided.
Is Fathom AI notetaker safe?
Fathom lists SOC 2 Type II and stores all data in the United States. Its bot records audio and video, and its bot-free version on Mac has transcript-only and audio-only modes. By default it uses de-identified data to improve its own models, and you can opt out in My Settings, or for a whole organisation on Team Edition. Fathom says it cannot record silently, and its terms require you to get every participant's consent.
Is Fireflies AI safe to use?
Fireflies lists SOC 2 Type II and stores data in the United States by default, with private storage on Enterprise. Its privacy policy says it does not use personal information for AI model training and bars its vendors from doing so. It offers an email notice before meetings, a chat notice and, on its Enterprise plan, an auto-delete setting. You remain responsible for consent.
Do AI note takers use my meetings to train AI?
It depends on the vendor. As at October 2026, Otter's policy says it trains its own models on de-identified audio and on transcripts. Fathom trains on de-identified data unless you opt out. Fireflies says it does not train on your content. Microsoft and Zoom say customer content is not used to train their foundation or AI models. Read the current policy before relying on any of this.
What is the main security risk with an AI note taker?
The risk a firm can most easily control is an unapproved tool: a staff member adds a personal note taker to a client call and nobody has checked where the recording goes, whether it trains a model, or whether the client was told. A short written policy and an approved tool list deal with that.