Before you worry about what the AI wrote, work out how many people read it. Under the SEC’s marketing rule, an advertisement is a communication an adviser makes “to more than one person, or to one or more persons if the communication includes hypothetical performance,” that offers the adviser’s advisory services (17 CFR 275.206(4)-1(e)(1)). Extemporaneous live oral communications are excluded. So is information contained in a required regulatory filing.
Sort your own AI output across that line and the answer runs opposite to what most advisory practices are bracing for.
The AI-drafted meeting recap you send to one client, carrying no hypothetical performance, is not an advertisement. The paragraph on your website describing your process as AI-driven is, because your site is read by more than one person and it offers your services. Same technology, opposite regulatory treatment. The output every firm proofreads twice sits outside the rule, and the sentence a marketing contractor wrote two years ago sits inside it.
What the SEC has actually charged advisers with
The settled cases are the fastest way to see where the risk really sits, because they point in one direction.
On March 18, 2024 the SEC settled charges against two investment advisers, Delphia (USA) Inc. and Global Predictions Inc., for false and misleading statements about their purported use of AI, with the firms paying $400,000 in total civil penalties: $225,000 from Delphia and $175,000 from Global Predictions (SEC press release 2024-36). Delphia’s statements appeared in SEC filings, a press release and on its website. Global Predictions’ appeared on its website and on social media, where it called itself the “first regulated AI financial advisor” and promised “[e]xpert AI-driven forecasts.” Both firms were charged with violating the marketing rule.
On October 10, 2024 the SEC charged Rimar Capital USA, Rimar Capital LLC, Itai Liptz and Clifford Boro over false and misleading statements about Rimar LLC’s purported use of AI to perform automated trading, settling for $310,000 in total civil penalties, with Liptz separately consenting to pay disgorgement and prejudgment interest of $213,611 (SEC press release 2024-167). That one was an antifraud case rather than a marketing rule case, and it involved money raised from investors, so it is a different animal. It shares the one feature that matters here.
In none of these cases did a model make a mistake. Not one of these firms was penalized because an AI summarized a portfolio wrong, misread a document or produced a bad recommendation. Every firm named above was penalized for a sentence a human wrote describing the software. As SEC Chair Gary Gensler put it in the 2024 announcement, “Investment advisers should not mislead the public by saying they are using an AI model when they are not.”
That is the practical inversion. Firms spend their AI governance budget reviewing model output, which is the artifact least likely to be read by more than one person, and spend nothing reviewing the marketing copy, which is the artifact the rule is actually about.
The clause that changes how you write the sentence
One line in the marketing rule does most of the work, and it is not the one about untrue statements. An advertisement may not “include a material statement of fact that the adviser does not have a reasonable basis for believing it will be able to substantiate upon demand by the Commission” (17 CFR 275.206(4)-1(a)(2)).
Read that as a writing instruction rather than a legal standard. It means every factual claim you make about your AI has to come with an answer to “show me.” Not an intention to be honest. A reasonable basis, held at the time you published, for believing you could produce the evidence if asked.
“We use AI to help prepare client materials” is a claim you can substantiate with a subscription invoice and a description of the workflow. “Our AI analyzes market conditions to optimize your portfolio” is a claim that requires you to show something that analyzes market conditions and something that changes as a result. The gap between those two sentences is the entire distance between Delphia’s marketing and Delphia’s software.
There is a second clause worth pinning next to it. An advertisement may not discuss potential benefits “without providing fair and balanced treatment of any material risks or material limitations associated with the potential benefits” (17 CFR 275.206(4)-1(a)(4)). A page that says AI makes your research faster, and says nothing about what it cannot do, has a structural problem no amount of accuracy fixes.
The record clock nobody starts
Advisers must keep a copy of each advertisement they disseminate (17 CFR 275.204-2(a)(11)(i)(A)). Those copies are held “for a period of not less than five years, the first two years in an appropriate office of the investment adviser, from the end of the fiscal year during which the investment adviser last published or otherwise disseminated” the communication (17 CFR 275.204-2(e)(3)(i)).
Note where the clock starts. Not when you wrote the page. When you last disseminated it. A live web page describing your AI is being disseminated today, which means the five-year clock has not started yet and will not start until the page comes down or changes.
Client emails are a separate obligation and are not off the hook either: advisers keep copies of written communications sent relating to any recommendation made or advice given (17 CFR 275.204-2(a)(7)). An AI-drafted client email is usually not an advertisement. It is very often a record.
Four moves, in order
1. Write one paragraph describing what your AI actually does, in verifiable terms. Name the tool, the task, and what a human does before anything reaches a client. Keep it in the same folder as the evidence: the invoice, a saved example, the vendor’s own documentation of the feature. This paragraph is now the source text every other piece of copy has to agree with, and it takes an afternoon.
2. Sweep every place you have already described it. Your website, your Form ADV brochure, LinkedIn, any pitch deck sent to more than one prospect, email signatures, conference bios, and anything a contractor wrote while you were busy. Global Predictions was charged on statements that appeared on its website and on social media, so the sweep has to cover both. Read each claim against the paragraph from move 1 and delete anything you cannot show.
