On Monday, September 14, 2026, two things happened in the advisory industry and almost nobody connected them.
Anthropic launched Claude for Financial Advisors at the Future Proof Festival, with connectors into Charles Schwab, Orion, BlackRock, Addepar, Envestnet, iCapital, SS&C Black Diamond, Wealthbox, Wealth.com, Vanguard and Zocks. Schwab Advisor Services is the first RIA custodian in the lineup. Advisors can now point an AI assistant at live custodial balances, positions, transactions and cost basis, at portfolio reporting, at CRM records, at estate and tax documents, and at meeting transcripts.
The same day, the SEC Division of Examinations published a risk alert on adviser annual compliance reviews, describing the specific ways firms had failed to conduct, scope, document and act on those reviews.
One of those events gives your advisors a large new capability. The other sharpens the standard your compliance program will be measured against. They landed on the same Monday, and most firms are only reading one of them.
What actually launched
Claude for Financial Advisors is not a chatbot with a finance theme. It is a set of connectors into the systems advisory firms already run, plus eight prebuilt workflow skills: advisor onboarding, alternative investments brief, compliance and AI policy, estate and tax brief, portfolio rebalance review, post-meeting notes and follow-up, pre-meeting prep, and prospect intake.
It also works with connectors firms already use, including Microsoft 365, Salesforce, DocuSign, Box, FactSet, S&P Global and Morningstar. Press coverage puts the cost in the range of seventy to one hundred twenty dollars per user per month, and firms requesting a license before September 30, 2026 receive a one-time usage credit.
Anthropic is direct about the boundary. In its own product documentation, it states that investment recommendations, client communications, compliance determinations and other regulated activities remain subject to human review and approval. That sentence is worth reading twice, because it is the whole compliance question in one line. The vendor has told you where its responsibility stops.
The rule that already governs this
There is no new AI rulebook coming to rescue you, and you do not need one. FINRA said so plainly in Regulatory Notice 24-09, issued June 27, 2024: the securities laws and FINRA rules are technology-neutral, and they apply to generative AI and large language models exactly as they apply to any other technology. The notice created no new requirements and relieved firms of nothing.
That is the correct frame, and it is more demanding than it first sounds. Adopting AI does not add a rule to your program. It adds new ways to break the rules you already have. Every obligation you had on September 13 you still had on September 15, except now some of the work producing your client communications, your meeting records and your portfolio analysis is being done by a system that is confident when it is wrong.
Six things your program owes now
One. A written AI policy. If your firm uses AI in advisory, marketing, operational or supervisory activities, Rule 206(4)-7 requires policies reasonably designed to prevent violations in that activity. For state-registered advisers the NASAA Model Rule reaches the same place. For broker-dealers, FINRA Rule 3110 requires a supervisory system that reasonably covers the technology, including governance, model risk and data integrity. A policy that does not mention AI is not a policy that covers AI.
The policy needs to answer concrete questions. Which tools are approved, and for what. Which tools may touch client data and which may not. What a person must verify before output goes anywhere. Who approved the connector into your custodian, and what scope it was given.
Two. Human review before anything reaches a client. This is the single control that carries the most weight, and it is the one most likely to erode quietly. It erodes because AI output reads well. A summary that is fluent and plausible gets skimmed rather than checked, and the figure inside it goes out unverified.
The reviewer has to confirm that every factual assertion is supported, that every figure traces to a source the firm can produce, that nothing promises a result, and that required disclosures are present. And the review has to be recorded, because a review nobody can evidence did not happen as far as an examiner is concerned.
Three. Marketing Rule review that is yours, not the vendor's. Claude for Financial Advisors includes a compliance skill that screens client-facing language against the SEC Marketing Rule and flags potential issues. That is a genuinely useful input. It is not your review.
Rule 206(4)-1 obligations sit with the firm. If an AI tool screens a piece, misses something, and the piece goes out, the firm answers for it. A vendor's screening feature is a first pass that helps your reviewer, and treating it as the supervisory record itself is the mistake we expect to see written up first.
