M365 PlaybooksPurview · SharePoint · Teams · Exchange· Refine · Automatic labelling
Label content automatically, starting in simulation
Purpose
Close the gap manual labelling always leaves: files stored before labels existed, and new content people forget to label or label too low. Automatic labelling finds sensitive content by what it contains and labels it, in the apps while people work and in SharePoint, OneDrive and Exchange without anyone opening a file. Done carefully, it makes label coverage something the organization can rely on rather than hope for.
Overview
Labels recommended or applied in Office apps, and service-side policies that label stored files and email, tuned in simulation first.
- Effort
- About 15 to 25 person-days
- Duration
- 8 to 12 weeks
- Complexity
- High
- Who's in the room
- Purview admin, security analyst, taxonomy owner, content owners for pilot sites, communications lead
- Licensing
- Client-side and service-side automatic labelling need Microsoft 365 E5, Microsoft Purview Suite, Microsoft 365 E5 Information Protection and Governance, or Office 365 E5. Enterprise Mobility + Security E5 includes client-side automatic labelling only. It isn't included in E3 or Business Premium.
Risk of not doing it
Older and unlabelled content stays invisible to label-based protection, and coverage depends on people remembering to label, which they don't do consistently.
Key activities
The plan in order: what must be true first, what to decide, then how to configure the pilot and production.
Prerequisites
Confirm manual labelling is stable
Automatic labelling amplifies whatever your labels already do. Before automating, check with the Maintain stage that labels are widely used, downgrade patterns are understood, and the taxonomy hasn't changed recently. If people still confuse two labels, fix that first; automation will only confuse them at a larger scale.
Check platform settings and roles
Service-side labelling of SharePoint and OneDrive files needs sensitivity labels for Office files to be turned on, and PDFs are skipped unless PDF support is on too. Auditing must be on. To see the contents of matched items in simulation, reviewers need the Data Classification Content Viewer role, which global admins don't have by default; turning a policy on needs a compliance administrator role.
Decisions
Decision: Recommend or apply labels in Office apps?
Client-side automatic labelling works while people write, based on sensitive information types or trainable classifiers. The app can recommend a label, which the person can accept or reject, or apply it. Microsoft's secure-by-default guidance suggests recommending at lower thresholds (1 to 9 matches) and applying automatically at higher ones (10 or more). Recommendations also teach people what each label is for. Client-side automatic labelling has fewer options than service-side policies: no location scoping or exceptions, and no support for PDFs or images. Exact data match can be used, but only alongside at least one other sensitive information type.
Options
- Apply automatically on any match
- Recommend at low thresholds; apply automatically at high thresholds
- No client-side automatic labelling
Recommended default: Recommend at low thresholds; apply automatically at high thresholds
Decision: Where do service-side policies start?
Service-side policies label files at rest in SharePoint and OneDrive, and email in transit in Exchange (not mail already in mailboxes; for email they label the message, not its attachments). Microsoft suggests securing the most sensitive data first, by site context as well as by sensitive information type, and running in batches of sites starting with the most sensitive. A policy can name up to 1,000 sites, and an adaptive scope can cover up to 50,000. Conditions can combine sensitive information types and trainable classifiers with whether content is shared inside or outside the organization, file extension, document name, document properties and size; Exchange rules can add sender, recipient and domain. Trainable classifiers only work with content that isn't encrypted.
Options
- All locations at once
- Most sensitive sites first, then batches by sensitivity
- Email only
Recommended default: Most sensitive sites first, then batches by sensitivity
Decision: May automatic labels replace manual ones?
By default, automatic labelling never replaces a label someone applied by hand. A service-side option can replace a lower-priority manual label, for email only or for all locations. Labels applied automatically or by a default are replaced when a higher-priority automatic label applies. Overriding people's choices is powerful but costly to trust; keep the default unless a specific risk justifies it. If you do turn on replacement for a specific risk, tell people before it starts and explain in the guide why their label may change; an unexplained change to a deliberate choice is what erodes trust fastest.
Options
- Keep manual labels
- Replace lower-priority manual labels on email only
- Replace lower-priority manual labels everywhere
Recommended default: Keep manual labels
Pilot configuration
Run every policy in simulation
A service-side policy can't label anything until it has run at least one simulation, which can take about 12 hours. Simulation isn't completely silent: it can still generate activity alerts. It shows one policy at a time, so conflicts between policies only show once they all run, and Exchange results reflect only mail sent during the run. Simulations are capped at 20 million matched files.
Review matches with content owners and tune
Review a sample of matched items with the people who own the content. Tune conditions until owners agree with what the policy finds, then rerun the simulation. Rerunning doesn't reclassify content, so use on-demand classification if you need fresh results. Turn on client-side recommendations for a pilot group in parallel and collect their feedback on wrong suggestions. Agree in advance what 'good enough' looks like: for example, owners accept at least nine out of ten sampled matches. Without a threshold, tuning can go on indefinitely.
Production configuration
Turn on in batches
Turn policies on for the first batch of sites and email, then add batches in order of sensitivity. A tenant can label up to 500,000 files a day across up to 100 policies, so large batches take time; check service desk and owner feedback before each new batch. Files that are open, checked out, or in libraries that require check-out are skipped until they're available. Labels that encrypt files in SharePoint and OneDrive through automatic labelling must use permissions the admin assigns, with no expiry, and automatic labelling doesn't apply visual markings to documents. Check both before choosing which label a policy applies.
