Purview · SharePoint · Teams · Exchange· Refine · DLP and Copilot

Use labels in DLP and Copilot

Purpose

Turn labels from a marking into a control. Once content is labelled reliably, data loss prevention can stop the riskiest actions on it (sharing Highly Confidential files outside the organization, copying them to USB) and Copilot can be kept away from what it shouldn't use. This is where the earlier work on taxonomy, policies and automatic labelling pays off: one decision about a label protects that content everywhere it goes, including in AI.

Overview

DLP rules that act on labels, Copilot kept from processing the most sensitive labelled content, and people told how Copilot treats labels.

Effort
About 10 to 15 person-days
Duration
6 to 10 weeks
Complexity
Medium
Who's in the room
Purview admin, security analyst, taxonomy owner, Copilot service owner, communications lead
Licensing
Labels as DLP conditions follow your DLP licensing. Using DLP to stop Copilot processing labelled files and email needs Microsoft 365 E5, Microsoft Purview Suite, Microsoft 365 E5 Information Protection and Governance, or Office 365 E5. Label inheritance into Copilot-generated content needs Microsoft 365 Copilot alongside E3, E5 or Business Premium.

Risk of not doing it

Labelled content can still be shared outside the organization or summarized into places it shouldn't go, so labels describe risk without reducing it.

Key activities

The plan in order: what must be true first, what to decide, then how to configure the pilot and production.

Prerequisites

  1. Confirm labels are reliable enough to act on

    DLP and Copilot controls based on labels are only as good as the labelling underneath. Check coverage and downgrade patterns with the Maintain stage, and make sure automatic labelling covers your most sensitive locations. A rule that blocks Highly Confidential content does nothing for sensitive files nobody labelled. It helps to agree a simple readiness test with the taxonomy owner, such as most files on the most sensitive sites carrying a label, before any blocking rule is switched on.

    Who
    Taxonomy owner with the security analyst
    Produces
    A go-ahead based on current coverage
  2. Know how Copilot treats labels

    Copilot only uses content the user can access, and for encrypted content the user needs both the VIEW and EXTRACT usage rights; with VIEW alone, Copilot can link to the file but won't summarize it. Copilot Chat shows the labels of its sources and the highest-priority one for the response. In Word, PowerPoint and Outlook, new content Copilot creates inherits the highest-priority label of its sources, and an inherited label can replace a lower-priority label someone applied by hand. Container labels aren't inherited by items, and labels on Teams meetings and chats aren't currently recognized.

    Who
    Copilot service owner with the Purview admin
    Produces
    A short brief on Copilot and labels for stakeholders
  3. Make sure Copilot can read labelled files properly

    Microsoft recommends turning on sensitivity labels for Office files in SharePoint and OneDrive; otherwise the encrypted files Copilot can work with are limited. Files encrypted with user-defined permissions can't be used by Copilot unless they're open or referenced directly, so admin-defined permissions are the better fit for content people will use with Copilot.

    Who
    Purview admin
    Produces
    Platform settings confirmed for Copilot

Decisions

  1. Decision: Which actions should DLP block on labelled content?

    Labels can be conditions in DLP policies for Exchange email, SharePoint, OneDrive, devices, Microsoft 365 Copilot, and Fabric and Power BI workspaces, for Word, Excel, PowerPoint and PDF files. Microsoft's secure-by-default guidance pairs a default label with a DLP rule that blocks sharing with external users and blocks "Anyone" links. Endpoint DLP can act on labels when files are copied to USB, to network shares, printed or uploaded. Legacy .doc, .xls and .ppt files aren't supported for label conditions, so older formats need sensitive information type conditions instead. Start with a small number of clear rules; each extra rule adds overrides and exceptions to manage. Write each rule's purpose in one sentence, so overrides can be judged against it.

    Options

    • Block every external action on every label above General
    • Block external sharing and risky device actions for the top tier; warn for the next tier
    • Monitor only
    Who
    Security analyst with the taxonomy owner
    Produces
    A short list of label-based DLP rules
  2. Decision: Should Copilot process your most sensitive labels?

