Your Firm Is Already Using AI. Here’s How to Make It Defensible.

Somewhere in your firm right now, an advisor is pasting text into an AI chatbot. Maybe it’s meeting notes. Maybe it’s a draft client letter. Maybe (and this is the one that should keep you up at night) it’s a spreadsheet with account numbers in column C.

You can’t policy-memo your way out of this. Employees use AI because it works. The question for a CCO isn’t “should we allow AI?” It’s “can I demonstrate to an SEC examiner that AI use at my firm is governed, monitored, and documented?”

If your firm runs on Microsoft 365, and if you’re a $300M–$3B RIA you almost certainly do, the good news is that you already own, or can license into, most of the control infrastructure you need. One honest caveat up front: several of the controls below assume Microsoft 365 E5 or add-on licensing. An E3 shop should read this as a deployment map and a licensing conversation, not a turnkey recipe. This article walks through how to deploy Claude (Anthropic’s AI assistant) inside the Microsoft ecosystem in a way that holds up under Reg S-P, Reg S-ID, Rule 206(4)-7, and Rule 204-2.

A note on scope. We use Claude as the working example because it’s the third-party model our clients ask about most, but the same principles apply to any external LLM. We’re also deliberately not writing about Microsoft Copilot. Copilot is the obvious in-family choice, and Microsoft’s own guidance covers it. The harder question, and the one firms actually bring us, is how to integrate a third-party AI into the Microsoft stack without losing the governance story. That’s this article.

The framing that matters: Microsoft is the control plane. The AI is a governed workload. Get that relationship right and the exam narrative follows from evidence you already collect. Get it backwards, AI first and controls later, and you’ve built a data exfiltration channel with a friendly interface.

First Decision: Where Does the AI Actually Live?

Before any technical controls, you pick a trust boundary. There are three realistic options for a Microsoft-centric RIA:

1. Claude Enterprise (browser/app). Anthropic’s business tier. Your users sign in through your identity system, business data isn’t used for model training, and you get admin audit logs. Data leaves your tenant to Anthropic, under contract. This is the fastest path and the most common.

2. Claude via the API. Your firm (or your IT provider) builds internal tooling that calls Claude programmatically. Maximum control over data flow, but you’re now maintaining software. Realistic only if you have engineering capacity most RIAs don’t.

3. Claude inside Microsoft Foundry (Azure). Anthropic’s models are now available through Microsoft’s Azure AI Foundry platform, which brings procurement, billing, and network controls into your existing Azure relationship. Two caveats before this becomes your answer to “where does our data go?”: Foundry offers both Azure-hosted and Anthropic-hosted configurations, so verify which one you’re actually buying, and in either configuration Anthropic doesn’t vanish from your data flow diagram or your vendor file. What you gain is residency options, network control, and one fewer procurement relationship, not “the data never leaves us.” Verify current regional availability and feature parity before committing; this is a fast-moving product area.

(There’s a fourth path: running an open-weight model entirely in-house, with no cloud provider at all. We covered it in What It Actually Looks Like to Run a Local LLM at Your Firm. That route trades frontier-model capability for maximum data control. This article is for firms that want the frontier capability, with the Microsoft stack as the leash.)

Whichever you choose, the critical corollary: sanction one, block the rest. A governed enterprise tenant means nothing if employees can still log into personal AI accounts from work devices. More on enforcement below.

The Microsoft Components, and What Each One Does

Here’s the architecture, component by component, in the order an examiner would walk it.

Entra ID: Who Gets In

Entra ID (formerly Azure AD) is your identity system, and it becomes the front door to Claude. Single sign-on means every Claude session is tied to a named, MFA-verified employee identity. SCIM provisioning means offboarding can be automated: disable the Entra account and Claude access goes with it, no ticket required. “Can be” is doing real work in that sentence. Test what actually happens to a departed user’s active sessions, and put that test in your 206(4)-7 file.

Conditional Access is where this gets powerful: you can require that Claude is only reachable from compliant, managed devices, from expected locations. An employee’s personal laptop at home? Blocked at the identity layer before a single prompt is typed.

Regulatory hook: Reg S-P’s safeguards rule requires administrative and technical controls over access to customer information. Access reviews on the Claude enterprise app become a standing 206(4)-7 testing artifact.

Microsoft Purview: What Data Gets Through

Purview is your data governance layer, and it does two jobs here.

Sensitivity labels and endpoint DLP. You’ve (hopefully) already labeled documents containing nonpublic personal information. Purview Endpoint DLP, running on managed devices in supported browsers, can warn on or block sensitive content being pasted or uploaded into AI websites. One wrinkle to configure around: personal and enterprise Claude live on the same domain, so DLP by itself can’t tell your sanctioned workspace from an employee’s personal account. The working pattern is DLP at the domain level plus Entra SSO enforcement, so the only Claude anyone can sign into from a firm device is yours. Test the combination and keep the evidence. The employee pasting that spreadsheet gets stopped in the moment, and the attempt gets logged.

AI activity monitoring. Purview’s data security posture management for AI gives you visibility into what users actually send to AI services, including, as of this writing, audit visibility into Claude Enterprise specifically. Know the boundary: this integration is visibility and auditing, not DLP or eDiscovery inside Claude, and it produces evidence only after you configure it (some capabilities are pay-as-you-go). Still, it’s the difference between hoping your policy is followed and demonstrating it.

