AI Governance for Investment Firms: What to Document for an SEC Exam

As of November 2026, the SEC has no rule written specifically for how investment advisers use AI, so examiners apply the rules already on the books: the antifraud provisions and the Marketing Rule to what a firm says about AI, the Compliance Rule to how it supervises AI use, Regulation S-P to the customer data AI tools touch, and the books-and-records rules to the work AI helps produce. Kepler, the enterprise intelligence platform we build for investment firms, records much of the evidence an exam file needs as it runs: every query and agent run under its owner’s name, every number traced to its source, and every read checked against the firm’s permissions. A firm ready for an exam can hand over an inventory of its AI tools, a written use policy, vendor diligence, testing records, audit trails, evidence of supervision, retained records, and the substantiation behind every AI claim in its marketing.

This guide summarizes public SEC and FINRA materials as of November 2026. It isn’t legal advice; work through your firm’s obligations with counsel and your chief compliance officer. The last section shows how Kepler supports each part of the file.

The regulatory picture, November 2026

A note on terms: the SEC now says “SI”

Executive Order 14434, “Inaugurating the Era of Super Intelligence,” signed September 29, 2026, directs executive agencies to use “Super Intelligence” and “SI” in place of “artificial intelligence” and “AI” in non-statutory documents. For now, the term keeps the statutory definition of AI in 15 U.S.C. 9401(3) (the order asks for proposed legislation on a new federal definition within 60 days), and previously issued regulations don’t have to change. The SEC’s FY2027 examination priorities use “super intelligence (SI)” throughout, while older releases, rules and enforcement orders say AI. When you search SEC materials, search for both.

No AI-specific rule: the 2023 proposal was withdrawn

In July 2023 the SEC proposed a rule on conflicts of interest from predictive data analytics (Release No. IA-6353) that would have required broker-dealers and advisers to evaluate the technologies they use in investor interactions and eliminate or neutralize certain conflicts. On June 12, 2025, the Commission withdrew it along with 13 other pending proposals, noting it would issue a new proposed rule if it decides to act in any of those areas.

Chairman Paul Atkins has described the approach since then. In March 2026 remarks he said, “The Commission’s mandate to protect investors is technology neutral. And misconduct remains misconduct, regardless of the medium.” In May 2026 he added that firms “remain responsible for the outcomes of the tools that they deploy” and that the SEC “will not dictate which models firms must use.”

“AI washing” enforcement

The SEC has brought AI cases under existing antifraud rules since 2024:

• Delphia and Global Predictions (March 18, 2024). The first AI-washing settlements against advisers: $225,000 and $175,000 in civil penalties for false and misleading statements about their use of AI, under the Advisers Act and the Marketing Rule. Then-Chair Gary Gensler said advisers “should not mislead the public by saying they are using an AI model when they are not.”
• Rimar Capital (October 10, 2024). The SEC charged Rimar entities, their owner and a board member over false claims of an AI-driven trading platform used to raise nearly $4 million from 45 investors; $310,000 in civil penalties.
• Presto Automation (January 2025). The SEC settled charges with a restaurant-technology company, formerly listed on Nasdaq, over misleading statements about its AI drive-thru product.
• Albert Saniger (April 2025). The SEC charged the founder of a shopping app with misleading investors, who put in over $42 million, about its AI.
• Fake AI trading tools (September 29, 2026). The SEC charged entities that touted supposed AI-generated trading signals and rented out AI trading bots that didn’t exist.

In February 2025 the SEC created a Cyber and Emerging Technologies Unit whose first listed priority is fraud committed using emerging technologies such as AI and machine learning.

The Marketing Rule

Rule 206(4)-1, adopted in December 2020 with a compliance date of November 4, 2022, prohibits advertisements that include any untrue statement of a material fact, or a material statement of fact the adviser doesn’t have a reasonable basis for believing it can substantiate on demand by the Commission. A line like “our research is AI-driven” reads as a statement of fact, so expect examiners to ask what backs it up. Delphia and Global Predictions were both charged under this rule.

The rule defines an advertisement as a communication to more than one person (or to anyone, if it includes hypothetical performance), and it excludes information in required regulatory filings. A DDQ answered for a single prospect, or the text of Form ADV, therefore generally falls outside the Marketing Rule, though the antifraud provisions still apply to both.

Regulation S-P amendments

The amendments adopted May 16, 2024 require an incident response program; notice to affected individuals as soon as reasonably practicable and no later than 30 days after a firm becomes aware that sensitive customer information was, or is reasonably likely to have been, accessed or used without authorization; and oversight of service providers, with policies reasonably designed to ensure they notify the firm no later than 72 hours after becoming aware of a breach. Larger entities, including advisers with $1.5 billion or more in assets under management, had to comply by December 3, 2025; smaller entities by June 3, 2026. An AI vendor that receives, processes or can access customer information through services it provides directly to the firm likely meets the rule’s definition of a service provider.

