AI Deployment in SEC-Registered Firms: Five Pillars and a Control Stack

Who This Is For

This article is written for the managing partner, the CEO, the CCO, and the COO at an SEC-registered investment adviser: the executives who do not need to become AI experts, but who do need to make defensible decisions about AI inside their firm.

You do not need to know how a large language model works. You do not need to evaluate prompt engineering or fine-tuning. You do not need to pick a vendor.

What you do need to understand, at the executive level, is three things:

  • Why your staff want to use AI, and why they are already using it whether you have approved it or not.
  • How AI actually moves through your firm: which tools, which data, which people, which channels.
  • What has to be in place, as policy, as documentation, as technology, as evidence, so that AI use inside your firm is governed, not improvised, and so that you can demonstrate that governance to an SEC examiner, a board, or a sophisticated client.

That is what the MTradecraft AI framework is for. It is a lens for executives to look at AI the same way they already look at any other operational risk inside the firm: define what is permitted, build the controls, produce the evidence, and put a name on the program.

The rest of this article walks through that framework in two parts: five compliance pillars that define the program, and a control stack that enforces it technically. By the end, you should be able to ask your IT provider, your CCO, and your staff the right questions, evaluate the answers, and know what a complete program looks like even if you are not the person who builds it.

Why This Matters Now

Artificial intelligence has moved from a curiosity to an operational reality inside SEC-registered investment advisers. Portfolio managers are summarizing research with consumer chatbots. Compliance officers are drafting policies with AI assistants. Marketing teams are generating client communications with productivity copilots. Operations staff are pasting client account data into whatever free tool a colleague recommended over lunch.

Most of this is happening without a written policy, without a documented vendor review, without monitoring, and often without the executive team’s knowledge. That is the gap this framework is designed to close.

The SEC has not issued a dedicated AI governance rule for advisers. In June 2025, the Commission formally withdrew its proposed predictive data analytics rule, and any future rulemaking on that subject will require a new proposal and a new comment period. That does not mean AI is unregulated at your firm.

Existing obligations under Rule 206(4)-7, Rule 204-2, Regulation S-P, the Marketing Rule, and the adviser’s fiduciary duty already reach virtually every practical AI use case. The Division of Examinations made the point directly in both its 2025 and 2026 examination priorities, naming AI as an emerging focus area for investment advisers, including the accuracy of AI-related representations, the adequacy of policies and procedures supervising AI use, and the protection of client information when third-party AI tools are involved.

The question is not whether AI is regulated at your firm. The question is whether you can demonstrate, with evidence, that your AI use is governed.

What follows is the framework MTradecraft uses when assessing AI deployment at RIAs. It has two parts: five compliance pillars that define the program, and a control stack that enforces it technically. MTradecraft is vendor-independent and does not resell technology; the framework describes capabilities, not products.

The Five Pillars

1. Governance

Governance is the foundation. Without it, every other control is improvisation.

Three questions establish whether a firm has AI governance:

  • Is there a written AI policy approved at the appropriate level of the firm?
  • Are permitted and prohibited use cases defined in concrete terms (not “use AI responsibly”)?
  • Has ownership been assigned to a named individual accountable for the program?

A policy that lives only in the CCO’s head is not a policy. A policy that says “employees should be careful with AI tools” is not a policy either. The artifact an examiner expects to see is a document that names approved tools, defines prohibited inputs (client PII, non-public portfolio data, MNPI), identifies approved use cases, and assigns an owner.

Ownership matters. Most firms assign the CCO by default. That is acceptable, but the CCO needs technical input from whoever administers the firm’s productivity environment or the firm’s IT provider. AI governance that ignores the technical layer becomes paperwork.

The AI Inventory (aka AIBOM). A governance program needs an artifact that an examiner can ask for and the firm can produce in one document. MTradecraft calls this the AI Inventory, sometimes referred to elsewhere as an AI Bill of Materials. At minimum it identifies, for every approved AI tool:

  • The product name and vendor
  • The business owner and approved users
  • Permitted use cases and prohibited data types
  • Data classification limits (public, internal, confidential, restricted)
  • Vendor due diligence status and date of last review
  • Logging source (where activity records live)
  • Date of last governance review

If a tool is in use at the firm but not on the inventory, that is the finding. If a tool is on the inventory but no one can produce evidence of vendor due diligence, that is also the finding. The inventory forces the firm to make AI use visible to itself.

2. Data Protection

This is where most firms fail. The risk is not theoretical: it is the moment an employee pastes a client’s portfolio statement into a free AI tool to “summarize it for the meeting.”

Two questions:

  • Is client data, MNPI, or other sensitive information being entered into AI prompts?
  • Are redaction, obfuscation, or technical controls in place to prevent it?

The first question requires honest assessment. Most firms assume the answer is no. Most firms are wrong. Employees use whatever tool reduces friction, and free consumer AI services are the least friction-laden option available.

