TL;DR: An AI notetaker takes more than a summary with it: it takes the audio, the word-for-word transcript, often a voiceprint, and everything that was said about people who weren't there. In Québec, telling people about the collection is an obligation (Law 25), and handing the recording to a service outside Québec first requires a privacy impact assessment and a written agreement. A vendor's terms can change, and a lawsuit against Otter.ai is moving forward in California. A pipeline that never leaves your premises does exist (Whisper, MIT licence, plus a local language model), but it makes mistakes too: a review by someone who attended stays the last step.
Disclaimer: this text explains a legal framework in plain language, without being legal advice. The sections cited come from Québec's Act respecting the protection of personal information in the private sector, the Civil Code of Québec and the Criminal Code. The exact scope of an obligation in your organization is something to confirm with legal counsel.
In this article
- What the bot takes with it when it leaves the room: the audio, transcript and voiceprint leave with the tool, and the absent people you discuss agreed to nothing.
- The host's yes doesn't cover the whole table: a lawsuit against Otter.ai is moving forward, and in Québec each person must be told the purposes and means of collection.
- When the recording leaves Québec: a service outside Québec requires a privacy impact assessment and a written agreement.
- Terms of service move, your contract should stay put: Zoom rewrote its terms in 2023 after a controversy, hence six questions whose answers belong in the contract.
- The assistant bundled with your suite is not a way out: Copilot's no-transcript mode exists, but prompts may still be kept with an American vendor.
- The pipeline that never leaves your premises: Whisper and pyannote.audio under the MIT licence, plus a local language model, on hardware you control.
- AI makes things up too, even when you host it: about 1% of Whisper transcriptions contained made-up sentences, so someone who attended reviews before sending.
- A six-line meeting rule: announce, turn off auto-join, exclude some meetings, destroy the audio, review, and refuse outside assistants.
- At Blue Fox: the rule and the assessment first, then a pipeline hosted in Québec that puts the minutes into Odoo.
Monday, 9:02 a.m. In the participant list of your leadership meeting, between the CFO and the outside advisor, there's a line nobody invited: "Notetaker." The advisor connected a note-taking assistant to their calendar three months ago, and the assistant follows the calendar. It listens to the discussion about the restructuring, the client who pays late, the employee on sick leave. At 10 a.m., it sends a flawless summary to everyone. The summary is the visible part.
This kind of scene is not hypothetical. In 2024, an engineer told the Washington Post that after a Zoom call with investors, Otter had sent him the transcript of the meeting, but also the conversation the investors kept having after he left, with confidential details about a business. The deal didn't go through.
We've already compared cloud note-taking tools in our overview of meeting assistants. This post asks the questions that come before choosing a tool: what leaves, who agreed to it, what Québec law expects of you, and what the version that stays in-house looks like.
What the bot takes with it when it leaves the room
Minutes fit on one page. What comes before them weighs a lot more: the full audio file, the word-for-word transcript, the participant list, and often a voiceprint that lets the tool recognize each person from one meeting to the next. The lawsuit against Otter.ai alleges precisely that the service keeps conversation content, voiceprints and other personal information, including to train its speech recognition models. These are allegations, not facts established by a court. They still describe well what such a tool technically has in its hands.
A meeting also contains people who aren't in it. You talk about the client disputing an invoice, the candidate you picked, the colleague on sick leave. None of them agreed to anything: not the transcript, not the service, not its terms. For an SMB, it's often the most sensitive content of the week, more than email, precisely because people speak freely.
The host's yes doesn't cover the whole table
The lawsuit filed against Otter.ai on August 15, 2025 in California starts there. According to the complaint, as reported by NPR, the tool asks the meeting host for permission to record by default, but not the other participants. And when the host has linked their Google Meet, Zoom or Teams calendar to the service, the assistant can join the meeting without the affirmative consent of anyone, host included. The cases were consolidated, and on August 13, 2026, Judge Eumi K. Lee refused to dismiss the core of the complaint: the claims under the U.S. federal wiretap law, California's anti-interception law, and Illinois's biometric privacy law, which covers voiceprints, are moving forward. Other claims were dismissed. Nothing has been decided on the merits: the case has simply cleared the stage where a judge could have stopped it outright.
In Canada, the Criminal Code doesn't make a recording an offence when one of the people taking part in the conversation consents to it (s. 184(2)(a)). Many people stop there, and that's a mistake. Québec's Civil Code lists intentionally intercepting or using a private communication among invasions of privacy (art. 36). And Québec's private-sector privacy act, modernized by Law 25, requires anyone collecting personal information from a person to inform them, at the time of collection, of the purposes and the means of that collection (s. 8). A voice, statements tied to a name, an opinion about an employee: that's personal information.