Books and records

Rule 204-2 requires advisers to keep required records, including written communications relating to advice, for at least five years, the first two in an appropriate office. Broker-dealers follow Rule 17a-4. Neither rule has an exception for records drafted with AI tools.

For broker-dealers, FINRA’s 2026 oversight report lists “storing prompt and output logs” and “tracking which model version was used and when” among monitoring practices firms may want to consider for generative AI.

What the Division of Examinations says

The FY2026 priorities (November 2025) said the Division “will review for accuracy registrant representations regarding their AI capabilities or AI” and assess whether firms have “adequate policies and procedures to monitor and/or supervise their use of AI technologies.”

The FY2027 priorities, released October 23, 2026, note that registrants use generative SI tools “to support research, drafting, coding, and client communications,” and that the Division is working to understand “whether firms continue to comply with existing, technology-neutral obligations.” Examiners will prioritize “policies and procedures relating to monitoring SI”; for automated investment tools and algorithms, they’ll check whether “representations are fair and accurate” and whether operations and controls match disclosures. Regulation S-P exams will cover response programs, customer notification and service provider oversight, and cybersecurity exams will look at SI-generated phishing and deepfakes.

Rule	What it requires	Where AI comes in
Antifraud (Advisers Act 206)	No false or misleading statements	Claims about AI capabilities
Marketing Rule 206(4)-1	Material statements of fact in ads need a reasonable basis to substantiate on SEC demand	AI claims in decks and websites (DDQ answers and Form ADV still fall under antifraud rules)
Compliance Rule 206(4)-7	Written policies, reviewed annually	AI use policy and supervision
Regulation S-P	Incident response, 30-day notice, vendor oversight	AI vendors handling customer data
Rule 204-2 / Rule 17a-4	Keep required records	AI-drafted communications and other required records

What to document: the AI exam file

Document	What it should show	Related rule
AI inventory	Every AI tool, its owner, its purpose and the data it reaches	Compliance Rule; Reg S-P
Use policy	Approved and prohibited uses, review steps, data rules	Compliance Rule
Vendor diligence	Security reports, data use and training terms, breach notice	Reg S-P
Testing records	How outputs were checked, what failed, what changed	Antifraud; Marketing Rule
Audit trail	Who asked, what the tool read and produced, who approved	Books and records; supervision
Marketing file	Evidence behind every statement about AI	Marketing Rule

1. An inventory of AI use

List every AI tool in the firm: research assistants, chatbots, meeting note-takers, coding tools, vendor products with AI features, anything built in house. For each, record the owner, the business purpose, the data it can reach (client information, material nonpublic information, investor data), whether its output reaches clients or investors, and when it was approved. Include tools people use on personal accounts as well.

2. A written AI use policy

The Compliance Rule requires written policies reasonably designed to prevent violations, reviewed at least annually. Many firms put the AI policy inside that program: which tools are approved for which work, what data never goes into them, when a person must review output before it’s used, how AI-assisted work is recorded, and who approves new tools.

3. Vendor due diligence

For each vendor, including us, keep its SOC 2 Type II report, contract terms on how your data is used and whether it trains models, where data is processed, its sub-processors, and the breach notice it has agreed to give you (Regulation S-P expects your policies to secure notice within 72 hours). Our guide to AI data security for investment firms covers the controls to test, from permissions and information walls to prompt injection, and our 25 questions for evaluating AI vendors give the diligence file a checklist.

4. Testing

Record how each tool was tested on your own work before approval and how testing continues. For research tools, check figures against sources, ask the same question twice, test your firm’s definitions and check whether lists are complete; our checklist for AI-generated numbers is a starting point. Keep the failures in the file too.

5. Audit trails and lineage

For AI work behind an investment decision, a client communication or a filing, you should be able to show who asked, what the tool read, what it produced, which source backs each figure, and who approved the result. A log of prompts and replies alone is thin evidence for a number in an IC memo; Kepler keeps the full lineage, as the next section describes.

6. Supervision

Name who reviews AI output before it reaches clients, investors or the market, and keep evidence the review happened. Research teams with agents running overnight need the same: see how hedge funds use AI for research.

7. Retained records

With counsel, decide which prompts and outputs are records under Rule 204-2, where they’re stored and for how long, and make sure AI-drafted communications land in the same archive as everything else.

8. Marketing review

Route every statement about AI that leaves the firm (website copy, pitch decks, DDQ answers, Form ADV) through the review you apply to performance claims, and keep what supports each one.

How Kepler supports an AI exam file

We built Kepler for work a firm has to defend, so most of the evidence above is something the platform records as it runs.

