RIA AI Compliance: The Questions an Examiner Will Actually Ask

For the first few years of generative AI, examination staff mostly asked about it when a firm advertised it. That has changed. AI questions now show up in routine examinations of registered advisers that have never marketed a model in their lives, because exam staff know the tools are already inside the firm whether the compliance manual mentions them or not.

We build AI governance for advisers with examination as the bar, so we spend a lot of time on what a defensible answer actually looks like. Here are the questions we prepare firms to answer, and where the answers usually fall apart.

Why AI is on the exam agenda at all

The SEC's Division of Examinations has named AI among its examination priorities, including how advisers use it and what they tell clients about it. The Commission has also brought enforcement actions against advisers over AI claims they could not substantiate, the pattern staff call AI washing.

Notice what did not happen: there is no comprehensive new AI rule for advisers. Examiners do not need one. Rule 206(4)-7 already requires policies and procedures reasonably designed to prevent violations of the Advisers Act. The Marketing Rule already governs claims about how you invest. Rule 204-2 already covers your books and records. Your fiduciary duty already covers the advice, however it was produced. AI questions are simply those existing obligations pointed at a new set of tools.

Question 1: What AI tools are in use at your firm?

This is the opening question, and it is quietly the hardest one. Examiners are not only asking about the tools you bought. They are asking whether you know what is actually in use, including the AI features embedded in software you already run: the assistant in Microsoft 365, the notetaker in your meeting platform, the AI summaries in your CRM, and whatever staff signed up for on personal accounts.

A defensible answer is a current, written inventory: each tool, what it is used for, what data it can touch, whether it is approved, and who owns the decision. A weak answer is a list of two sanctioned tools at a firm where staff are quietly using ten. The gap between those lists is shadow AI, and it is the single most common finding when we assess a firm for the first time.

Question 2: Show us your AI policies and procedures

Examiners will ask for the policy, when it was adopted, and how it connects to your compliance program under Rule 206(4)-7. Then comes the question that separates real governance from paper: how do employees know what it says?

A defensible answer is a policy that names the approved tools, classifies data by what each tool is authorized to touch, sets review requirements for AI-assisted work, and comes with evidence of training and employee acknowledgment. A weak answer is a template downloaded the month the exam letter arrived, with no training records behind it. A policy nobody is trained on can be worse than no policy, because it documents a standard the firm is not meeting.

Question 3: What do you tell clients and prospects about AI?

This is the Marketing Rule question, and it cuts in both directions. If your deck, website, or RFP responses describe AI-driven analysis, examiners will ask you to substantiate it. If the reality is an analyst with a spreadsheet and an occasional chatbot session, that claim is a problem, and it is exactly the fact pattern behind the AI-washing enforcement actions. The opposite direction matters too. If AI meaningfully drives your process and your disclosures say nothing, expect questions about whether clients understand how their money is managed.

The defensible position is boring on purpose. Say what you actually do, keep a substantiation file behind every AI claim in your materials, and have compliance review AI language the same way it reviews performance claims.

Question 4: How did you vet these vendors?

For every AI tool on the inventory, examiners want to see due diligence that happened before adoption, not after. The questions that matter: where does the data go, is it retained, is it used to train models, who are the subprocessors, and what security posture stands behind the answers. For an adviser, whether your inputs are used for training is not a technical footnote. It determines whether client non-public information is leaving your control.

A defensible answer is a short, written diligence memo per tool with the vendor terms attached. It does not need to be long. It needs to exist, be dated before the rollout, and match the plan the tool is actually on, because an enterprise agreement and a consumer account from the same vendor can carry completely different data terms.

Question 5: What about your books and records?

Records you are required to keep under Rule 204-2 do not stop being records because AI drafted them. Client communications written with an assistant still get retained and reviewed. And where an AI output becomes the basis for a recommendation, expect questions about how that work is documented. The point is not that a new category of record was invented. It is that AI-assisted work has to flow into the retention and review you already do, and firms that adopted tools without thinking this through usually find gaps.

Question 6: Who supervises AI-assisted work?

Someone has to be accountable for work product, and it cannot be the tool. Examiners want to see that AI-assisted output gets human review appropriate to its risk, that responsibility is assigned by role, and that the supervision is actually happening, not just described. For dual registrants, the same logic runs through FINRA Rule 3110 on the brokerage side: the supervisory system has to cover how the work actually gets done now.

Where firms actually fail

Three patterns cover most of the failures we see. Shadow AI, where the real tool inventory is twice the official one and includes personal accounts nobody vetted. The paper policy, adopted but never trained, so the first employee an examiner interviews contradicts it. And non-public information in public tools, where a well-meaning employee pastes client data or material non-public information into a consumer chatbot to save an hour. Each one is cheap to fix before an exam and expensive to explain during one.

How to get ahead of it

The sequence matters less than starting. Build the inventory first, because everything else keys off it. Then right-size the policy to how your firm actually works, classify data by what each tool is authorized to touch, and train people until the certification is honest. If you want a structured starting point, an AI readiness assessment surfaces the gaps in weeks, and our AI governance work is built around the examination questions above. We do this for financial services firms specifically because the bar there is not internal comfort. It is examination.

Key takeaways

• AI is on the SEC examination agenda under existing rules. No new AI rule was needed.
• The first question is the tool inventory, and shadow AI is the most common gap.
• A policy only counts with training and employee acknowledgment behind it.
• Marketing claims about AI need substantiation, in both directions.
• Vendor diligence, records, and supervision must cover AI-assisted work the way they cover everything else.

Talk it through

Preparing for an exam with AI on the agenda? Start with a 30-minute call.