How to Turn Your AI Policy Into Conditional Access and DLP Rules

Why does AI policy enforcement matter in an SEC or FINRA exam?

For an RIA, broker-dealer or dual-registered firm, AI policy enforcement is the part an examiner tests. A written AI policy shows intent, and technical rules show it works. Regulation S-P requires every covered institution to keep written policies and procedures that address administrative, technical and physical safeguards for customer information, reasonably designed to guard against unauthorized access to or use of that information. An AI policy that lives only in the compliance manual covers the administrative piece and leaves the technical piece open.

The supervisory rules point the same way. Advisers Act Rule 206(4)-7 requires an SEC-registered adviser to review, at least annually, the adequacy of its policies and the effectiveness of their implementation. FINRA Regulatory Notice 24-09 reminds broker-dealers that FINRA rules, including the supervision requirements of Rule 3110, are technology neutral and apply when a firm uses generative AI. The SEC Division of Examinations listed emerging technology and AI, cybersecurity and the Regulation S-P amendments among its fiscal 2026 examination priorities. "We have a policy" is the start of the conversation with an examiner, and "here is the control behind each line" is the end of it.

Which policy statements map to which controls?

Most AI policies at advisers and broker-dealers come down to six or seven sentences. Each one has a matching Microsoft 365 control:

• "Use only firm-approved AI tools." Mark unapproved AI apps as Unsanctioned in Microsoft Defender for Cloud Apps so firm devices block them, and limit approved tools to firm accounts (see how to block personal AI accounts on firm devices).
• "Use AI only on firm-managed devices with multifactor authentication." A Conditional Access policy on each approved AI app that requires MFA and a compliant device.
• "Never enter client nonpublic personal information into an unapproved AI tool." Microsoft Purview endpoint DLP rules on paste and upload to browsers, plus inline DLP for unmanaged AI apps in Edge for Business.
• "Copilot must not process restricted files." A DLP policy on the Microsoft 365 Copilot location that excludes items carrying a restricted sensitivity label.
• "Personal devices get limited access." A Conditional Access session policy through Conditional Access app control that blocks or monitors downloads and uploads on unmanaged devices.
• "New AI tools need approval before use." Entra ID user consent settings that send app connection requests to an admin consent workflow, plus a vendor review of data retention and training terms before approval.
• "AI use is supervised." Audit logs and alert review, with dated records of who reviewed what.

Write the control name next to each sentence in the policy itself. That one change turns the annual review from a reread of the text into a test of whether each control still runs. If the policy text still needs work, start with how to write an AI acceptable use policy your tools can enforce.

How do you build the Conditional Access rules?

Start with one policy per approved AI app that requires MFA and a compliant device, then add a second policy for unmanaged devices. Microsoft's Conditional Access deployment guidance says the tenant needs Microsoft Entra ID P1, P2 or a trial license, and lists the available grant controls, including MFA, a device marked compliant, an approved client app and an app protection policy. Session controls include Conditional Access app control and sign-in frequency.

A working set for a small advisory firm looks like this:

• Policy 1: all users, target the approved AI app (for example Microsoft 365 Copilot, or an enterprise ChatGPT or Claude workspace connected to Microsoft Entra ID for single sign-on), require MFA and a compliant device.
• Policy 2: users on unmanaged devices, same app, route the session through Conditional Access app control so downloads and uploads can be blocked or monitored.
• Policy 3: block legacy authentication and any sign-in path that skips the first two policies.

Microsoft recommends running each policy in report-only mode for at least one week before enforcement, and warns that combining block with all resources in one policy can lock out administrators. Keep emergency access accounts excluded from every policy and write down why. Conditional Access app control can block download, cut, copy and print of sensitive documents on unmanaged devices and block upload of unlabeled files with sensitive content, so a personal phone or home computer gets a narrower session than a firm laptop.

How do you build the DLP rules?

DLP rules enforce the data sentence of the policy: client information stays out of unapproved AI tools. The Purview endpoint DLP paste-to-browser activity offers three actions (Audit, Block with override and Block) and works in Edge, Chrome and Firefox on Windows with the Purview extension, and in Safari on macOS. For browser-based AI apps, inline DLP in Edge for Business can block sensitive text and file uploads sent to unmanaged generative AI apps on Intune-managed Windows devices.

Build the rules around the data your firm actually holds: Social Security numbers, account numbers, dates of birth, and custodian or client identifiers. Then stage enforcement:

1. Audit for two to four weeks to see where the data goes and which teams send it.
2. Move to Block with override for approved workflows, which asks the user for a business reason and logs it.
3. Move to Block for unapproved AI sites.

Check licensing before you plan. As of October 2026, Microsoft's Purview service description lists Endpoint DLP under Microsoft 365 E5 and the Purview suites, and lists Microsoft 365 Business Premium only for DLP in Exchange, SharePoint and OneDrive. If Copilot is approved, a DLP policy on the Copilot location can exclude labeled files and emails from Copilot processing; Microsoft notes updates can take up to four hours to apply.

What about AI tools you cannot sign in to or label?

Block them at the device. In Defender for Cloud Apps, apps marked Unsanctioned sync to Microsoft Defender for Endpoint, and their domains are blocked on onboarded devices, with up to three hours of delay. This needs Defender for Cloud Apps with Defender for Endpoint (or Microsoft 365 E5) and Microsoft Defender Antivirus network protection in block mode. Apps marked Monitored show users a warning with a link to your approved list and let them continue, which is a reasonable first step while staff adjust.

Point the block page's support link to an internal page that lists approved AI tools and explains how to request a new one. That keeps the block and the approval process connected, and every request becomes a record that the approval process runs.

What evidence will an examiner ask for?

Expect a request for the policy, proof that the controls exist, and proof that someone reviews them. Regulation S-P written policies and procedures are required records: Rule 204-2(a)(25) for advisers and Rule 17a-4(e)(14) for broker-dealers. FINRA's 2026 oversight report lists effective practices that include written GenAI policies, formal review and approval of use cases, and storing prompt and output logs.

A practical evidence folder holds:

• The AI policy with the control mapped to each sentence.
• Dated exports or screenshots of each Conditional Access and DLP policy, showing it is on.
• A monthly summary of DLP alerts and overrides, with the reviewer's name and date.
• The list of sanctioned, monitored and unsanctioned AI apps, with approval dates.
• The annual review memo, noting which controls were tested and what changed.

What should you do this quarter?

1. Rewrite the AI policy into short, testable sentences, one rule each.
2. Assign a control and an owner to every sentence.
3. Confirm licensing for Entra ID P1 and for the DLP features you plan to use.
4. Turn on Conditional Access for approved AI apps in report-only mode, review for a week, then enforce.
5. Run DLP in audit mode, then move to Block with override and Block on a set schedule.
6. Mark known unapproved AI apps as Unsanctioned and publish the approved list.
7. Start the evidence folder now so the next annual review has data in it.

How s90 helps

s90 maps your AI policy to Conditional Access, DLP and app controls in Microsoft 365, runs the report-only and audit phases with you, and keeps the evidence folder current for your annual review and your next exam. Our agreement includes 72-hour breach notification and 24-hour evidence production, so the records an examiner asks for are on hand. If you want to see how your current policy maps to controls, start with a short review call.