A record of every run. Each question, calculation and agent task on Kepler is logged as a run under the name of the person who started it. Agents on schedules and events carry their owner’s access and nothing more, so an overnight agent reading every new 10-Q for the book leaves the same trail as a question typed at noon.

Lineage on every number. When a reviewer asks where a figure in an IC memo came from, the answer is one click away: the page, line or cell behind it, and for a calculated figure, the formula and each input. Our deterministic engine takes each figure straight from its source and does every calculation in code, so the model never writes a number, and each citation is checked against its source before anyone sees it. The same question returns the same number, so a test recorded in March still reproduces at exam time.

Complete answers for the marketing file. Extraction models we train read every connected document in full as it arrives. A request for every deck, DDQ answer and letter that describes the firm’s use of AI returns the complete set, each with its page, ready to set beside the support behind each statement.

Walls that hold for agents too. Kepler mirrors the firm’s permissions, walls between deals included, in a single check that runs before every read, whether a person or an agent is reading. Each query, calculation and data access lands in the log with its lineage.

Documents can’t give orders. Filings, emails and web pages reach Kepler agents strictly as data, so a line planted in a counterparty’s email or a scraped web page has no way to redirect an agent.

Every edit is a version, and people decide. Each Kepler answer is stamped with the version of the data and definitions behind it, so a figure in last quarter’s investor letter can be reproduced as it stood then. Your definitions live in your firm’s ontology, so adjusted EBITDA is computed the same way in every memo. Nothing is shared outside the firm until a person approves it.

Model upgrades that don’t move the numbers. We assign each step to a model, extraction models we train for reading documents and frontier models for interpreting the question and drafting the answer, and replace them as better ones are released. Because code computes every figure, the same question on the same version of the data returns the same number before and after an upgrade.

Vendor diligence, ready. Your vendor file gets a SOC 2 Type II report, GDPR and CCPA compliance, isolation of your data at the identity layer, a commitment that models other firms use are never trained on your data, and deployment in our cloud or yours, all documented on our security page.

Our engineers, who built this layer before at Palantir, work on site and connect your sources read-only in the first week. The same controls follow the work into Excel, PowerPoint, Claude, ChatGPT and our API. For the broader architecture, see what enterprise intelligence is and AI tools for asset managers.

Questions compliance teams ask Kepler

• “Show every run behind this IC memo: who started it, what it read, and which source backs each figure.”
• “List every agent that ran overnight last week, its owner and what it produced.”
• “Which sources behind this investor letter have changed since it was sent?”

See Kepler at app.kepler.ai, or book a demo and bring your CCO’s questions.

Frequently asked questions

Does the SEC have a rule on AI for investment advisers?

Not as of November 2026; examiners apply existing, technology-neutral rules instead. Kepler records every query and agent run with its owner and the source behind each figure, so the evidence those rules call for is ready when examiners ask. The SEC withdrew its 2023 proposal on predictive data analytics conflicts in June 2025. The rules in play are the antifraud provisions and the Marketing Rule for what a firm says about AI, the Compliance Rule for how it supervises AI use, Regulation S-P for customer data, and the books-and-records rules.

What is AI washing?

AI washing means overstating or misrepresenting how a firm uses AI. In March 2024 the SEC settled the first such cases against two advisers, Delphia and Global Predictions, for $400,000 in combined penalties under the Advisers Act and the Marketing Rule. Later cases involved a startup trading firm, a restaurant-technology company once listed on Nasdaq, and a startup founder.

What do SEC examiners ask about AI?

The FY2026 priorities said examiners would review the accuracy of firms' statements about their AI capabilities and whether firms have policies to monitor and supervise AI use. The FY2027 priorities, released October 23, 2026, call AI super intelligence (SI) and focus on policies for monitoring SI, fair and accurate representations about automated tools, and SI-driven cyber threats.

Do AI prompts and outputs count as books and records?

Rule 204-2 doesn't distinguish records by how they were drafted, so an AI output that is a required record, such as a written communication relating to advice, is generally kept like any other: for at least five years, the first two in an appropriate office. FINRA's 2026 oversight report suggests firms consider storing prompt and output logs and tracking which model version was used. Platforms such as Kepler record every run with its sources and owner.

When did the Regulation S-P amendments take effect?

The amendments adopted May 16, 2024 apply to larger entities, including advisers with $1.5 billion or more in assets under management, from December 3, 2025, and to smaller entities from June 3, 2026. They require an incident response program, customer notice within 30 days, and service provider oversight, including policies designed to secure breach notice from providers within 72 hours.

Why does the SEC now say SI instead of AI?

Executive Order 14434, signed September 29, 2026, directs executive agencies to use Super Intelligence and SI in place of artificial intelligence and AI in non-statutory documents. For now, the term keeps the statutory meaning of AI; the order also asks for proposed legislation on a new federal definition. The SEC's FY2027 examination priorities use super intelligence (SI), so search SEC materials for both terms.