AI for Wealth Management (2026): What a Small RIA Can Actually Use, and the Two Rules That Decide What It Can Say

AI for wealth management: the short answer

If you run a small advisory firm, the useful version of AI in 2026 is not a platform with "wealth" in its name. It is a $20-to-$25-a-month business seat, four repeatable jobs, one job it never does alone, and two rules you already live under. The four jobs: meeting prep and recap from your own notes, client drafts from your approved templates, first-read summaries of long documents, and a shared inbox cleared by drafts a human approves. The job it never does alone: advice, or anything that reads like a performance claim. The two rules: the SEC Marketing Rule, which does not care who or what wrote the draft, and the books and records rule, which wants a copy of it for five years. Everything below is the small-firm version, from someone who installs this for owner-led businesses and is not your compliance officer. Your CCO makes the compliance call; this page tells you what to bring to that conversation.

What the vendors mean by "AI for wealth management" and why it is not your problem yet

Search the phrase and you get enterprise: portfolio analytics inside a custodian platform, next-best-action engines, AI layered on a CRM built for a thousand advisors. Those products exist for firms with a technology budget and a project manager to run the rollout. A ten-person RIA does not need any of it to get the first year of value.

The first year of value is boring. It is the ninety minutes before a client review that go into pulling notes together, the forty-five minutes after it that go into the recap email, the hour of reading a trust document to find the two clauses that matter, and the inbox that eats the operations person's morning. A general assistant on a business plan does all four of those today, with your templates and your notes, for the price of a lunch per seat. Start there. The platform decision can wait until the workflows are proven.

The four jobs a small firm should give AI this quarter

1. Meeting prep and recap, from your notes

Before a review, paste last meeting's notes, the agenda, and any open items, and ask for a one-page brief: what changed, what was promised, what to ask. After the meeting, paste your raw notes and ask for a recap in the firm's voice, with action items and owners. The assistant does not need the client's account data to do either of these well; it needs your notes, which is the point. The recap goes to the client only after an advisor reads it, and it is filed where the firm files client correspondence.

2. Client communications from approved templates

Market-moment notes, onboarding checklists, "here is what happens next" emails, annual review scheduling. Give the assistant the approved template and the specifics, and it produces a clean draft in seconds. The discipline is the template library: the firm writes and approves the templates once, and the assistant only ever fills them in. That is also how you keep a model from inventing a claim, because the claims live in language a person already approved.

3. First-read summaries of long documents

Statements from an outside custodian, a trust or estate document, a 401(k) plan proposal, a 60-page insurance illustration. Ask for a summary, the defined terms, the dates, and the three questions an advisor should ask. This is a first read, not the read. The advisor still reads the document; the assistant just tells them where to look first. On a business plan that does not train on your data, this is one of the highest-return uses in the firm.

4. The shared inbox, with a person on the approve button

Scheduling requests, document requests, "did you get my form," "what is your address." An assistant that drafts replies for a human to approve clears the routine half of an operations inbox before lunch. The rule I install everywhere: the assistant drafts, a person sends. Nothing reaches a client without a human click. That single rule answers most of the supervision questions before they are asked.

The one job it never does alone

Advice. Anything that recommends a product, a strategy, a rebalance, or an allocation to a specific client is the advisor's work, and the assistant's role stops at helping the advisor think. Same for anything that reads like a performance claim: "our clients averaged," "this strategy returned," "you would have." A model will produce those sentences fluently and without a basis, and the Marketing Rule is explicit that a material statement needs a reasonable basis you can substantiate on demand. The fix is structural, not a prompt: performance and recommendation language is never generated, only copied from approved material.

Rule one: the Marketing Rule does not care who wrote the draft

The SEC's marketing rule for investment advisers, 17 CFR 275.206(4)-1, lists what an advertisement may not do. Read on eCFR on September 15, 2026, the general prohibitions include any untrue statement of a material fact or an omission that makes a statement misleading, and any material statement of fact the adviser does not have a reasonable basis to believe it can substantiate on demand. Testimonials and endorsements are allowed only with disclosures: that the person is a client or is not, whether they were compensated, and any material conflict. Third-party ratings and hypothetical performance each carry their own conditions.

