October 26, 2026 · 11 min read · Trust & Compliance

AI Meeting Notes for Compliance Officers in Financial Services: Keeping Recordings Off the Cloud

Published October 26, 2026

Key takeaways

• FINRA's 2026 Oversight Report (released December 9, 2025) added a dedicated GenAI section that identifies summarization as the top GenAI use case among member firms.
• SEC Rule 17a-4 and FINRA Rule 4511 apply to AI-generated meeting notes exactly as written — the tool is new, the obligations are not.
• Cloud AI notetakers create records on vendor servers; firms are responsible for capture, retention, and production regardless of where the artifact lives.
• On-device transcription is an architectural choice that eliminates the vendor-server retention surface — it is not a compliance badge.
• Compliance officers should evaluate processing location, retention defaults, training use, and record-production capability before approving any AI meeting tool.

Quick answer: For compliance officers at SEC- and FINRA-regulated firms, the risk with cloud AI meeting assistants is not the summary — it is where the audio, transcripts, and prompts live. FINRA's 2026 Oversight Report applies existing supervision and recordkeeping rules to GenAI as written. On-device transcription keeps those artifacts on the advisor's device, under firm control, with no vendor server copy to produce, subpoena, or breach.

Compliance officers at SEC- and FINRA-regulated firms are being asked a new version of a very old question: when an AI assistant writes up the notes from a client meeting, who holds the record, where does it live, and can the firm produce it when the examiner asks? The honest answer in 2026 is that the regulators have not written an AI-specific rulebook — they have made clear that the existing one already covers it. For compliance leaders, that means the choice of AI notetaker is less about summary quality and more about what the vendor leaves behind on its servers.

What FINRA Actually Said in Its 2026 Oversight Report

On December 9, 2025, FINRA published its 2026 Annual Regulatory Oversight Report. For the first time, the report contained a standalone section on generative AI that substantially expands on last year's treatment. The core message, as the Debevoise analysis summarizes, is that GenAI may implicate rules relating to supervision, communications, recordkeeping, and fair dealing — not that any new regime applies.

The report singles out one use case above all others. According to ArchiveIntel's summary of the 2026 report, FINRA identifies summarization as the most common GenAI use case among member firms, and the capture and retention of notes produced by AI meeting assistants is flagged as a top priority for compliance officers. In plain English: the auto-generated meeting write-up your advisor's AI assistant produces is treated as what it is — a record of that conversation.

The Rules That Already Apply

Three rules do the heavy lifting here, and none of them mention AI.

• SEC Rule 204-2 under the Investment Advisers Act of 1940 — the books-and-records rule for registered investment advisers — requires most records to be kept for at least five years, with the first two readily accessible.
• FINRA Rule 4511 mandates, in the words of a recent Global Relay enforcement summary, that firms create and maintain "legible, true, accurate, and complete" records of all business communications.
• SEC Exchange Act Rule 17a-4 governs how broker-dealers preserve electronic records, how long they must stay accessible, and what the firm has to hand over when the SEC or FINRA comes knocking.

ACA Global's analysis of the 2026 Report puts it as directly as the report itself does: GenAI does not alter firms' existing obligations around recordkeeping, supervision, outsourcing, or fair dealing with customers. The tool changes; the obligation does not.

Why Cloud AI Notetakers Create a Compliance Surface

Here is where the architectural choice bites. When an advisor uses a cloud-based AI meeting assistant, three artifacts are created and stored on a vendor's servers: the audio recording, the transcript, and (in many products) the prompts and model outputs. All three can qualify as records. All three now live on someone else's infrastructure.

The SEC's 2022 amendments to Rule 17a-4 explicitly addressed the cloud question. The amended rule contemplates broker-dealers maintaining electronic recordkeeping systems on servers owned or operated by a third party — a cloud service provider — provided the broker-dealer retains control of the system and independent access to the records. But as a First Mark Solutions analysis of the amended rule stresses, this is not a blanket cloud exception: if the firm has to ask the provider to transfer records, decrypt them, or restore access before it can reach them, the alternative undertaking is not the right fit.

