How to Use AI as a Financial Advisor: What Works, and the Rules That Do Not Bend

A recommendation to a client has your name on it, not the model's. That single fact should decide where AI belongs in a financial advisor's day and where it does not — and it holds regardless of how good the drafting gets, because the obligation was never about drafting in the first place.

Start with the obligations, because they narrow the options

If you work through a FINRA-member broker-dealer, FINRA Regulatory Notice 24-09 is explicit that existing rules — including Rule 2210 on communications with the public — apply to generative AI output exactly as they apply to anything a person wrote by hand. A client email, a market commentary, a social post: if a human-written version would need review under your firm's supervisory procedures, an AI-drafted version needs the same review, not a lighter one because a tool produced the first pass.

Suitability and fiduciary duty work the same way. Whether a recommendation is actually appropriate for a specific client's goals, time horizon and risk tolerance is a judgement about that client, not a generic answer a model can supply from a prompt with no context attached. If you hold a CFA charter, Standard V(A), Diligence and Reasonable Basis requires a reasonable and adequate basis for a recommendation — a standard that predates generative AI and applies to it without modification. An AI-drafted rationale is a source you investigated, not a substitute for having investigated.

Confidentiality narrows things further. Client account details, holdings and financial circumstances are exactly the kind of information Standard III(E), Preservation of Confidentiality, covers regardless of the channel — a general consumer AI tool is not a confidential channel unless your firm has specifically reviewed and approved it as one. Other client-facing, regulated professions have landed on the same line independently: how to use AI as an accountant and how to use AI as a financial analyst both draw it in almost the same place, for the same reason.

What it is genuinely good at

1. 01Meeting prep. Pulling together a client's account history, recent contact notes and any open action items into a short pre-meeting brief, so you walk in already knowing where you left off.
2. 02Summarising a meeting into next steps. You already had the conversation and made the judgement calls; ask for a short, structured recap and action list rather than writing it from a blank page afterward.
3. 03Explaining a concept in plain language. Turning "sequence-of-returns risk" or "backdoor Roth conversion" into a version a specific client can actually follow, at the reading level and length you specify — the model is good at rephrasing something you already understand correctly, not at deciding what the client should do about it.
4. 04First-draft plan structure. The shape of a financial plan document is fairly generic even when the numbers and recommendations inside it are not. Let the model draft the skeleton and headings; put your analysis and the client's actual numbers into it yourself.
5. 05Drafting routine client communications. A market-update newsletter or a quarterly check-in email, reviewed and approved the same way any other client communication is under your firm's compliance process.

A worked example

Weak: "Explain to my client why their portfolio is down this quarter." With no actual portfolio data attached, this produces fluent, generic market commentary that may not reflect what actually happened in that client's specific account — a plausible-sounding explanation is not the same as a correct one, and a client reads confidence as accuracy.

Better: give the model the client's actual asset allocation and the quarter's return by asset class, and ask for a short, plain-language draft that attributes the change to the specific holdings that moved, in the tone you would use with that client. You still read it against the numbers before it goes out — but now there is something specific to check it against, rather than a paragraph of confident generalities with nothing underneath.

A generic explanation and a correct one can read identically confident. The only way to tell them apart is to check the specific numbers underneath.

Where it will quietly get you into trouble

• —Market data with a stale cutoff. A model's training data has a date. A current rate environment, a recent fund closure, or "who currently manages this fund" can come back wrong in a fluent, confident way if the honest answer needed a live lookup the model did not do.
• —Return projections and performance claims. Asking a model to project future returns or characterise past performance is asking it to generate exactly the kind of statement securities regulations are strictest about — this is not a shortcut, it is a compliance risk with your name attached.
• —Generic advice standing in for a suitability analysis. "What should a 45-year-old with $200k save for retirement" has no client in it. Suitability requires that specific client's actual numbers, goals and constraints — a template answer dressed as guidance is the opposite of what the obligation requires.
• —Asking the model to check its own draft. Models trained on human approval tend to produce approval-shaped answers when asked to grade their own output — Anthropic has measured and named this tendency sycophancy. Review a draft yourself, against the actual numbers and your own judgement, rather than asking the same tool if it looks right.
• —Client data in an unvetted tool. Account numbers, holdings and personal financial details belong only in a tool your firm has specifically reviewed for that use — not in a general consumer AI assistant, however good the drafting is.

The connecting thread is the same one covered generally in what AI is actually bad at: these systems fail confidently rather than visibly, which is a harder failure to catch than an error message. How to check an AI answer when you are not the expert covers the habits that catch it anyway.

Getting started without a project plan

1. 01Confirm which tool, if any, your firm has actually reviewed for client information, and use nothing else for anything beyond public, non-client material. If nothing has been reviewed, raise that first — it is the actual blocker, not a detail.
2. 02Pick one recurring task you can judge instantly — a meeting-notes summary or a plan-document skeleton, not a recommendation. Find the repetitive part is a short test for choosing it if the answer is not obvious.
3. 03Do that task both ways for a couple of weeks — your usual way and with an AI-drafted first pass — and keep the time. You want evidence it actually helps, not a feeling that it does.
4. 04Build a small set of prompts that reliably produced something checkable rather than confident filler, and reuse them rather than starting from scratch each time.

The part that actually protects your job

The tasks most exposed here are the ones furthest from judgement about a specific client — drafting a summary, structuring a document, rephrasing a concept. What is left is the part clients are actually paying for: a suitability judgement about their specific situation, and being the accountable person who stands behind it when markets move against a plan. That part does not automate the same way, because a client is not paying for a paragraph — they are paying for someone accountable to have checked it. The same conclusion holds from the analyst side in how to use AI as a financial analyst and from the hiring side in AI-proof careers.

The short version

Use AI for meeting prep, drafting summaries, and turning a plan you already built into client-ready language. Do not use it for return projections, generic advice standing in for a real suitability analysis, or anything touching client data in a tool your firm has not reviewed. Confirm what is actually approved before you put anything client-specific near a general tool, and treat a confident-sounding draft as a first pass to check, not a finished answer to send.

Coursium teaches exactly this layer — practical use of the tools plus the checking habit that has to go with it. Stay ahead of AI by learning them on your phone.