In practice, the defensible minimum fits in one sentence said at the start of the meeting: this meeting is transcribed by this tool, to produce the minutes, which will go to these people. With a real chance to say no. This post isn't legal advice: for a borderline case (health, human resources, litigation), talk to your person in charge of the protection of personal information or to your lawyer.
When the recording leaves Québec
Most cloud notetakers process audio in the United States or in Europe. Québec law doesn't forbid that transfer, it sets conditions on it. Before communicating personal information outside Québec, a business must conduct a privacy impact assessment (PIA) that takes into account the sensitivity of the information, the purposes for which it will be used, the protection measures, including contractual ones, and the legal framework of the destination State (s. 17). The communication must be covered by a written agreement. And the rule also applies when the business entrusts someone outside Québec with collecting, using or keeping that information on its behalf: that is exactly what a transcription service does.
Subscribing to such a tool is also, in the law's terms, a project to acquire an information system involving personal information. That triggers a PIA from the outset, with the person in charge of the protection of personal information consulted (s. 3.3). The assessment must be proportionate: a few team meetings with nothing sensitive don't call for the same exercise as a human resources committee. But it has to exist, and a free trial switched on by an employee on a Tuesday afternoon never had one. We've covered what Law 25 really changed for SMBs, and why a vendor's jurisdiction matters in our article on the Cloud Act.
Canada's privacy authorities, including Québec's Commission d'accès à l'information, signed principles for generative AI in December 2023 that point the same way: valid consent when consent is the legal basis, appropriate purposes, assessments such as PIAs, and outputs as accurate as their use requires. That last point comes back further down.
Terms of service move, your contract should stay put
In the summer of 2023, Zoom found itself in the middle of a controversy: an update to its terms of service made earlier that year was read as permission to train its AI models on customer content. On August 7, the company published a clarification, then changed its terms: it does not use customer audio, video, chat, screen sharing or attachments to train its own models or third-party models. The lesson isn't about Zoom, which corrected course. It's about the mechanism: what protects you this morning is a text the vendor can rewrite.
Otter, for its part, doesn't hide the training. Its policy says it gets explicit permission, through a box the user checks, to train its systems on meeting transcripts after "de-identifying" them. The complaint specifically disputes that this de-identification removes confidential information. Above all, notice who checks the box: the account holder, not the people they're talking to.
The useful habit is to ask every vendor the same questions and to require written answers, in the contract or the data processing agreement, not in a help page that can change tomorrow.
| The question | The reassuring answer | The answer that should stop you |
|---|---|---|
| Where is the audio processed and stored? | In Canada, ideally in Québec, named in the contract | "In the cloud," with no country given |
| Are recordings used to train models? | No, by default and by contract | Yes, once "de-identified," or through a pre-checked option |
| How long are they kept? | A period you set, with verifiable deletion | As long as the account exists |
| Does the tool join calendar meetings on its own? | Never without an explicit action by the organizer | Yes, as soon as the calendar is connected |
| Who gets the summary automatically? | The organizer, who reviews it before sharing | Every participant, external ones included |
| Which subprocessors touch the data? | A published list, with the countries they operate in | No list, or "our trusted partners" |
The assistant bundled with your suite is not a way out
If your organization lives in Microsoft 365, Copilot in Teams seems to settle the question: no third-party tool, and the data stays in your tenant. Microsoft even offers a setting, "Only during the meeting," where Copilot produces notes and tasks without transcription or recording. It's a sensible choice for a delicate meeting, with an accepted downside: after the meeting, Copilot doesn't show up in the recap, and anything worth keeping has to be copied before the end.
The same help page, however, contains a warning worth reading carefully: depending on your organization's retention policies, Copilot prompts and responses may be kept for compliance purposes, even without transcription, and found through an eDiscovery search. "No transcript" therefore doesn't mean "nothing kept." And the tenant, however well protected, sits with an American vendor: the jurisdiction question doesn't go away because the tool is included in the licence.
The pipeline that never leaves your premises
Automatic minutes are made in three stages, and each one can run on hardware you control, on your premises or with a hosting provider in Québec.
Transcription comes first. Whisper, the speech recognition engine published by OpenAI, is under the MIT licence, code and model weights included: you download it and run it with no account and no subscription. It comes in six sizes, from "tiny" to "turbo," which need roughly 1 to 10 GB of video memory. For a manager, it boils down to this: a well-equipped workstation is enough for the small models, the best ones call for a dedicated graphics card.
