StratiFi Blog — Insights for RIAs and Broker-Dealers

Shadow AI Is Already in Your Firm — Here's What the SEC Will Ask

Written by Akhil Lodha | 9/11/26, 9:30 PM

Shadow AI — personnel using AI tools the firm never approved — is already happening inside most RIAs, whether or not leadership knows it. The SEC's question in an exam won't be "did you allow it?" It will be "did you know about it, and what did you do?" A firm that can't answer has a supervision problem under Rule 206(4)-7, not a technology problem.

Your firm doesn't have an AI adoption decision to make. It already made it — informally.

Most founders and CCOs believe AI adoption is a choice sitting on next quarter's agenda. Here's the reframe: your advisors made that choice for you months ago. They're drafting client emails in ChatGPT. They're summarizing meeting notes with a transcription bot. Someone in ops is pasting spreadsheet data into a chatbot to "clean it up." None of it went through compliance, because none of it felt like a compliance event. It felt like saving twenty minutes.

That's shadow AI, and it's not a hypothetical. It's the predictable result of putting genuinely useful tools one browser tab away from overworked people. And it's broader than chatbots: meeting-notes transcribers, AI "polish" features inside email clients, machine-learning capabilities quietly added to the CRM in a product update, browser extensions that summarize PDFs. Some of it your firm technically pays for — without ever having reviewed the AI features shipped inside software you approved years ago.

Shadow AI in RIA compliance is the use of unapproved, unsupervised AI tools by firm personnel for work involving firm or client information. The danger isn't that the tools exist. It's that their use is invisible to your compliance program — no inventory, no review, no records — which converts everyday convenience into unsupervised activity.

Why is shadow AI a supervision problem, not an IT problem?

Because the Advisers Act doesn't care where the tool came from — it cares whether the activity was supervised. Rule 206(4)-7 requires written policies and procedures reasonably designed to prevent violations of the Advisers Act, implemented in practice, not just on paper. When an advisor uses an unapproved AI tool to produce client-facing work, three exposures open at once:

Client data leaves your perimeter. Nonpublic client information typed into a consumer AI tool may be stored, logged, or used to train models outside your control. That intersects Regulation S-P and its safeguarding obligations — and after the amended rule's compliance dates, your incident response obligations too.

Advice gets produced outside your review chain. If AI-assisted recommendations or client communications never pass through documented review, your supervision program has a blind spot exactly where examiners now look.

Records don't exist. Books-and-records obligations under Rule 204-2 assume the firm can produce required communications and documentation. Work done in a personal chatbot account generates records the firm doesn't hold and can't produce.

The pattern rhymes with the off-channel communications problem — years of SEC penalties driven not by the existence of texting, but by business happening where firms couldn't see or preserve it. Shadow AI is off-channel work product.

What will the SEC actually ask about unapproved AI tools?

Expect artifact requests, not abstract questions. The SEC's FY2026 exam priorities state that examiners will assess advisers' policies for monitoring the use of AI technologies — monitoring, not merely permitting or prohibiting. In practice, requests in this area look like:

  • An inventory of all AI tools used by the firm and its personnel — the phrase "and its personnel" is where shadow AI surfaces
  • The firm's policies governing personnel use of third-party AI tools, including generative AI
  • A description of how the firm monitors for unapproved AI use, not merely whether it prohibits it
  • Training records showing personnel were informed of the firm's AI rules
  • Any incidents involving client data and AI tools, and how they were handled

Notice the trap in the third item. A firm with a strict written ban and zero monitoring is arguably worse off than a firm with a permissive policy it actually enforces. The unenforced ban is documentary evidence that the firm identified a risk and then didn't supervise it.

What does the shadow AI problem cost when it surfaces?

The direct cost is exam findings and remediation. The indirect cost is worse: once examiners find one unsupervised workflow, the presumption of control is gone, and everything gets a second look. Extended exams consume the compliance calendar for months. And if client data actually leaked through a consumer AI tool, you may be in breach-notification territory — with the added embarrassment that the "vendor" was never vetted because it was never disclosed.

There's an internal cost too. Firms that respond to shadow AI with blanket bans push usage further underground and lose their best people's productivity gains. The advisors don't stop; they just stop telling you.

How do you bring shadow AI into the light?

Treat it like the supervision problem it is — with a program, not a memo:

  1. Survey without punishment. Run an amnesty-style inventory: what tools are people using, for what? You cannot govern what you refuse to know.
  2. Approve a real alternative. People use shadow AI because it solves real problems. Give them approved tools that solve the same problems, or the shadow economy continues.
  3. Write an acceptable-use policy that matches reality. Which tools, which use cases, what may never touch an AI tool (client PII, holdings, anything nonpublic), and what review applies to AI-assisted output.
  4. Monitor and document. Periodic attestations, spot checks, and — critically — a system of record where reviews and exceptions are logged as they happen.
  5. Fold it into the annual review. Shadow AI risk belongs in your Rule 206(4)-7 annual review, with the [drift and policy-breach discipline you apply to portfolios](/blog/ips-drift-style-drift-policy-breach-ria-compliance/) applied to tooling.

The last two steps are where spreadsheets quietly fail. Supervision that lives in someone's inbox produces no evidence. StratiFi's ComplianceIQ takes the opposite approach: exceptions get flagged, reviews get logged, and every supervisory action creates an audit trail automatically — so when the exam letter asks how you monitor AI use, you produce a report instead of a narrative. It's the same architecture that handles portfolio supervision, which means AI oversight doesn't become yet another disconnected tool. For the broader exam context, our breakdown of the SEC's 2026 exam priorities covers where AI sits in the current cycle.

Frequently Asked Questions

What is shadow AI in a financial advisory firm?

Shadow AI is the use of AI tools — chatbots, transcription services, document summarizers — by firm personnel without firm approval, oversight, or recordkeeping. It typically involves consumer-grade tools accessed through personal accounts, invisible to the firm's compliance program.

Is it a violation for an advisor to use ChatGPT at an RIA?

Not inherently. The exposure depends on what data enters the tool, whether the use is supervised, and whether required records are kept. Entering nonpublic client information into an unapproved consumer AI tool creates safeguarding and supervision risk; drafting a generic paragraph with no client data is a far smaller issue. Firm policy should draw exactly that line.

Should an RIA just ban AI tools entirely?

A total ban is usually the weakest option. Bans without monitoring create documented, unenforced policies — which examiners read as supervision failures — while pushing usage underground. A defined set of approved tools plus monitoring is more defensible and more realistic.

How can a small RIA detect shadow AI use?

Start with attestations and an amnesty inventory, then add periodic spot checks of client-facing work product and, where feasible, network-level visibility into AI service usage. Perfection isn't the standard; a documented, reasonable monitoring effort is.

Most firms discover their shadow AI problem during an exam. The better order is to discover it yourself — start the inventory this week.