Four. Books and records for AI-assisted work. Rule 204-2 and Exchange Act Rule 17a-4 do not care which software drafted the document. If an AI tool produces a client communication, a meeting summary, or an analysis your firm relies on, that output is a record and it has to be retained.
This one catches firms off guard because AI-generated meeting notes feel like scratch paper. They are not. If the note informs advice or documents a client interaction, it is a record, and it needs to sit inside your retention program rather than in a tool nobody archives.
Five. Vendor diligence on every connector. Each connector you switch on is a service provider with access to client data. Regulation S-P's service-provider expectations apply, and so does your own vendor oversight policy.
Ask the questions before you connect, not after. What does the vendor do with your inputs, and are they used to train models. Where is data processed and stored. What can the connector read, and how is that scope limited. What happens to your data when you terminate. Then write the answers down, because the diligence you cannot evidence is diligence you did not do.
Six. Training, and a test in your annual review. Your people need to know how these tools fail before they rely on them. AI produces confident, well-formatted, incorrect output, and an advisor who has never seen that happen will not be looking for it.
Then it has to be tested. Which brings us back to the risk alert that landed the same day.
Why the timing matters more than it looks
The SEC's September 14 risk alert described firms that skipped annual review years entirely, that reviewed periods longer than twelve months, that left pivotal risk areas out of scope, that relied on training and attestations instead of actually reviewing anything, that kept no documentation of testing, and that identified problems and then never fixed them.
Read that list next to a firm that adopted an AI assistant in September and did not update its policies, its training, its recordkeeping or its testing plan. You have a new material activity, touching client data and client communications, that appears nowhere in the compliance program. That is the definition of a pivotal risk area left out of scope, and the staff just told you they are looking for it.
The firms that handle this well will name AI in the annual review plan, test it, document what they found, and fix what did not work. The firms that do not will discover the gap when someone else finds it for them.
What this is not
This is not a reason to avoid the tools. An advisor who can prepare for a client meeting in ten minutes instead of ninety is delivering better service, and refusing useful technology is not a compliance strategy.
It is also not a reason to panic-write a policy nobody follows. A two-page document that says the firm takes AI seriously, adopted and never referenced again, satisfies nothing. What examiners test is whether the written procedure describes what the firm actually does, and whether you can show the work.
How Compliance Approved handles this
We ship a written Artificial Intelligence and Automated Tools policy, mapped to your registration type, citing Rule 206(4)-7, Section 206, Rule 204-2, the Marketing Rule, the NASAA Model Rule, FINRA Rules 3110 and 2210, and Regulatory Notice 24-09. It covers approved tools, prohibited uses, the human review requirement, vendor diligence, connector authorization, recordkeeping, an annual testing procedure, and training. It is a working document with the testing steps written out, not a statement of intent.
Our AI marketing review reads your client-facing material against the Marketing Rule and returns findings with the specific language at issue, so the review is yours and the record is yours. Our communications archive takes email and Teams directly through a Microsoft 365 connector, and takes SMS, chat and social by import or API from whatever archive you already keep. Everything that lands is screened for risk, unflagged correspondence is sampled, and the supervisory record an examiner asks for is kept. That is where AI-drafted client communications end up whether or not anyone planned for it.
Vendor and service-provider oversight tracks each connector as what it is. The training library carries the firm-element material your people need. And the annual review module runs the firm's own data against the September 14 risk alert, flagging missing review years, periods over twelve months, published policies that require annual testing but appear nowhere in scope, and prior-year findings that were never dispositioned.
The short version
Your advisors are going to use these tools, and they should. FINRA already told you the rules do not change. What changes is the surface area those rules apply to, and the honest question is whether your program covers an activity that did not exist in your firm a week ago.
Write the policy. Require the review and record it. Do your own Marketing Rule review. Retain the output. Diligence the connectors. Train the people. Then test it in your annual review and write down what you found.
None of that is exotic. It is the same compliance program you already run, extended to a system that now touches your client data.