Next, speaker separation, called diarization, attributes each passage to a voice. The reference library, pyannote.audio, is also under the MIT licence. Its community model downloads after you accept terms of use, and the publisher also offers a paid service that runs on its own servers. The choice deserves to be made knowingly: the local model is the one that keeps the audio in-house.
Finally, a language model turns the transcript into structured minutes: topics discussed, decisions, follow-ups. That model can run locally too, as we showed with Ollama for local, private AI and with an assistant built on Ollama and Open WebUI.
The cost changes nature. Instead of a per-user subscription that climbs with the team, you pay for a machine, or a hosted server, and the time to set it up. For an organization holding a few meetings a week, an existing workstation with a better graphics card can be enough. For one that transcribes every day, a dedicated server makes sense.
AI makes things up too, even when you host it
Hosting the pipeline fixes the leak. It doesn't fix the errors. Researchers from Cornell and other American universities analyzed Whisper and presented their results at the ACM FAccT conference in June 2024: about 1% of transcriptions contained entire made-up sentences that weren't in the audio, and 38% of those hallucinations included harmful content, such as violence or false personal information. It happened more often with speakers who talk with long pauses. Cornell notes that OpenAI has improved the model since and that the rate has gone down: the figure describes the version studied in 2023, but the mechanism hasn't disappeared.
Add the classic weaknesses of transcription: proper names, industry jargon, strong accents, overlapping voices. The summary step then brings its own errors: a language model can attribute a decision to the wrong person, turn a hypothesis into a commitment, or drop the caveat that changed everything.
The rule that follows fits in one sentence: someone who attended reviews the minutes before they go out, and reviews the decisions and assigned tasks first, because those are the lines people will act on. AI removes the chore of transcribing. It doesn't sign the minutes for you.
A six-line meeting rule
The most common risk isn't a tool badly chosen by management. It's a tool nobody chose: an employee or a partner who connects their own assistant to their calendar, exactly the pattern described in our post on shadow AI. A short rule, written down and known to everyone, settles most cases, and it fits naturally into your AI usage policy.
What we recommend putting in writing:
- Every transcription is announced at the start of the meeting, with the tool used and who will receive the minutes.
- No assistant joins a meeting on its own: automatic joining from the calendar is turned off.
- Some meetings are never transcribed: one-on-one HR conversations, health, litigation, negotiations.
- The audio is destroyed after a set period, once the minutes are approved.
- Someone who attended reviews the minutes before they're shared.
- An outside participant who shows up with their own assistant is asked to remove it, otherwise the meeting isn't transcribed.
At Blue Fox
Our approach starts from the previous section: the rule first, the tool second. We help write the meeting rule and the PIA that goes with it, then set up a minutes pipeline hosted in Québec, where the audio, the transcript and the summary stay under your control. The minutes land in our Meetings module for Odoo, next to the agenda, with decisions and assigned follow-ups ready to review.
For an organization that would rather keep a cloud tool, we also do the reverse exercise: put the vendor through the six questions, and say plainly whether its answers hold up.
Are your meetings already transcribed somewhere, or about to be? Let's talk about your meetings.
Sources
- NPR, August 15, 2025: the lawsuit filed against Otter.ai, consent asked of the host only, the training checkbox, and the investor story reported by the Washington Post.
- In re Otter.AI Privacy Litigation, order of August 13, 2026: the allegations about voiceprints and training, and the claims Judge Eumi K. Lee kept or dismissed.
- Act respecting the protection of personal information in the private sector: the PIA for an information system project (s. 3.3), informing people at collection (s. 8) and communication outside Québec (s. 17).
- Criminal Code, section 184: interception allowed with the consent of one of the people taking part in the communication.
- Civil Code of Québec, article 36: intercepting or using a private communication as an invasion of privacy.
- Principles for responsible, trustworthy and privacy-protective generative AI technologies: the December 7, 2023 text signed by the federal, provincial and territorial authorities, including the Commission d'accès à l'information.
- Zoom, clarification on its terms of service and AI: the commitment not to train models on customer content, published August 7, 2023.
- Microsoft, use Copilot without transcribing or recording a Teams meeting: the "Only during the meeting" setting and the possible retention of prompts and responses.
- Whisper: the MIT licence on code and weights, the six model sizes and the video memory each one needs.
- pyannote.audio: the MIT licence, the community model behind terms of use, and the publisher's paid service.
- Koenecke et al., "Careless Whisper: Speech-to-Text Hallucination Harms": the rate of hallucinated transcriptions and the share of harmful content.
- Cornell Chronicle, June 11, 2024: the FAccT presentation and the model's improvement since the study.