SEC and FINRA exams: supervision, books and records, Reg S-P, and shadow AI

The SEC and FINRA apply the rules you already follow to AI. FINRA's Regulatory Notice 24-09 says its rules are meant to be technology neutral, and that the notice creates no new requirements.^[1] In June 2025 the SEC dropped its 2023 proposal on conflicts of interest in predictive data analytics, which would have covered AI tools used with investors.^[2]

That puts the weight on four areas examiners already know how to test: supervision, books and records, client data under Reg S-P, and what you tell clients about your AI. Shadow AI, meaning staff using AI tools the firm never approved, touches all four.

What are SEC examiners looking for on AI in 2026?

The SEC's exam division released its fiscal 2026 priorities on November 17, 2025. On AI, examiners will check that what firms say about their AI is true. They will also check whether firms have adequate policies and procedures to monitor or supervise their AI use. That includes AI used for fraud checks, back-office work, anti-money laundering, and trading.^[3]

For automated advice, examiners also check that claims are fair and true, that the firm works the way its disclosures say, and that the advice fits each investor's profile. The cybersecurity section adds training and controls for new risks tied to AI, and it names data loss prevention and access controls as focus areas.^[3] Two advisers have already paid over AI claims. In March 2024, Delphia and Global Predictions settled SEC charges of false and misleading statements about their use of AI, with $400,000 in combined penalties.^[4]

The priorities speak of firms' "use of AI technologies." When staff use a public chatbot for client work, that is the firm using AI, whether or not compliance approved the tool.

How does supervision apply when staff use AI?

FINRA Rule 3110 requires a broker-dealer to supervise its people with a system and written procedures reasonably designed to achieve compliance. Notice 24-09 applies that to generative AI.^[1] For advisers, Rule 206(4)-7 requires written policies and procedures reasonably designed to prevent violations by the firm and its staff. The firm must review them at least once a year.^[5]

FINRA's 2026 oversight report, published December 9, 2025, says firms should work out how the rules apply while they evaluate a generative AI tool, before testing or rollout.^[6] Its suggested practices include logs of prompts and outputs, tracking model versions, and human review of output. For AI agents, it adds limits on what an agent can reach, human checkpoints, and a record of what the agent did.

Advice standards apply to AI-assisted work the same way. Reg BI covers a broker-dealer's recommendations to retail customers, and advisers owe clients a fiduciary duty.^[3] FINRA Rule 2210's content standards apply to communications a tool drafted.^[1] Once an AI draft reaches a client, the firm owns it.

Is an AI chat a business record?

It depends on the content. A broker-dealer must keep messages about its business for at least three years, the first two close at hand. An adviser must keep written messages about advice, orders, money movements, and performance for at least five years, the first two in its office. Registered advisers to hedge funds and other private funds follow the same rule.^[7]

A client email drafted in a chatbot is a record once it's sent. Whether the prompt and the chat around it count is less settled. Rule 17a-4(f) lets a broker-dealer keep electronic records in a format that can't be rewritten or erased. It can also use a system that logs every change and deletion with a time stamp.^[7]

A personal chatbot account keeps those chats on the vendor's servers, outside your archive and out of reach when an examiner asks for them. Where your data goes when staff use AI traces that path in detail.

What does amended Reg S-P require after an incident?

The SEC amended Reg S-P in May 2024 for broker-dealers, funds, registered advisers, and transfer agents. Larger firms had to comply by December 3, 2025, and smaller firms by June 3, 2026. Every covered firm now needs a written plan to spot, handle, and recover from any access to or use of customer data without permission.^[8]

If sensitive customer data was, or likely was, accessed or used without permission, the firm must tell the people affected. It has to do so as soon as it can, and within 30 days of learning of it. The firm can skip notice if it finds the data hasn't been, and likely won't be, used in a way that causes substantial harm or inconvenience. The rule also requires written records that show the firm meets its safeguards and disposal duties.^[8]

The rule includes no duty to report a breach to the SEC. For service providers, it requires due diligence and monitoring. Your procedures must make sure each provider tells you within 72 hours of learning of a breach in a customer data system it runs for you. A provider can send customer notices for you, but making sure they go out stays your job.^[9] The 2026 priorities say Reg S-P exams will focus on policies and procedures, internal controls, vendor oversight, and governance.^[3]

An AI vendor the firm hires that receives or stores prompts, files, or chat logs holding client data fits the rule's definition of a service provider.^[9] When a staff member pastes a client's tax return into a public chatbot, your plan has to say whether that counts as unauthorized use of sensitive customer information. Settle that with counsel before it happens.

Why is shadow AI an exam problem even without a breach?

Supervision and recordkeeping apply whether or not anything leaks. If staff use AI your procedures never mention, you can't show you supervise that use or keep what it produces. In a 2025 University of Melbourne study with KPMG of over 48,000 people in 47 countries, 57% of employees said they hide their use of AI and present AI-generated work as their own.^[10]

In IBM's 2025 Cost of a Data Breach study, run by the Ponemon Institute, one in five organizations it studied reported a breach tied to shadow AI.^[11] For an adviser, the files staff might ask a chatbot to summarize include estate plans, tax returns, trust documents, and account statements. All of it is customer information under Reg S-P.

How do you find shadow AI and close the gap?

Find out what people use, write the rules down, and give staff an approved tool that does the same job. People turn to chatbots to get work done, so any limit needs a tool behind it that can do that work. The steps below cover the first pass.

1. Look for it. Check web proxy or DNS logs for AI sites and data loss prevention alerts for client data headed to them. Then check expense reports for AI subscriptions, browser extensions, and meeting platforms for AI note-takers on client calls. Last, ask staff directly, with no penalty for honest answers.
2. Write it into the procedures. Name the approved tools, the data allowed in them, who reviews output before it reaches a client, and how compliance checks that people follow the rules. Test all of it at the annual review.
3. Decide what counts as a record. Say which AI interactions you keep and where, and make sure your archive can produce them on request.
4. Add AI vendors to your Reg S-P program. Run due diligence, put the 72-hour notice in the contract, and get a written answer on whether the vendor keeps or trains on your data.
5. Check your AI claims. Read every place you describe AI to clients or investors, including answers on due diligence questionnaires, and confirm each claim is true today.

An approved tool should keep client data in your environment, follow your existing permissions, and log every interaction where compliance can search it. Lumen, the private AI platform from Cognetryx, runs on your own servers. Answers follow each person's existing permissions through SSO over SAML or OIDC, so an adviser sees only the client files they can already open. Each answer cites its source. The Compliance Portal logs chats, tool calls, agent runs, approvals, meetings, consent events, deletion requests, exports, permission and config changes, and sign-ins, and your compliance team can search it. Ekho, the meeting transcription tool, runs on your own GPU and tracks each participant's consent.

Keith Kennedy, CISSP

Founder, Cognetryx

Keith is an IT thought leader with nearly 20 years of experience architecting secure technology solutions for regulated industries. He holds a CISSP certification and has advised enterprise companies on SEC/FINRA, HIPAA, and GDPR compliance.