Nothing in that rule mentions software, which is the point. An AI-drafted blog post, LinkedIn update, or client letter is an advertisement or a communication on exactly the same terms as a human-drafted one. What changes with AI is the volume and the speed, which means the review step is more important, not less. The practical setup for a small firm as of 2026:

• One approval path for anything public. AI-drafted marketing goes through the same person and the same checklist as everything else. If you do not have a checklist, the Marketing Rule's prohibitions are the checklist.
• No generated claims. Numbers, results, and comparisons come only from approved source material that the firm can substantiate. The assistant formats them; it never originates them.
• Testimonials stay manual. The disclosure conditions are specific enough that an assistant should never be the one assembling a testimonial into an ad.

Rule two: books and records wants a copy, for five years

Under 17 CFR 275.204-2, an investment adviser keeps a copy of each advertisement it disseminates, and its books and records are maintained in an easily accessible place for not less than five years from the end of the fiscal year during which the last entry was made, the first two years in an appropriate office. Electronic storage is permitted, with procedures to keep the records complete, true, and legible when retrieved.

For AI this means two things. First, every draft that becomes an advertisement gets filed the same way as before; the chat window is not a filing system. Second, decide now where client correspondence produced with an assistant lives. The firms that get this right export the approved recap into the CRM or the document system the moment it is sent. The firms that get it wrong discover at exam time that the only copy is in someone's chat history.

What FINRA said about generative AI, in its own words

If any part of your business is a FINRA member firm, Regulatory Notice 24-09, published June 27, 2024, is the one to read. FINRA says its rules are intended to be technology neutral and apply when firms use generative AI just as they apply to any other tool. It points to Rule 3110, which requires a reasonably designed supervisory system, and says that if a firm uses Gen AI as part of that system, its policies and procedures should address technology governance, including model risk management, data privacy and integrity, and the reliability and accuracy of the model. It also says the content standards of Rule 2210 apply whether a communication is generated by a human or a technology tool, and that firms should evaluate Gen AI tools before deploying them.

Rule 2210 itself draws the line a small firm needs to know: a "retail communication" is one distributed to more than 25 retail investors in any 30-day period, and an appropriately qualified registered principal must approve each retail communication before it is used. Correspondence, 25 or fewer, sits under Rule 3110's supervision and review. So an AI-drafted market note that goes to 200 clients is a retail communication with principal approval, and an AI-drafted reply to one client is correspondence under supervision. Same as before the model existed.

The plan decides the data terms

This is the decision most firms get wrong first, because it is the one made by whoever signed up. As of September 2026, the business plans from the three major vendors state that customer data is not used to train models by default, and they give an owner an admin console:

• Claude Team: $20 per seat per month billed annually, $25 monthly, for teams of 2 to 150. No training on Team data by default. The plan details are on Claude Team plan.
• ChatGPT Business: Standard seats at $20 per user per month annually, $25 monthly, 2-seat minimum. No training on Business data by default. Self-serve Business does not include a business associate agreement, which matters if you touch health information.
• Microsoft 365 Copilot Business: $21 per user per month on an annual subscription, as an add-on to a Microsoft 365 business plan, for firms that live in Outlook and Teams.

Consumer plans are where the trouble lives. A personal account an advisor signed up for on a weekend has different terms, no admin control, and no offboarding when that advisor leaves. Buy the business plan, assign seats by name, and write the never-paste list into the one-page policy: account numbers, Social Security numbers, full statements, and anything your privacy policy promises to protect. The template I give clients is on the one-page AI acceptable use policy.

What it costs a small firm, as of September 2026

A five-person firm on annual seats is about $100 to $125 a month before setup. The whole bill, seats plus usage plus setup, is laid out in how much does AI cost. For a firm that also runs its own books, the accounting version of this page is AI for accounting firms, and the install is the same shape.

AI for financial advisors: the one-advisor version, with the math

Everything above is written for a firm. If you are one advisor, or two advisors sharing an operations person, most of it shrinks to a single question: is one business seat worth it on meeting prep and recaps alone? Here is that question with a published wage on it instead of a vendor's promise.

O*NET OnLine, the U.S. Department of Labor's occupation database, lists the 2025 median wage for personal financial advisors at $50.51 an hour ($105,070 a year), read October 10, 2026. The workload numbers below are assumptions, labeled as such. Swap in your own.