    A DLP policy for the Microsoft 365 Copilot and Copilot Chat location can stop Copilot using files and email with chosen labels: they may still appear in citations, but Copilot doesn't use their content. It covers stored and open files and email sent from 1 January 2025, not calendar invites; it needs its own policy using the custom template; label and sensitive information type conditions can't be in the same rule; and changes take up to four hours. Some parts, such as blocking based on information in prompts, are in preview.

    Options

    • Let Copilot process everything a user can open
    • Exclude the top tier from Copilot processing
    • Exclude every label above General
    Who
    Taxonomy owner with the Copilot service owner
    Produces
    A decision on which labels Copilot may process
  3. Decision: How will you find oversharing before Copilot does?

    Purview's Data Security Posture Management runs data risk assessments of SharePoint sites, including a weekly assessment of the top 100 sites, and can apply a sensitivity label or remove sharing links from items it finds. It also offers recommendations to create label-based DLP for Copilot and auto-labelling policies. The earlier DSPM for AI is now called the classic version.

    Options

    • Rely on users reporting oversharing
    • Run data risk assessments and act on the highest-risk sites
    Who
    Security analyst
    Produces
    A recurring oversharing assessment

Pilot configuration

  1. Run DLP rules in simulation

    Create the label-based DLP rules and the Copilot policy in simulation mode, review matches with the taxonomy owner, and adjust. Check for legitimate work that would be blocked, such as sending Confidential documents to a regular auditor, and decide on exceptions before anyone sees a block. The Copilot location supports simulation too. Use the simulation period to tell Copilot users what will change, so content appearing only as a link doesn't look like a fault.

    Who
    Security analyst
    Produces
    Tuned rules with agreed exceptions
  2. Test Copilot with labelled content

    As a pilot user, ask Copilot about files with each label: one you can open, one with VIEW but not EXTRACT, one excluded by the DLP policy. Create a document from a Highly Confidential source and check it inherits the label. Write down what people will see so the guidance matches reality. Repeat the tests after each policy change, allowing for the up to four hours the Copilot policy takes to apply.

    Who
    Copilot service owner
    Produces
    Tested scenarios and guidance text

Production configuration

  1. Enforce in waves

    Turn on policy tips first, then enforcement, wave by wave, starting with groups like the pilot. Hold each wave until overrides and service desk tickets settle. For the Copilot policy, remember changes can take up to four hours to apply, so test after each change rather than immediately. Keep a record of agreed exceptions with their owners and review dates. Report progress to the Copilot service owner after each wave.

    Who
    Purview admin
    Produces
    Label-based DLP and the Copilot policy enforced
  2. Tell people how Copilot and labels work together

    Explain in plain language why a Copilot summary can arrive more sensitive than expected, why some files appear only as links, and what to do if a label seems too high. Add it to the labelling guide and to Copilot onboarding. Give the service desk the same guidance, so they can answer 'why can't Copilot summarize this?' consistently.

    Who
    Communications lead
    Produces
    Published guidance on Copilot and labels
  3. Make it enforceable

    DLP rules act on labels wherever the label goes (email, sites, devices) and the Copilot location policy keeps the most sensitive labelled content out of Copilot's processing, without relying on each person's judgement. Coverage still depends on content being labelled, which is why automatic labelling comes first.

    Enforced by a mechanism.

Trial to adoption

Rollout approachBy risk tier · Pilot size: A cross-section of 5 to 10% of users, including Copilot users and people who share externally

Blocking actions on labelled content affects real work. Starting with the top tier and the riskiest actions, tested in simulation, protects what matters most while exceptions are found and agreed.

Make your pilot representative
  • Every major role, not only IT and enthusiasts
  • More than one location or time zone
  • Heavy users of the workload you are changing
  • People who handle sensitive information
  • At least one team you expect to be sceptical
  1. Brief the pilot groupChange: Inform

    Explain which actions will be blocked or warned on, and how Copilot treats labels. Include people who share files with partners every day and heavy Copilot users, since they will meet the new rules first.

