Shadow AI Is Now an SEC Exam Finding. Here's What That Means for RIAs.

In December 2025, the SEC published its 2026 examination priorities for registered investment advisers. AI governance shows up across multiple categories: cybersecurity, emerging technology, automated investment tools, and operational resiliency. The guidance is direct - examiners expect to see whether firms have adequate policies and procedures to monitor AI use, including unsanctioned use by staff.

That word "including" is doing a lot of work. It means the examination isn't just asking whether you've deployed AI. It's asking whether you know what your people are doing with it on their own.

What examiners are actually looking for

According to guidance published by Goodwin and Harvard Law's Corporate Governance forum tracking the SEC's priorities, examiners in 2026 expect to review three things specifically: the accuracy of firms' AI disclosures, whether there are adequate policies and procedures to monitor AI use across both investment and operational functions, and whether staff have received training on appropriate use.

FINRA is aligned. Their 2026 Regulatory Oversight Report dedicates a full section to generative AI, treating it as a supervised technology that demands the same compliance rigor as any other critical system. Their framing is explicit: technology-neutral rules continue to apply when employees use GenAI, and firms must evaluate how those tools comply with applicable regulations before staff start using them - not after.

Most firms have not done that evaluation. Their staff started using AI tools anyway.

Why investment advisers face specific exposure

Wealth management handles some of the most sensitive personal data in financial services. Estate plans, tax returns, trust documents, portfolio records, beneficiary designations - all of it is non-public personal information (NPI) covered under Regulation S-P. The 2024 amendments to Reg S-P tightened the requirements significantly: firms must now notify affected clients within 30 days of a data breach, and they must report to the SEC within the prescribed window.

An adviser pasting a client's estate planning documents into a public chatbot to generate a summary is not obviously a breach. But that data has left the firm's network, it's been processed by a third-party model, and it's now potentially part of a training dataset the firm has no visibility into. Whether or not an incident results, the supervision failure is real: an employee used client NPI with an unsanctioned tool, and the firm's written supervisory procedures either didn't cover it or weren't enforced.

That combination - a data handling gap plus a supervision gap - is exactly what an examiner is trained to find.

The supervision obligation CCOs are sitting on

The CCO owns the written supervisory procedures. If staff are using AI tools that aren't addressed in those procedures, that's the exam finding. It doesn't require a breach. The absence of a policy covering a tool employees are actively using is itself the problem.

FINRA makes this point without much softening: "firms must consider how they will comply with applicable regulations when evaluating GenAI tools prior to testing and deployment." Not during. Not after an incident. Prior. For firms that haven't run that evaluation yet, the answer to "which AI tools are your people using?" is almost certainly not "none." The data on shadow AI adoption in regulated industries is consistent: employees adopt available tools to do their jobs faster, and they do it quietly when they expect the answer to a formal request to be no.

A 2025 KPMG global study found 57% of employees conceal their AI use from employers. That number is probably higher in regulated environments, where people correctly expect that admitting it creates problems.

What closing the gap actually looks like

There are two parts. The first is documentation: written supervisory procedures that address AI use explicitly, covering what's permitted, what's prohibited, and how compliance will monitor it. FINRA recommends formal governance with clear ownership and pre-approval requirements for new use cases, including documented purpose, data sources, and control design. That's the policy layer.

The second part is harder. A policy that says "don't use unsanctioned AI" without a sanctioned alternative is a policy that asks your people to do their jobs slower. They won't. The research is consistent on this: the only intervention that actually reduces shadow AI adoption is giving people a tool they're allowed to use.

For an investment adviser, that tool needs to meet a specific bar. Client data can't leave the network. Every query and response needs to be logged for examination readiness. Access needs to map to the firm's existing permission structure, so the right documents are only visible to the staff assigned to the right clients. A private AI environment that runs inside the firm's own infrastructure - without routing data through external APIs - is the architecture that satisfies all three requirements at once.

It also means that when an examiner asks whether you have policies and procedures to monitor AI use, the answer is yes, and the logs are right there.