Two things this table does not say. It does not say the recovered time turns into revenue; that depends on whether you spend it on prospects or on the inbox. And it does not change a single compliance obligation: a solo RIA still files the advertisement, still substantiates the claim and still keeps account numbers out of the chat. Note that the Team plan needs two seats, so a true solo practice either buys two (one for the operations person) or uses a single-user plan, where model training is a setting you have to turn off yourself, and gets that choice approved before any client notes go in. What a solo advisor can skip is the firm machinery: there is no approval chain to design when the approver is you.

Claude for Financial Advisors: Anthropic's plugin, and the audit-log question

In September 2026 Anthropic added Claude for Financial Advisors, a plugin installed inside Claude (not a new model or app). It connects to Addepar, BlackRock, Charles Schwab, Envestnet, iCapital, Orion, SS&C Black Diamond, Vanguard, Wealthbox, Wealth.com and Zocks, and adds workflows for meeting prep and follow-up, portfolio reviews, proposals and a Marketing Rule pre-check. Anthropic's own page says Claude doesn't give investment advice, doesn't place trades or move money, and holds CRM updates and emails until you approve them, which lines up with the four jobs above.

One thing changes the plan section above: Anthropic recommends Enterprise for RIAs because it's the only Claude plan with audit logs and compliance controls. Enterprise has a 20-seat self-serve minimum, so a 6-person firm pays about $4,800 a year in seats before usage, against about $1,440 on Team. If your approved recaps and emails are already filed to the CRM, Team may satisfy your CCO; if they want the tool itself to keep a log, it's Enterprise. The full comparison and the usage math are on Claude for financial services.

Bank-owned trust and wealth departments answer to a bank examiner as well, so the vendor checklist from the 2023 interagency third-party guidance applies to them too; it's walked through stage by stage in AI for banks.

Who can help a wealth management firm roll out AI?

That's the work I do, for the firm side. I'm Justin McKelvey, an AI consultant in Austin, Texas, and I set up Claude, ChatGPT Business and Microsoft 365 Copilot for teams of 10 to 200: the plan chosen for where your advisors already work, the never-paste list written with your CCO, your approved templates loaded so the assistant only ever fills them in, the filing step that sends approved recaps to your CRM instead of leaving them in a chat window, and the advisors and operations staff trained. The AI Readiness Assessment is $2,500 flat for a two-week written plan, credited in full if we build it together within 90 days. A done-for-you rollout starts at $4,500 and takes about two weeks, vendor seats separate; the six steps are on AI implementation. A larger firm with several offices gets a written custom scope after the Assessment. The first 30-minute call is free. What I don't do is the compliance call itself: your CCO signs off, and I build to what they decide.

Want it planned and installed for you? Every price is on one page: AI consulting for wealth management, from the $2,500 Assessment to a rollout from $4,500, with what Reg S-P's service-provider rule means for an AI vendor.

What to do this week

1. Pick the four jobs and name an owner for each. Meeting recap, template drafts, document first-reads, inbox. One person per job, and the rule that a person sends everything.
2. Write the never-paste list with your CCO. One page. What goes in the chat, what never does, and where approved outputs get filed. Give it to the team before anyone gets a seat.
3. Buy business seats for the daily users only. Then run the four jobs for thirty days on your own templates and notes before anyone proposes a platform.

If you would rather have the workflow list, the policy, and the order of operations written for your firm before you buy anything, that is what the AI Readiness Assessment is: $2,500 flat, two weeks, a written roadmap, and the fee becomes your deposit if we build it together. Either way, the first year of AI at a small advisory firm is four boring jobs done well and two rules you already follow. The platform can wait.

Sources

• eCFR, 17 CFR 275.206(4)-1, Investment adviser marketing (read September 15, 2026)
• eCFR, 17 CFR 275.204-2, Books and records to be maintained by investment advisers (read September 15, 2026)
• FINRA Regulatory Notice 24-09 and FINRA Rule 2210, finra.org (read September 15, 2026)
• O*NET OnLine, Personal Financial Advisors (13-2052.00), median wages 2025, onetonline.org (read October 10, 2026)
• Anthropic, Claude pricing, claude.com/pricing (read October 10, 2026)
• Anthropic, Claude for financial advisors, claude.com/solutions/financial-advisors (read October 12, 2026)
• OpenAI ChatGPT Business and Microsoft 365 Copilot Business prices, from each vendor's pricing page (read September 2026)