3. Start keeping dated copies, then check the retention math. Two real options, both priced publicly. Stillio takes automatic scheduled screenshots of pages you nominate, at $29 a month for up to 5 pages, $79 for up to 25, $199 for up to 100 and from $299 for unlimited, with a 14-day free trial. Read its retention line closely before you rely on it: all plans include 36 months of screenshot retention, which is shorter than the five years the rule asks for, so export the captures into your own storage rather than treating the subscription as the archive. Visualping solves the adjacent problem of noticing when a page changes without telling you, with a free tier covering 150 checks a month across 5 pages at hourly frequency, and Personal 1K at $14 a month for 1,000 checks across 10 pages. The dedicated compliance archiving vendors aimed at advisers, including Smarsh, Pagefreezer and MirrorWeb, quote rather than publish, so budget a sales conversation if you want that tier.
4. Put the limitation in writing next to the benefit. Wherever your copy claims AI speeds something up, add the plain sentence about what it does not do and who checks it. This is the cheapest of the four moves and the one that answers the fair-and-balanced requirement directly.
What this does not cover
This is a description of published rules and enforcement actions, not legal advice, and your compliance consultant or counsel is the person who should sign off on your actual copy. The marketing rule at 206(4)-1 applies to SEC-registered advisers; Rimar Capital LLC was state-registered, and state-registered firms work under their own regime alongside the federal antifraud provisions. The amended marketing rule has been in force for a while now, with a compliance date of November 4, 2022, and the SEC staff maintain a running Marketing Compliance FAQ, last updated January 15, 2026, which states plainly that staff FAQs have no legal force or effect.
Beyond the advisory rulebook, the general prohibition on unfair or deceptive practices applies to AI claims across every industry, which we covered in this guide to AI ethics and regulations for businesses, alongside the wider 2026 AI compliance picture for firms that are not advisers.
So where does the Anthropic news fit
Underneath all of this sits the platform layer, moving quickly. On February 24, 2026 Orion announced it was “expanding its ongoing collaboration with Anthropic” following Anthropic’s introduction of financial services plug-ins, with Reed Colley, President of Orion Advisor Technology, saying the firm looked forward to “working with these new plug-in capabilities as well as other enhancements as they become available” (Orion newsroom). Anthropic’s own Claude for Financial Services offering, announced July 15, 2025, connects to data providers including FactSet, Morningstar, PitchBook and S&P Global and is sold through its sales team rather than off a price list.
For a small practice, the useful thing about that news is not the capability. It is that none of it changes who owns the sentence. Whatever your custodian, your TAMP or your portfolio software builds on top of a model, the description of it on your website is your advertisement, substantiable by you, retained by you. Platform announcements generate marketing copy inside advisory firms faster than they generate workflows, and the copy arrives first.
One more caution that belongs to this layer rather than the rulebook: an adviser feeding client information into a general-purpose chatbot should know exactly who can read it, which is a different question from whether the output is any good, and we worked through who can actually read a work chat separately.
The part I would argue for
Everything above is documented. This next bit is opinion, and it runs against the way AI in advice usually gets sold.
The substantiation requirement is quietly the most pro-human clause in the rulebook. It says that somebody at your firm has to be able to stand up and explain what the software does, in specifics, on demand. You cannot satisfy that with a vendor’s brochure or a model’s own confident summary of itself. It requires a person who understands the workflow well enough to describe its limits, and that person becomes more valuable the more of the drafting the machine takes over, not less.
Which is why the version of this that worries me is the firm that adopts the tooling and thins out the people who would have been able to answer for it. That firm has not reduced its costs. It has published claims it can no longer substantiate and kept them live, restarting the five-year clock every day, with nobody left who could reconstruct why the sentence was written.
Open your own website right now and find the sentence where you describe your technology. Ask what you would hand over if someone asked you to show your work on it. If the answer takes longer than a minute to assemble, that is the sentence to fix first, and it is a great deal cheaper to fix than to defend.
Frequently asked questions
Is an AI-drafted email to one client an advertisement under the SEC marketing rule?
Generally no. The rule defines an advertisement as a communication made “to more than one person, or to one or more persons if the communication includes hypothetical performance,” that offers advisory services (17 CFR 275.206(4)-1(e)(1)). A one-to-one client email carrying no hypothetical performance falls outside that definition, though it may still be a required book and record if it relates to a recommendation made or advice given (17 CFR 275.204-2(a)(7)).
What has the SEC actually fined investment advisers for regarding AI?
For what they said about AI, not for what AI did. Delphia (USA) Inc. and Global Predictions Inc. paid $400,000 in total civil penalties in March 2024 over false and misleading statements about their purported use of AI, split $225,000 and $175,000 (SEC press release 2024-36). In October 2024 the Rimar Capital entities, Itai Liptz and Clifford Boro settled for $310,000 in total civil penalties over similar misrepresentations (SEC press release 2024-167).
How long do I have to keep a web page that describes my firm’s use of AI?
Copies of advertisements are kept “for a period of not less than five years, the first two years in an appropriate office of the investment adviser, from the end of the fiscal year during which the investment adviser last published or otherwise disseminated” the communication (17 CFR 275.204-2(e)(3)(i)). The clock runs from last dissemination, so a page that is still live has not started it.
What does it mean to substantiate an AI claim on demand?
An advertisement may not include “a material statement of fact that the adviser does not have a reasonable basis for believing it will be able to substantiate upon demand by the Commission” (17 CFR 275.206(4)-1(a)(2)). In practice that means holding, at the time you publish, the evidence that would show the claim is true: the tool, the workflow it runs in, and what a person does before the output reaches a client.
