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

Why the Compliance Officer, Not the Advisor, Owns This Decision

Across broker-dealers, RIAs, hedge funds, and asset managers, AI meeting assistants are now in the room on nearly every client call. The vendor pitch is simple: automatic transcription, action items, CRM sync. The compliance problem is harder: an AI-generated summary of a conversation about recommendations, advice, or material non-public information is a business record the moment it exists, and your examiner will treat it that way regardless of which laptop it came from.

That is the backdrop against which FINRA released its 2026 Annual Regulatory Oversight Report on December 9, 2025 — a nearly 90-page document that, for the first time, carries a dedicated standalone section on generative AI. The headline for compliance officers is this: the rules do not move; the examiner's vocabulary does. If your firm has not yet written the AI meeting notetaker into its written supervisory procedures (WSPs), the governance gap is already open.

What Changed in 2025–2026 — and What Did Not

Three things changed in the last twelve months, and none of them was a new AI-specific rule.

First, the regulators themselves confirmed the technology-neutral stance. Goodwin's analysis of FINRA's 2026 Report describes it as spotlighting generative AI and cybersecurity as 2026's key areas of concern, while reiterating that firms remain responsible under existing rules. McGuireWoods' client alert notes that FINRA identified "summarization and information extraction" — the meeting-note use case — as the single most common GenAI deployment among member firms.

Second, enforcement for electronic recordkeeping failures has not slowed. According to Legal Dive's reporting on the January 2025 sweep, the SEC fined nine investment advisers and three broker-dealers a combined $63 million for off-channel communications violations. As Global Relay noted, since 2021 the SEC's "communications compliance crusade" has levied over $3 billion in penalties against more than 100 firms. In a single prior sweep the SEC charged 26 firms a combined $390 million for widespread failures to maintain and preserve electronic communications. An unapproved AI notetaker capturing client discussions outside the firm's archive is precisely the kind of unmonitored channel this program was built to find.

Third, FINRA has quietly shifted toward personal liability. Corporate Compliance Insights documents that in October a former Wells Fargo Advisors broker was fined and suspended for off-channel messaging and deleting the evidence, and in 2026 FINRA barred an individual outright from associating with any member firm for off-channel use. Where the SEC sweep largely targeted institutions, FINRA is now holding individual registered persons personally accountable. The advisor who installed an unapproved cloud notetaker last quarter is the person whose BrokerCheck record changes.

What AI Meeting Output Actually Is, in Regulatory Terms

Strip the marketing away and the AI meeting assistant produces four artifacts the compliance officer has to classify: (1) the raw audio of the client call, (2) the transcript, (3) the summary with action items, and (4) in some cases the prompt logs the model consumed. Each of those lives somewhere, for some retention period, under someone's supervision — or it is a gap.

Two regulations do the heavy lifting here. SEC Rule 17a-4, as amended effective January 3, 2023, sets the electronic recordkeeping standard for broker-dealers; the SEC staff FAQ explains that firms may now use either write-once-read-many (WORM) storage or an audit-trail alternative that preserves original records with time-stamped modifications. FINRA Rule 4511 is the general books-and-records rule for broker-dealers, defaulting to a six-year retention period where no other period applies. For RIAs, Advisers Act Rule 204-2(a)(7) covers written communications relating to recommendations and advice for five years, the first two in an easily accessible place.

Here is the pivot: the duty attaches to the record, not to the tool that produced it. An AI-generated summary of a client call about rebalancing a managed account is a written communication about advice. It is a book and record. The fact that a language model drafted it does not change its classification. Your WSPs either account for where it lives and how it is supervised, or they do not.

The Cloud-vs-On-Device Decision for Meeting Notes

Every AI notetaker makes an architectural choice about where audio is processed. That choice determines your compliance surface — meaning the set of third parties, data copies, and legal interfaces your firm has to supervise, document, and defend.

Note the last row. On-device processing does not eliminate the recordkeeping obligation — if the summary is a record, your firm still has to retain and produce it under data-minimization principles that European firms apply alongside US rules, and under 17a-4 and 4511 at home. What on-device does is collapse the number of parties who hold the raw material to one: the firm.

The Off-Channel Communications Lens

The off-channel enforcement program is the single most useful mental model for AI meeting notetakers. Here is why: the SEC's theory in those cases was not that WhatsApp or iMessage was intrinsically bad; it was that firms allowed business communications to occur on channels their archives did not capture, which prevented examiners from reconstructing the record.

An unapproved cloud notetaker fits that pattern exactly. An advisor downloads it, lets it join a client Zoom, and now a verbatim transcript of a conversation about recommendations sits on a vendor's server — outside the firm's archive, outside its supervisory procedures, and outside its ability to produce it on demand. If an examiner asks "show me the complete record of your communications with client X during Q2," the firm cannot answer honestly without touching that vendor.

This is why vendor due diligence under Regulation S-P and the firm's written supervisory procedures must specifically name which AI meeting tools are approved and where their outputs go. For context on the broader shift toward private, on-device architectures, see our earlier analysis of AI meeting notes for asset managers handling MNPI.

The Compliance Officer's AI Notetaker Checklist

A defensible approval for any AI meeting notetaker — cloud or on-device — should document the following before deployment. Treat each item as a question the examiner will eventually ask.

