How to Label Client Files So AI Tools Know What to Leave Alone

AI assistants are now built into the email, documents and browsers advisers use every day. They are fast because they read whatever the signed-in user can open. For an RIA, broker-dealer, insurance agency or family office, that includes account applications, copies of driver's licenses and transfer forms. Sensitivity labels for financial firms tell those tools which client files to leave alone.

Why do AI tools need sensitivity labels to know what to leave alone?

AI tools need labels because, without them, a client's account application and a marketing one-pager look the same: both are files the user is allowed to open. A label adds a second signal that travels with the file and that AI and data loss prevention (DLP) controls can act on.

Microsoft states that Copilot "only surfaces organizational data to which individual users have at least view permissions." That is a reasonable default, but in most small firms nearly everyone can open nearly everything. Microsoft also notes that a sensitivity label is stored in the file's metadata, so "the label stays with the content, no matter where you save or store it," and other apps and services can read it.

Which client files should you label first?

Start with the files that hold sensitive customer information as Regulation S-P defines it. Those are the files whose exposure can trigger customer notification.

The amended rule's definition at 17 CFR 248.30(d)(9) gives examples such as Social Security numbers, government ID numbers, biometric records, and account identifiers combined with information that would allow access to the account. At most advisory and brokerage firms those show up in new account paperwork, ID copies collected for customer identification, transfer and ACAT forms, tax documents and trust paperwork.

Regulation S-P separately requires written policies and procedures reasonably designed to protect customer information against unauthorized access or use. Labels alone do not meet that requirement. They give your AI and DLP controls something to act on. If you have not mapped where this data lives, start with how to map sensitive customer information for Reg S-P.

How many sensitivity labels does a small financial firm need?

Three or four is enough for most firms. Microsoft's own guidance says effectiveness is "noticeably reduced when users have more than five main labels."

A workable set for a financial firm:

• Public: approved marketing and published material.
• Internal: firm business with no client data.
• Client Confidential: anything containing client information.
• Client Restricted: sensitive customer information such as ID copies and account applications, with encryption turned on.

Each file carries one sensitivity label at a time, and Microsoft recommends ordering the list so the most restrictive label sits at the bottom. Keep each label's description to one sentence an operations associate can apply without asking.

What do labels actually do to AI tools?

Labels let Copilot and DLP tools skip or block sensitive files, but only after you configure the controls that act on them. What happens next depends on the tool.

• Microsoft 365 Copilot. When a label applies encryption, Copilot "honors the usage rights granted to the user." A DLP policy using the Copilot location can also exclude items with chosen labels, so Copilot does not use their content in a response. Microsoft notes those items can still appear in citations, that email coverage starts with messages sent on or after January 1, 2025, and that files uploaded directly into a prompt are not scanned.
• Public AI sites in the browser. Microsoft Purview can block typed or pasted text and file uploads to unmanaged AI apps in Edge for Business on Intune-managed Windows devices. Microsoft's example detects bank account and routing numbers rather than labels, and not every AI app is supported, so test against the tools your staff use.
• Other software. Because the label sits in clear-text metadata, CRM, portal and AI vendors can read it and apply their own rules. Ask each vendor whether it does.

Unlabeled files get none of the label-based protections. Browser DLP can still catch typed or pasted Social Security and account numbers, but Copilot exclusions and encryption only follow a label. That is why the rollout matters more than the label design.

Which Microsoft licenses do you need?

Microsoft 365 Business Premium includes manual labeling. Automatic labeling, default labels on SharePoint libraries and Endpoint DLP are listed under E5-level plans in Microsoft's Purview service description.

The Edge for Business AI controls are billed on a pay-as-you-go or per-user basis. Licensing changes often, so confirm current terms with your Microsoft partner before you design around a feature.

How do you roll labels out without slowing the office down?

Roll labels out in small steps, starting with the files that carry the most risk:

1. Publish the three or four labels to a pilot group, usually the CCO and operations staff, for two weeks.
2. Label the client folders that hold account opening and ID documents first. On E5-level licensing, a default library label can do this for every new file in the folder.
3. Turn on encryption only for Client Restricted, then test custodian uploads, e-signature and scanning workflows, since encryption can change how other systems open a file.
4. Add a Copilot DLP rule that excludes Client Restricted from processing.
5. Add browser DLP for public AI sites so pasted Social Security and account numbers are blocked or flagged.
6. Publish labels to everyone with a short demo and a one-page guide.
7. Save the label definitions, policy settings and a test result as evidence for your Reg S-P policies.

For the policy side of this work, see how to turn your AI policy into Conditional Access and DLP rules.

What should you do this quarter?

You can make real progress in one quarter without touching every file:

• Pick your label names and write the one-sentence definitions.
• List the five folders or systems that hold the most sensitive customer information.
• Check your Microsoft licensing so you know which steps are automatic and which are manual.
• Run the pilot and label the account opening folders.
• Turn on one Copilot or browser DLP rule and record the result.

How can s90 help?

s90 designs and deploys sensitivity labels, Copilot controls and browser DLP for financial firms in Los Angeles and Orange County, and documents the settings as evidence for Reg S-P and exam requests. If you want to know what your AI tools can reach today, we can run that check with you.