Consumer AI notetakers generally are not built to that standard. Their terms grant the firm a login, not independent recordkeeping access. Their retention defaults are set by the vendor. Their export formats may or may not satisfy a request for a reasonably usable electronic format. The compliance question is not whether a vendor is "compliant" — it is whether the firm can account for the artifacts that vendor is creating.

The Off-Channel Communications Analogy

Compliance officers do not need to imagine what enforcement looks like when records live in the wrong place. They have watched the texting-fines campaign for five years. According to a Jatheon compliance review, the SEC and FINRA have collected more than $3.5 billion in texting fines since 2021, and the enforcement wave now reaches investment advisers, senior executives, and repeat offenders.

The pattern is directly transferable. A January 30, 2026 FINRA action fined a broker-dealer $750,000 for failing to supervise and retain at least 3,560 business-related text messages, citing violations of FINRA Rule 3110, FINRA Rule 4511, and SEC Exchange Act Rule 17a-4. The same statutory frame applies to AI meeting notes that live on a vendor server the firm does not fully account for.

SEC Chair Paul Atkins has signaled a different posture on the off-channel campaign specifically. As one 2026 industry analysis notes, Atkins described the texting-fines wave as "exactly how regulators should not act" and signaled a change in approach. The penalties may be paused; the recordkeeping rules have not changed. A notetaker that stores client conversations in a system the firm does not capture raises the same statutory question the texting cases did.

Cloud AI vs On-Device AI: The Compliance Officer's Comparison Table

The clearest way to see the difference is to look at what the firm actually has to account for in each architecture.

Dimension	Cloud AI Notetaker (Otter, Fireflies, Zoom AI Companion, etc.)	On-Device AI Notetaker (Basil AI)
Where audio is processed	Streamed to vendor servers	Processed on the advisor's device via Apple Speech Recognition
Where the transcript is stored	Vendor cloud by default	Local device + the firm's chosen archive (e.g., Apple Notes, firm DMS)
Default retention of audio	Vendor-defined, often indefinite	None by vendor — firm controls
Model-training on customer content	Varies by vendor and plan	Not applicable — audio never leaves device
Rule 17a-4 third-party undertaking exposure	Yes — vendor is a third party preparing or maintaining records	No vendor server copy to undertake
eDiscovery / subpoena surface	Vendor can be served directly for records	Records sit with the firm
Who the firm must do vendor diligence on	The AI vendor and its sub-processors	The device OS vendor (Apple) only
"Compliant" claim possible	No vendor is SEC/FINRA-compliant; it is the firm's determination	Same — architecture reduces surface, firm still owns compliance

Material Non-Public Information and the MNPI Problem

Asset managers, research analysts, and investment bankers routinely discuss material non-public information in meetings. Any cloud AI notetaker ingesting that audio becomes a secondary custodian of MNPI. The vendor's access-control model, its sub-processors, its data-residency defaults, and its model-training clauses all become relevant inputs to the firm's information-barrier program.

The 2026 FINRA Report explicitly asks firms to assess the third-party vendor's use of GenAI in their products or services and to include contract language that prohibits firm or customer sensitive information from being ingested into a third-party vendor's open-source GenAI tool. For MNPI-sensitive meetings, the simplest way to satisfy that language is architectural: do not send the audio to a third party in the first place.

What Compliance Officers Should Actually Check

Rather than asking vendors whether they are "compliant," ask them where the artifacts live and what the defaults are. A workable intake checklist:

1. Processing location. Where is the audio transcribed — on the device or on a vendor server? Request a plain-English answer, not a SOC 2 report.
2. Retention defaults. How long are audio, transcripts, and prompts retained by default? Who can change that setting? Is it enforceable per user or per firm?
3. Training use. Does the vendor use customer content to train, fine-tune, or evaluate models? Read the current terms, not the marketing page.
4. Record production. Can the firm produce records on demand from its own systems, or must it go through the vendor? If the latter, has the vendor filed an alternative undertaking under Rule 17a-4(i)?
5. Prompt and output logs. If the vendor keeps them, where? If the firm has to supervise prompts, as the 2026 Report's monitoring expectations suggest, how does it get access?
6. Sub-processors. Which third parties receive the audio or transcript — speech-to-text providers, LLM providers, analytics vendors, hosting providers? Each one is an addition to the firm's vendor-oversight surface.