1. Where is the audio processed?

If the answer is "a vendor cloud," you have a data-processor relationship to document and supervise. If the answer is "on the advisor's device," you have a device-management question instead, which your MDM program likely already answers.

2. What does the vendor retain, for how long, and under what default?

Zero-day retention options exist at some vendors but are not the default. Audio deleted after processing is a different exposure than audio stored indefinitely. Document the configuration, not just the capability.

3. Does the vendor train on client data?

Model-training defaults vary. Opt-outs exist but are often buried in enterprise contracts. For MNPI-adjacent conversations the answer must be no, in writing, with contractual remedies.

4. Where does the AI-generated note end up as a business record?

A note that lives only inside the vendor's app is a recordkeeping gap. The firm's archive needs the note, with its metadata, in a format Rule 17a-4 recognizes as either WORM or audit-trail compliant.

5. Who supervises the AI's output before it becomes the record of advice?

FINRA Rule 3110 requires reasonable supervision. The 2026 Report specifically emphasizes human review of model outputs. The advisor is not the supervisor of their own notes.

6. Is the advisor's use of the tool consistent with the WSPs?

An approved tool used in an unapproved way is still a finding. The written procedures must match the practice.

Comparing the Market: What the Major Notetakers Actually Keep

The practical question for a compliance officer evaluating this market is whose architecture matches the firm's risk tolerance. The vendor policies below are the ones a compliance officer would actually read.

The point of the table is not to declare a winner. It is to make the architectural trade-off visible so the compliance officer can match the tool to the firm's actual risk appetite. For a deeper comparison of individual vendors, see our notetaker comparison guide.

SEC Rule 17a-4: WORM, Audit Trail, and What Your Archive Must Do

The 2022 amendments to SEC Rule 17a-4 gave broker-dealers an alternative to WORM: an audit-trail system that preserves original records in a manner permitting reconstruction if any element is modified or deleted. The amendments also require that records be furnished to the Commission in a reasonably usable electronic format.

Practically, this means an AI-generated meeting summary that qualifies as a business record must sit in a system that is either WORM or produces a complete time-stamped audit trail of edits. A notetaker that lets any user quietly rewrite the summary after the fact — with no immutable original and no audit trail — is not a recordkeeping system; it is a draft tool. The downstream archive is where the record obligation is satisfied.

This cuts both ways for cloud and on-device tools. Neither architecture is a 17a-4 archive on its own. Both need integration with the firm's archive of record. The question is whether the raw audio and the first-pass transcript — the material behind the summary — exist in a form the firm can retrieve, or exist only on a third party's servers under terms the firm did not write.

How Basil AI Solves This

Basil AI is built around a single architectural choice: audio is captured and transcribed on the device, using Apple's Speech framework running on the Apple Neural Engine. No audio file is uploaded to a Basil server. No transcript is uploaded to a Basil server. There is no Basil cloud copy to subpoena, breach, train on, or produce in discovery, because the architecture does not create one.

For a compliance officer, this inverts the usual diligence conversation. Instead of asking "what does the vendor do with our client audio," the question becomes "what does our firm do with the notes the advisor creates on their own device." That is a question the WSPs, the MDM program, and the archive of record are already built to answer.

What Basil does not do, and what no AI meeting tool can do, is make the compliance determination for the firm. The firm still writes the policy, decides which meetings get recorded, documents client consent, routes the resulting notes to its archive of record, and supervises their use. Basil provides the architectural fact — on-device processing, no vendor data copy — on which the firm's compliance program can rest. Compliance remains the firm's determination. For the technical architecture behind on-device capture, see our deep dive on Apple Intelligence and on-device meeting notes, and for a practical competitor view see our analysis of what the Otter.ai eavesdropper ruling means for notetaker buyers.

What to Put in Your Written Supervisory Procedures This Quarter

Three concrete additions a compliance officer can make to the WSPs this quarter — none of which require waiting for new rulemaking:

Approved AI meeting tools list. Name every AI notetaker the firm permits by product name and version, with the approved use case. Everything not on the list is prohibited. This matches the model the firm uses for approved communications channels today.

Routing to the archive of record. For each approved tool, document where its output goes to satisfy the recordkeeping obligation — the email archive, the CRM, the dedicated compliance archive. The output cannot be allowed to live only inside the tool. Rule 17a-4 does not care which product generated the note; it cares whether the firm can produce it.

Supervision checkpoint. Specify who reviews AI-generated meeting summaries before they become the firm's record of a client interaction. The 2026 Oversight Report is explicit that ongoing human monitoring of model outputs is essential; the WSPs must say by whom and how often.

The Bottom Line for Compliance Officers

The 2026 regulatory posture is clear: existing rules reach generative AI exactly as far as they reach everything else. Nothing in that framework prohibits cloud AI notetakers, and nothing in it blesses on-device ones. The compliance officer's job is to match the tool's architecture to the firm's risk tolerance, document the choice in the WSPs, route the output to the archive of record, and keep the examiner's reconstruction question answerable: "show me the complete record of what you communicated with this client."

On-device processing is not a compliance guarantee. It is an architecture fact — the vendor server copy does not exist — that narrows the surface a compliance officer has to supervise. In an enforcement environment where off-channel programs have collected over $3 billion since 2021 and FINRA is now charging individuals personally, a narrower surface is worth the diligence.