    Move on when: Pilot members know what to expect and how to report

  2. Simulate and tuneChange: Explain

    Run DLP rules and the Copilot policy in simulation and agree exceptions. Review matches with the taxonomy owner and the owners of affected content before deciding exceptions.

    Move on when: False positives are tuned and exceptions recorded

  3. Turn on tips, then enforcement, for the pilotChange: Explain

    Show policy tips for two weeks before blocking. Read every override reason during the tips period; they show which exceptions are legitimate.

    Move on when: No legitimate work stays blocked

  4. Extend in wavesChange: Explain

    Widen scope wave by wave. Start with groups like the pilot, then the rest of the organization, holding each wave until overrides settle.

    Move on when: Each wave settles with no rise in escalations

  5. Make guidance routineChange: Sustain

    Include labels in Copilot onboarding and the labelling guide.

    Move on when: People can explain what a label means for Copilot

Common challenges

  • If youBlock on labels before labelling is reliable

    What tends to happenSensitive files without labels pass straight through, and the rules give false confidence.

    A better approach

    Get automatic labelling in place on the most sensitive locations first.

  • If youExclude every label from Copilot

    What tends to happenCopilot becomes useless for most real work, and people find other tools.

    A better approach

    Exclude only the top tier and let labels and permissions handle the rest.

  • If youLet people set their own encryption permissions

    What tends to happenRecipients get VIEW without EXTRACT, and Copilot can't use content it should.

    A better approach

    Use admin-defined permissions on labels people will use with Copilot.

  • If youSurprise people with inherited labels

    What tends to happenA summary arrives labelled Highly Confidential and its author removes the label.

    A better approach

    Explain inheritance before Copilot reaches the wider organization.

  • If youCombine label and sensitive information type conditions in one Copilot rule

    What tends to happenThe rule can't be saved that way, and the policy is delayed while it's redesigned.

    A better approach

    Put label conditions and sensitive information type conditions in separate rules.

  • If youExpect the Copilot policy to cover everything Copilot sees

    What tends to happenCalendar invites and email sent before 2025 aren't covered, and people assume they are.

    A better approach

    State the policy's scope plainly in the guidance and cover other risks with labels and permissions.

Checkpoints

Pilot

  • Is labelling reliable enough on the locations the rules cover?
  • Did every rule and the Copilot policy run in simulation?
  • Have Copilot scenarios been tested for each label?
  • Has the service desk been given the guidance on Copilot and labels?

Production

  • Are label-based DLP rules enforced for every wave?
  • Is the top tier excluded from Copilot processing as decided?
  • Is an oversharing assessment running regularly?
  • Is guidance on Copilot and labels published?

Success measures

Capture a baseline before you start, or you won't be able to show the change.

Measures, where you see them, and the direction that means success
MeasureWhere you see itGood direction
Highly Confidential items shared outside the organizationDLP matches in Activity explorer Down
DLP overrides and false-positive reportsPurview alerts and service desk Down
Overshared items found in data risk assessmentsData Security Posture Management Down
Copilot-created documents carrying the expected labelSpot checks of pilot users' documents Up

Ready-to-use text

Guidance: Copilot and labels
Copilot works with the same labels you do. It only uses content you're allowed to open, and when it creates something from labelled files, the new content takes the most sensitive of their labels. So a summary of a Highly Confidential report is Highly Confidential too. Some of our most sensitive content is kept out of Copilot entirely and will appear only as a link. If a label on something Copilot made looks too high, check where the content came from before you lower it.
Policy tip text for external sharing
This file is labelled Highly Confidential, so it can't be shared outside the organization. If you need to share it with a partner, contact [team] to agree a safe way.
Exception request
Rule: [DLP rule]. What the person needs to do: [action, e.g. send Confidential files to our external auditor]. Who: [people or group]. For how long: [dates]. Safer alternative considered: [yes/no, why not]. Approved by: [taxonomy owner], review date: [date].

Additional reading

Changelog

Changelog (1 version)
Published versions of this playbook
VersionDateAuthorChange
1Sev DerghazarianFirst version. Facts checked against Microsoft Learn on 2026-10-15.