For the broader framing — what the phrase "compliant AI meeting notes" actually means and does not mean — our deeper piece on what compliant AI meeting notes actually means walks through the four dimensions in detail.

Where Consumer AI Notetakers Get Compliance Officers in Trouble

The specific products are worth naming. Otter.ai's privacy policy describes a cloud-first architecture with broad rights to use customer content for product improvement. Fireflies.ai's privacy policy similarly describes server-side processing. Zoom's privacy statement addresses how meeting content is handled across its product suite. None of these are inherently disqualifying for every use case, but for a compliance officer at a FINRA member firm, all three raise the exact questions FINRA's 2026 Report asks firms to answer in writing: where is the data, who processes it, what is retained, who else can touch it.

The buyer-focused piece on granola alternatives for privacy-first users covers how "bot-free" and "cloud-free" are different claims — a distinction that matters enormously to a compliance officer evaluating these products.

How Basil AI Solves This

Basil AI is built on an architectural answer rather than a certification claim. Transcription runs on-device using Apple's Speech Recognition framework on the Apple Neural Engine. The audio does not stream to a Basil server for processing. Transcripts and summaries are generated locally and written to the user's chosen store — commonly Apple Notes, which syncs via the firm's existing iCloud configuration — rather than to a Basil vendor cloud.

What that means for a compliance officer:

• There is no Basil server copy of audio to subpoena, breach, or produce.
• Retention defaults are the firm's defaults, not a vendor's.
• Customer content is not used to train models — because the content never leaves the device.
• The vendor-oversight surface shrinks to the device OS, which the firm is already supervising for every other sanctioned app.

None of this makes Basil AI "SEC-compliant" or "FINRA-compliant" — as our longer analysis of AI meeting transcription for financial advisors explains, compliance is a firm determination, not a vendor badge. What on-device processing does is narrow what the firm has to determine. The capture, retention, and supervision obligations under Rule 204-2 and Rule 4511 still belong to the firm; the vendor-cloud exposure under Rule 17a-4(i) largely does not arise, because there is no vendor cloud holding the recording.

A Practical Written Supervisory Procedure for AI Meeting Notes

Based on the 2026 Report's expectations and the Rule 17a-4 framework, a workable WSP for AI meeting notetakers covers:

• Approved tools list. Every AI tool advisors are allowed to use, including the ones they signed up for on a phone, with its approved use written down. The 2026 Report's expectations are explicit on this point.
• Data restrictions. Which categories of information (PII, MNPI, PHI, client financial data) may not be ingested by which tools.
• Capture and retention. Where the firm will receive the AI-generated note, how it is indexed, and how long it is retained. For on-device tools, this typically means a routing step from the device into the firm's DMS or archive.
• Supervisory review. How a principal reviews AI-generated communications under FINRA Rule 3110, including sampling of prompts and outputs where feasible.
• Vendor diligence. For any tool that touches firm or customer data, documented assessment of processing location, retention, training use, sub-processors, and ability to produce records.
• Incident response. Procedure if an advisor uses an unapproved AI tool for a client meeting, including record capture after the fact.

The Direction of Travel

The 2026 Report is explicit that it "does not create any new legal or regulatory requirements or new interpretations of existing requirements." What it does is tell firms where examiners will be looking. The Global Relay summary of the FINRA 2026 Report notes that firms are expected to use the report to perform gap analysis of their compliance programs and incorporate the flagged risk areas into their risk assessments.

For compliance officers, the practical path is to treat the AI meeting assistant decision the same way the firm treated email archiving twenty years ago and text-message archiving five years ago: an architectural choice that determines how much supervision surface the firm takes on. The cloud-first consumer products put the surface on the vendor's infrastructure. On-device processing leaves it where the firm's existing supervision program already lives — on the advisor's managed device.

For a broader look at the technical architecture, see our deep dive on on-device AI meeting notetakers for Mac.

Download Basil AI