Skip to content

Territory 4 · Privacy duties

Meeting Transcription: Why It Calls for Its Own Evaluation

Meeting transcription brings together volume, sensitivity, identifiability and processing. What those four factors mean when Canadian advisors evaluate a tool.

Updated 11 min readSandra Lyne
A weathered slatted Prairie snow fence in open country, the drift formed on its lee side, the road beyond kept open.

Financial advisors use AI for many things. Scheduling assistance. Research queries. Draft writing. Meeting transcription. And there is a tendency to think about all of these in roughly the same category: "AI tools I use for productivity."

Transcription belongs in a category of its own.

Meeting transcription carries more client information, of a more personal kind, than any other AI use an advisor is likely to adopt. Seeing why means looking at four factors that meet in this one use, and at what removing names from a transcript does and does not achieve.

Privacy is the axis this page is about. On the securities-law axis, CSA Staff Notice and Consultation 11-348 (5 December 2024) (external site, opens in a new tab) says that "using an AI system directly in suitability determinations and trading decisions will be problematic unless a high degree of explainability is assured." A tool that drafts a record for the advisor to review does neither. Does CIRO Allow AI Meeting Notes? covers that side.


How much a single meeting contains

Consider what happens in a single scheduling query: a name, a date, maybe an email address. A research question might include a topic and some context. A draft request shares the information you choose to share.

Now consider what happens in a single client meeting.

In sixty minutes, a client might reveal their complete portfolio composition. Their health conditions. Their family structure, including the estranged adult child and the second marriage. The cardiac episode last year. The plan to retire in June 2027. The vacation property in Muskoka. Their mother's cognitive decline. The business interest they have not disclosed to their spouse.

A single 60-minute financial planning meeting can cover many distinct pieces of personal information about several people. One meeting can hold more client-identifying information than many other AI interactions combined.


What clients say in conversation

Other AI tools handle sensitive data too. But meeting transcription captures a specific kind of sensitivity: what clients share when they believe they are in confidence.

Clients do not self-censor in meetings the way they might in writing. They mention health conditions in passing. They reveal family dynamics while explaining estate planning goals. They share fears about capacity, about children's judgment, about what happens when they are gone. In email or written forms, clients are guarded. In conversation, disclosure happens in the natural flow. Meeting transcription captures information a client might never put in writing.

This is not business data. It is life data. Some of the most intimate details clients share with anyone outside their immediate family, captured verbatim, every session. One safeguard advisors reach for is the subject of the next section.


The re-identification reality under PIPEDA

Some advisors believe they have addressed the risk: "I scrub the names before uploading."

Re-identification works differently from what that step assumes.

A 2019 study in Nature Communications (external site, opens in a new tab) estimated that 99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes, such as age, gender, and marital status. A single client meeting can touch on more attributes than that. More recent research has demonstrated that large language models can re-identify individuals from anonymized text, making rich contextual transcripts even more vulnerable.

"A 67-year-old widow in Oakville with two adult children, one estranged, who had a cardiac episode last year, holds significant Suncor stock, and plans to retire from her consulting practice in June 2027."

That is not anonymous. That is one person. Removing the name made no difference.


Where the data goes

When advisors use popular transcription tools, where does that data travel?

Each vendor's own documents say where meeting data goes. As at 13 November 2026, Otter.ai's Privacy Policy (external site, opens in a new tab) names "Amazon Web Services, based in the United States" among the providers it relies on for computing and storage. Fireflies.ai's Privacy Policy (external site, opens in a new tab) states that it and its service providers "process and store personal information on servers located in the United States and other countries", and Fireflies offers Enterprise customers private storage (external site, opens in a new tab) "at your preferred location".

This matters because PIPEDA requires that Canadian personal information receive a comparable level of protection (external site, opens in a new tab) when processed abroad. Quebec's Law 25 (external site, opens in a new tab) requires a privacy impact assessment for any project to acquire an information system that involves personal information, and sets penalties in two tiers: administrative penalties up to $10 million or 2% of worldwide turnover, and penal fines up to $25 million or 4%.

And processing location is only part of the picture. Vendors' terms on model training differ. As at 13 November 2026, Otter.ai's Privacy Policy (external site, opens in a new tab) lists "training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)" among its uses of personal information. Rev's Terms of Service (external site, opens in a new tab) state that customer content "may be used for continuous training" of Rev's own speech recognition and AI models, and that it "will not be used for any generative AI model training." Fireflies.ai's Privacy Policy (external site, opens in a new tab) states: "We do not use personal information for AI model training". The terms vary, and the stakes are different in kind when the data is an hour of unfiltered client conversation.

Otter.ai is defending a class action in the United States. Brewer v. Otter.ai was filed in August 2025 in the United States District Court for the Northern District of California and now proceeds as In re Otter.AI Privacy Litigation (No. 25-cv-06911) (external site, opens in a new tab). The complaint alleges that Otter recorded conversations without consent and kept their contents to train its models. On 13 August 2026 the court granted Otter's motion to dismiss in part and denied it otherwise: the federal wiretap claim, the California wiretap claims, and the Illinois biometric claims go forward, and several other claims were dismissed with leave to amend. That order tests the pleadings only and does not decide whether the allegations are true. As at 13 November 2026 it is the most recent ruling found.