The second question has a range of answers from “we trust our employees” (not a control) to “we route AI traffic through an enterprise environment with prompt inspection and content filtering” (a control). The middle ground includes training, acceptable use policies that explicitly list prohibited data types, and technical enforcement through the control stack described below.

For firms subject to the 2024 amendments to Regulation S-P, the data protection question carries a second weight. If sensitive customer information is entered into or exposed through an unauthorized third-party AI service, the firm should treat that as a potential Regulation S-P incident and document its assessment. If unauthorized access to or use of sensitive customer information occurred or is reasonably likely to have occurred, and the firm cannot conclude after reasonable investigation that the information is not reasonably likely to be used in a manner that would result in substantial harm or inconvenience, customer notification may be required as soon as practicable and no later than 30 days. The amended rule now applies to every covered firm: larger entities had to comply by December 3, 2025, and smaller entities by June 3, 2026.

3. Deployment Model

Not all AI is the same. The deployment model determines the risk profile.

  • Consumer SaaS: free or personal-tier AI services. Consumer-tier offerings generally lack the contractual protections, administrative controls, audit logs, retention commitments, and enterprise data-use restrictions an adviser would need for any client-related work.
  • Enterprise SaaS: business or enterprise versions of major AI services. Contractual data protection, commitments that customer data is not used for model training, administrative controls, audit logging. Appropriate for most RIA use cases with proper configuration.
  • Private or managed cloud: AI services that operate within the firm’s cloud control plane or a contracted service boundary rather than through a public consumer interface. Stronger control over identity, logging, network access, and data handling, but more technical governance is required.
  • Local deployment: on-premise or air-gapped models running on infrastructure the firm controls. Maximum control. Appropriate for firms with specific data residency requirements or that handle particularly sensitive information.

Model selection matters within each category. A firm using a productivity copilot is not using “AI generally”; it is using a specific product with a specific identity context, permission model, service boundary, logging regime, and data flow into a specific tenant. Each of those needs to be documented. “We use a copilot” is not an answer an examiner will accept; an answer that names the product, identifies how identity is enforced, describes what DLP controls apply, and points to where activity is logged is.

4. Vendor Due Diligence

This is the simplest pillar to articulate and the one most often overlooked:

An AI vendor is a service provider.

That means the same vendor due diligence framework that applies to your custodian, your portfolio accounting system, and your email provider applies to your AI vendor. SOC 2 reports. Data processing agreements. Security questionnaires. Confidentiality and intellectual property terms. Subprocessor disclosure. Incident notification commitments.

The 2024 amendments to Regulation S-P make this explicit. If a service provider maintains customer information or a customer information system and a breach occurs there, the covered institution remains responsible for its incident response and customer notification obligations. AI vendors that ingest client data are no different from any other vendor in this respect.

A short checklist for AI vendor DDQ:

  • SOC 2 Type II report on file
  • Data processing agreement signed
  • Subprocessor list reviewed (most AI vendors use multiple subprocessors)
  • Confirmation that customer data is not used for model training
  • Defined data retention and deletion procedures
  • Contractual incident notification commitment no later than 72 hours after the provider becomes aware of a breach involving unauthorized access to a customer information system maintained by the provider

That 72-hour requirement is not optional language. The amended Reg S-P requires covered institutions to have written policies and procedures that require their service providers to notify them as soon as possible, but no later than 72 hours, after becoming aware of such a breach. Existing vendor contracts may need to be amended to meet this standard.

5. Monitoring

If governance is the foundation, monitoring is the proof. Examiners ask for evidence. Monitoring produces it.

Two dimensions:

  • Prompts and usage: who is using which AI tools, how often, and with what data? At minimum, the firm should know which AI services have been accessed from corporate devices and corporate accounts. Visibility into prompt content is the higher-maturity answer.
  • Staff risk: which employees handle the most sensitive data, and which employees are the most active AI users? The overlap is where attention belongs.

Monitoring is not surveillance for its own sake. It is the mechanism that lets the CCO answer the examiner’s question: “How do you know your policy is being followed?”

Books and Records: A Pillar That Cuts Across All Five

Rule 204-2 was written long before generative AI existed, but it applies. For any AI use that contributes to advice, client communications, marketing materials, investment commentary, meeting summaries, or compliance decisions, the firm should determine whether the prompt, output, approval record, and final artifact are required books and records. The control design should identify where those records live, how long they are retained, and how they can be produced during an examination.

The practical implication: if an analyst uses an AI assistant to summarize a research call and that summary feeds into an investment recommendation, the firm needs to know whether the AI interaction is itself a record. If a marketing team member uses an AI tool to draft a quarterly letter, the prompt and output may need to be retained. This is a determination the firm should make deliberately, document, and apply consistently, not a question to leave open until an examiner asks it.