Regulatory hook: This is the mechanical heart of your Reg S-P program as applied to AI. Policies without enforcement are wallpaper; DLP is enforcement.

Defender for Cloud Apps: Killing Shadow AI

Defender for Cloud Apps is Microsoft’s cloud access security broker. Its job in this architecture is simple: discover every AI service touching your network, mark the personal/consumer versions as unsanctioned, and block them. One prerequisite worth naming: “unsanctioned” becomes an actual block only when Defender for Cloud Apps is wired to Defender for Endpoint on your managed devices, or to your firewall or secure web gateway. On its own it’s a label, not a wall. Wired up correctly, your sanctioned Claude tenant stays open and everything else gets the wall.

Take a screenshot of the block page. That screenshot goes in your exam file. It’s one of the cheapest, most persuasive artifacts you’ll ever produce.

Intune: Which Devices Qualify

Intune manages device compliance: encryption, patching, screen lock, endpoint protection. Paired with Conditional Access, it ensures AI access only happens from machines you control. This closes the “I’ll just use my phone” gap.

Exchange, SharePoint, and the Graph: When AI Touches Your Content

If you go further and connect Claude to mailboxes or document libraries (increasingly common), the rules tighten: app registrations scoped to least-privilege permissions, admin consent required, and, critically, AI access should inherit your existing SharePoint permissions, never bypass them. Two things to validate before you trust that sentence: confirm the connector authenticates as the individual user, so current permissions and revocations are honored, and remember that a sensitivity label traveling with a file doesn’t mean the AI enforces it. Enforcement stays in your Microsoft stack. If an employee can’t open a document, neither should their AI session. Verify that; don’t assume it.

Sentinel: Where the Story Gets Stitched Together

Microsoft Sentinel is your SIEM. Entra sign-in logs, endpoint DLP events, Defender discoveries, and Claude’s enterprise audit logs can all land in one place. “Can,” because each source has its own connector, licensing, and retention settings; “it’s all in Sentinel” is a configuration outcome, not a default. Know precisely which questions your logs can answer before an incident asks them: Claude’s audit logs record administrative and usage events, not full prompt content. What you get, wired correctly, is the ability to build detections for anomalous behavior (say, a user suddenly pushing unusual volumes of content into AI prompts at 11 PM) and to reconstruct who accessed what, when.

Regulatory hook: The amended Reg S-P requires notifying affected individuals as soon as practicable, and no later than 30 days after you become aware that unauthorized access to or use of sensitive customer information occurred or is reasonably likely, unless a reasonable investigation concludes substantial harm or inconvenience isn’t reasonably likely. You cannot make that determination, or notify, about what you cannot reconstruct. Sentinel plus retained audit logs is your reconstruction capability, and your incident-response and 206(4)-7 testing evidence. (Rule 204-2 is a separate analysis: specific AI-assisted records and communications may be required books and records depending on their content and use. Classify them deliberately rather than assuming the SIEM archive covers it.)

The Exam File: Six Artifacts

Our operating principle at MTradecraft: if it cannot be demonstrated to an SEC examiner with evidence, it did not happen. The technology above only matters insofar as it produces the following documentation:

  1. AI Acceptable Use Policy: names the sanctioned tenant, prohibits personal accounts, adopted by the firm, acknowledged by employees.
  2. Vendor due diligence file on your AI provider: SOC 2 report, data processing agreement, the no-training-on-business-data commitment, subprocessor list. The amended Reg S-P requires due diligence and monitoring of service providers that receive, maintain, process, or are otherwise permitted access to customer information, and whether your AI vendor meets that definition is decided by the architecture you chose above. It also requires procedures reasonably designed to ensure providers notify you within 72 hours of becoming aware of a breach of their systems. Get that into the contract. (BrainTrust Premium members: this belongs in your VDD workflow like any other critical vendor.)
  3. Data flow diagram: one page showing what NPI can and cannot reach the AI, and which Purview policy enforces each boundary.
  4. Periodic access reviews of the Claude enterprise application in Entra. Quarterly is our recommended cadence, a risk-based choice rather than a rule requirement.
  5. Incident response playbook with an AI-channel scenario, mapped to the S-P notification timeline. If your firm maintains covered accounts under Reg S-ID, add an AI-assisted impersonation scenario (cloned client voice, fraudulent wire instructions) to your identity-theft red-flags program.
  6. Log retention evidence: proof you could answer “who sent what, when” for any date in your retention window.

Notice what this list is: your existing 206(4)-7 program, extended to a new system. AI governance isn’t a new discipline. It’s your compliance program doing what it already does, pointed at a new target.

The One-Paragraph Version for Your Managing Partner

Microsoft decides who can use the AI, from which devices, with what data. Purview watches and blocks. Sentinel records. The compliance file proves all of it. The firm gets the productivity benefit of AI without opening an ungoverned channel for client data. When the examiner asks about AI, the CCO opens a folder instead of starting a scramble.


MTradecraft LLC is an independent cybersecurity compliance consultancy serving SEC-registered investment advisers, hedge funds, broker-dealers, and family offices from Dallas, TX. We sell no hardware, resell no services, and take no vendor commissions. Our recommendations answer to your exam file, not a channel partner agreement. If you want a governed AI deployment assessed end-to-end, reach out: brian@mtradecraft.com.

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 →