A second action, Cruz v. Fireflies.AI Corp. (external site, opens in a new tab), was filed on 18 December 2025 in the United States District Court for the Central District of Illinois and alleged collection of voiceprints without consent. The plaintiff withdrew it, and the court closed the case on 11 March 2026 without ruling on the allegations.


The intersection that creates categorical difference

Other AI tools share some of these characteristics. A research tool might process data internationally. A writing assistant might handle sensitive information. A scheduling tool might have training provisions.

But meeting transcription sits at the convergence of all of them:

Volume: More client information per use than any other AI use an advisor is likely to adopt.

Sensitivity: Some of the most intimate client information, shared in confidence, captured verbatim.

Identifiability: Enough attributes that a person can remain identifiable after the names are removed.

Processing: By a vendor, in some cases outside Canada, under terms on model training that differ from one vendor to the next.

This is not one risk factor. It is all of them, intersecting at the same point. That intersection creates categorical difference.


What evaluation looks like

Understanding why meeting transcription is different enables better evaluation. Not generic "is this secure?" questions, but specific criteria that address the actual risk.

The free Compliance Conversation Kit includes a Vendor Evaluation Checklist with Canadian-specific criteria for taking these questions to a compliance team.


Frequently asked questions

Which AI notetaker is PIPEDA compliant?

No AI transcription tool has been certified as PIPEDA compliant, because no such certification exists. PIPEDA compliance is an ongoing obligation of the organization using the tool, not a vendor feature. Advisors should evaluate each tool against PIPEDA's consent, data minimization, and cross-border transfer requirements.

Does Otter.ai use my data to train AI models?

Yes, according to its own policy. As at 13 November 2026, Otter.ai's Privacy Policy (external site, opens in a new tab) (effective 16 June 2026) lists among its uses of personal information "training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)". Otter states (external site, opens in a new tab) that it de-identifies user data before training and that no person reviews a recording without the customer's explicit consent. Allegations about this practice are part of In re Otter.AI Privacy Litigation (external site, opens in a new tab), a class action in the United States that began as Brewer v. Otter.ai in August 2025.

Is a voice recording personal information under PIPEDA?

Yes. PIPEDA defines personal information as any information about an identifiable individual, recorded or not. A meeting recording containing a client's name, financial circumstances, health information, and family details constitutes personal information regardless of format. Law 25 (external site, opens in a new tab) names biometric information as an example of sensitive personal information, and Quebec's privacy regulator (external site, opens in a new tab) lists the voice among the characteristics biometric systems analyse. A recording becomes biometric information when a system analyses the voice to identify or authenticate the speaker.

Can de-identified meeting data be re-identified?

Research shows that it can. A 2019 Nature Communications (external site, opens in a new tab) study estimated that 99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes. A single client meeting can touch on more attributes than that. In PIPEDA Findings #2026-001 (external site, opens in a new tab) (5 March 2026), the Office of the Privacy Commissioner of Canada wrote that "the risk of re-identification is not a static consideration and may increase over time".

Do CIRO regulations cover AI meeting recordings?

As at 13 November 2026, CIRO has published no rule or guidance note specific to AI. Staff Notice 11-348 (external site, opens in a new tab) sets out how existing securities law applies when market participants, including registered dealers and advisers, use AI systems. CIRO's Compliance Report for 2026 (Administrative Bulletin 26-0034, released 17 February 2026) (external site, opens in a new tab) says its examiners will be asking about the use of AI in dealers' operations. Meeting transcripts containing client instructions or recommendations likely qualify as business records subject to CIRO's documentation and retention requirements.



The framework for this category

Meeting transcription requires different evaluation than other AI tools because it occupies a different risk category. The criteria that might be sufficient for a scheduling tool or a research query do not address what is actually at stake with transcription.

CIRO's dealer members had about 109,500 Approved Persons (external site, opens in a new tab) as at 31 March 2025, and CIRO concluded 48 enforcement proceedings (external site, opens in a new tab) in its 2026 fiscal year (1 April 2025 to 31 March 2026), 39 of them against individuals. That record suggests careful professional judgment. The same careful judgment that serves clients well in investment decisions should extend to how their most sensitive information is handled.

The evaluation questions above are the place to start with any transcription tool, including the one already in use.

Written by

Sandra Lyne

Founder, Northern Catalyst

TagsPIPEDAPrivacyFinancial AdvisorsAI RiskMeeting Transcription

This is educational content based on published regulatory sources as at the page's last-updated date. It is not legal, compliance, or financial advice. Review specific compliance decisions with your compliance officer and qualified legal counsel before implementation.

Source: PIPEDA s. 5 and Schedule 1; Quebec Law 25

Re-verified