Marketing Rule and AI-Washing

AI also creates a disclosure risk. If the firm describes its use of AI in marketing materials, investor communications, pitch decks, Form ADV language, or website copy, those statements must be accurate, supportable, and not overstated. The SEC has already brought enforcement actions against investment advisers for so-called “AI washing”, overstating the role of AI in investment processes or capabilities, and the Division of Examinations has named the accuracy of AI-related representations as an examination focus in both 2025 and 2026. The same evidence binder that governs internal AI use should also support any external claim that the firm uses AI for investment research, portfolio management, compliance, or security.

The AI Control Stack

The five pillars define what to govern. The control stack defines how to enforce it. The layers below describe capabilities, not products. Most firms will assemble these capabilities from their existing technology platform; the specific tooling matters less than whether each capability is actually present, configured, and producing evidence.

Identity and Access

Every AI tool used by the firm should be accessed through corporate identity, with multi-factor authentication, and with administrative visibility.

The capability requirements:

  • A corporate identity provider that authenticates users into AI services. AI tools should not be accessed through personal accounts.
  • Single sign-on federation to every AI service that supports it. SSO produces logs. Personal logins do not.
  • Multi-factor authentication enforced at the identity layer.
  • Administrative ability to provision, deprovision, and audit AI service access centrally.

If an employee can sign into an AI tool with a personal account from a corporate device, the firm has lost the identity layer. That is the first control to close.

Data Loss Prevention

DLP is the technical answer to the data protection pillar. The objective is to detect and block sensitive data from leaving the firm’s environment through AI prompts, browser sessions, or uploads.

The capability requirements:

  • Detection of sensitive data types relevant to the firm: financial account numbers, Social Security numbers, client identifiers, portfolio data patterns, and custom-defined sensitive content.
  • Inspection of outbound traffic to AI service endpoints, including content-aware rules applied to prompts and uploads.
  • Policy enforcement options ranging from user warning to outright block, calibrated to the data classification involved.
  • Alerting and logging of DLP events to a location the firm can review and produce.

DLP that exists but is not configured for AI traffic specifically is DLP that does not address this risk.

Browser Control

A significant portion of AI usage happens in the browser. Controlling the browser controls the channel.

The capability requirements:

  • Enforcement of which AI services are accessible from corporate devices.
  • Control over copy/paste, file uploads, and downloads at the browser level.
  • Ability to permit enterprise versions of AI services while blocking consumer-tier versions of the same product.
  • Telemetry on browser-based AI activity feeding back into the monitoring layer.

Browser control is the layer where policy meets the user. It is also the layer where employee experience matters most. A control stack that makes legitimate work impossible will be circumvented.

Visibility

Visibility is the recognition that the firm cannot govern what it cannot see. Before any DLP rule or browser policy is deployed, the firm should know which AI services its employees are actually using.

A discovery pass, review of proxy logs, productivity-platform sign-in logs, browser telemetry, or a cloud application discovery report, typically surfaces a longer list than the CCO expects. That list becomes the inventory against which the policy is written and the baseline against which future drift is measured.

Logging and Monitoring

The final layer is centralized logging. Without it, the firm has controls but no record of whether they are working.

The capability requirements:

  • A central log destination that ingests identity events, DLP alerts, application activity, and AI-specific telemetry.
  • Correlation of AI service access with sensitive-data events, so the firm can answer the question: “What happened in our environment when this AI activity occurred?”
  • Retention of logs for a defined period consistent with the firm’s books-and-records obligations.
  • Documented review cadence with evidence that the review occurred.

Logging without review is logging for compliance theater. The Remote CISO function or equivalent should be reviewing AI-related alerts on a defined cadence and documenting the review.

What This Looks Like in an Examination

When an SEC examiner asks about AI at your firm, the answer should not be a conversation. It should be a binder:

  • The AI policy, dated and approved.
  • The AI Inventory listing every approved tool, owner, permitted use, and review date.
  • Evidence of vendor due diligence for each tool that touches firm or client data.
  • DLP policy documentation and a sample of logged events.
  • Identity and access configuration showing AI services federated to corporate identity with MFA.
  • A discovery report showing which AI services are in use across the environment.
  • Logs, with evidence that someone reviewed them on a defined cadence.
  • Training records and employee acknowledgments.
  • The books-and-records determination for AI outputs.

A program assembled this way answers the examiner’s real question, not “do you use AI?” but “can you demonstrate that your AI use is governed?”, with documents instead of assurances.


MTradecraft is vendor-independent and does not resell technology. If your firm is deploying AI and wants the framework mapped to your actual environment, that is the work we do.

Related Reading

The BrainTrust, free

This article is public. A free BrainTrust account opens every Insight, the 13 Sample Deliverables, the full Mock SEC Cyber Exam and the core compliance briefs. Name and email, no card.

Create Free Account →
Outside Testing

Want an outside test of your program? One-time engagements are available only to BrainTrust Premium members →