Most RIA AI-governance conversations start with the wrong question: “should we adopt AI?” The honest starting point is that adoption already happened, and you weren’t in the room. Microsoft built Copilot into the 365 ecosystem. Zoom added AI Companion. Your CRM, your note-taker, your portfolio commentary tool, your phone system’s transcription; each shipped AI features in the last two years. Whether a given feature is enabled, bundled, tenant-configurable, or one admin-click away depends on your licensing and settings, which is precisely the problem, because at most firms nobody has checked. Each feature changes where client information travels and what it may train.
Your AI policy may say the firm doesn’t use generative AI. Your vendor stack disagrees.
Why this is a vendor due diligence problem, not a technology debate
When a vendor adds AI features to a product that touches client data, three things may have silently changed underneath your existing diligence file: where data goes (a subprocessor you’ve never heard of may now receive meeting audio or document content), what it’s used for (training and “service improvement” rights hide in updated terms), and who can see outputs (a meeting summary emailed to all attendees is a new disclosure path for anything said in the meeting). Your SOC 2 from last year predates all of it. The diligence file is now describing a product that no longer exists.
The five questions to send every vendor that touches client data
- Which AI features exist in the product today, and which are enabled by default for our tenant?
- Is our data used to train models, yours or a third party’s, and is the answer contractual or just current practice?
- Which subprocessors receive our data for AI features, and did the subprocessor list change when the features launched?
- Can we disable the features at the tenant level, and does anything else break when we do?
- How are AI outputs retained, and do they enter our books-and-records obligation? (A transcription or AI summary of a client meeting may become a required record if it captures advice, recommendations, instructions, or other covered business communications; treat retention as the default until you’ve analyzed the content.)
Send these to the vendors you already have, not just the ones you’re evaluating. The answers go in the vendor file next to the SOC 2, and the unanswerable ones tell you which vendors haven’t thought about it either.
The supervision link
The SEC’s posture on AI has been consistent: existing obligations apply to new tools. Client data entering an AI feature is a Reg S-P safeguards question. AI-generated content reaching clients is an advertising and supervision question. Meeting transcripts and summaries are recordkeeping questions. None of this requires a new rule to be examinable now; it arrives through the document requests you already receive, asking what tools touch client data and how the firm supervises them. “We didn’t know the feature was on” is an answer; it is just not a good one.
The tenant audit that takes one afternoon
Inventory the AI surface you already own: the M365 Copilot licensing and data-access configuration, Zoom/Teams AI summaries, CRM AI features, browser extensions on employee machines, and whatever the phone system started transcribing. For each: on or off, who decided, where outputs land. Most firms find at least one feature enabled that nobody chose, and at least one place AI outputs are accumulating outside the retention system. Turn off what you can’t justify; document what you keep; revisit quarterly, because the vendors will keep shipping.
BrainTrust members: the AI Governance Workbook structures the tenant audit and the policy decisions that follow it, and the Vendor Due Diligence Questionnaire’s AI addendum carries the five vendor questions above in sendable form.