Plan for migrations
Microsoft recommends labelling files before they move to the cloud where you can, or labelling them automatically as they are uploaded. If content moves between tenants, labels from the source tenant often aren't recognized in the target, so plan a relabelling pass with automatic labelling. Agree this with whoever runs the migration before the first batch moves.
Feed what you learned back into Plan
A condition that matches far more than expected may mean a label is too broad; a label that nothing ever matches may not be needed. Bring simulation results and false positives to the taxonomy owner with examples, so the next Plan cycle starts from evidence. Refine feeds Plan.
Make it enforceable
Service-side policies label stored files and outgoing email without waiting for people, and client-side labelling applies labels at high-confidence matches as people work. Content that is open, checked out, encrypted in ways classifiers can't read, or in formats auto-labelling doesn't support stays outside it.
Enforced by a mechanism.
Trial to adoption
Rollout approachBy risk tier · Pilot size: 1 or 2 policies on 5 to 10 high-sensitivity sites, and client-side recommendations for a pilot group
Automatic labelling acts on thousands of items at once. Starting with the most sensitive sites, tuned in simulation with their owners, puts the value where the risk is and keeps mistakes small and visible.
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
Recruit content ownersChange: Inform
Explain what automatic labelling will do on their sites and how to report a wrong label. Choose owners of the most sensitive sites first, and explain that nothing is labelled until they have seen and agreed with the simulation results.
Move on when: Owners agree and know where to report
Simulate and tuneChange: Explain
Run policies in simulation, review matches with owners, and rerun until they agree with the results. Share a sample of real matched items with each owner rather than counts, so they judge the content, not the numbers.
Move on when: Owners accept the matches in the latest simulation
Turn on for the pilotChange: Explain
Switch pilot policies on and turn on client-side recommendations for the pilot group. Watch the service desk and the owners' channels daily for the first week; most wrong labels show up within days.
Move on when: Two weeks with no unresolved wrong labels
Extend in batches by sensitivityChange: Explain
Add site batches and email policies in order, checking feedback before each. Track which batches are complete, because the daily labelling limit can spread a large batch over several days.
Move on when: Each batch settles before the next starts
Tell everyone how automatic labels workChange: Sustain
Explain why labels may appear on files and how to correct one.
Move on when: People know what to do about a wrong automatic label
Common challenges
If youTurn a policy on straight after its first simulation
What tends to happenThousands of files get the wrong label in a day, and content owners lose trust in labels.
A better approachReview matches with owners and tune before turning anything on.
If youAutomatically apply an encrypting label broadly
What tends to happenPeople and partners suddenly can't open files they used yesterday.
A better approachStart with labels that mark rather than encrypt, and add encryption only where owners agree.
If youExpect simulation to be silent
What tends to happenAlerts fire during simulation and the security team chases matches that were never acted on.
A better approachTell the security operations team before each simulation run.
If youReplace manual labels by default
What tends to happenPeople see their deliberate choices overwritten and stop labelling carefully.
A better approachKeep manual labels unless a specific risk justifies replacing lower ones.
If youLaunch large batches all at once
What tends to happenThe daily labelling limit spreads the work over days, and nobody knows which sites are done.
A better approachPlan batches with the daily limit in mind and track which batches are complete.
If youRely on simulation results from weeks ago
What tends to happenCoverage results expire 30 days after the latest run, and content has changed since.
A better approachRerun the simulation shortly before turning a policy on.
If youPoint policies at every site from the start
What tends to happenMatches arrive from thousands of sites at once and owners can't review them, so nothing gets tuned properly.
A better approachStart with a handful of high-sensitivity sites, then add batches as owners confirm the results.
Checkpoints
Pilot
- Do reviewers have the role to see matched items?
- Did every policy run in simulation, with matches reviewed by owners?
- Was security operations told before each simulation?
- Did content owners agree a sampling threshold before tuning started?
Production
- Are policies on for every planned batch?
- Are client-side recommendations and thresholds published?
- Do migrations include automatic labelling?
- Have findings gone back to the taxonomy owner?
Success measures
Capture a baseline before you start, or you won't be able to show the change.
| Measure | Where you see it | Good direction |
|---|---|---|
| Stored files with a sensitivity label | Data explorer | Up |
| Wrong automatic labels reported | Service desk and content owners | Down |
| Recommended labels people accept | Pilot group feedback | Up |
| Sensitive files with no label | Data explorer | Down |
Ready-to-use text
Explaining automatic labels
You may notice that some files now have a label you didn't apply. Microsoft 365 adds labels automatically when a file contains information such as [examples]. If a label looks wrong, change it and give a short reason; we review these reasons to improve the rules.
Simulation review checklist
Policy: [name]. Locations: [sites]. Matches: [count]. Sample reviewed: [number] items with [owner]. Wrong matches: [count and examples]. Change made: [condition or threshold]. Rerun date: [date]. Owner agrees to turn on: [yes/no, date].
Note to security operations before a simulation
From [date], we're running an auto-labelling simulation on [sites]. It doesn't label anything, but it can raise activity alerts for matched content. Please treat alerts from policy [name] during this window as expected and don't escalate them unless they show something else.
Additional reading
Plan
Configure
Changelog
Changelog (2 versions)
| Version | Date | Author | Change |
|---|---|---|---|
| 2 | Sev Derghazarian | Narrowed to automatic labelling. Sites, groups and Teams, library defaults, and DLP and Copilot now have their own playbooks in this topic. | |
| 1 | Sev Derghazarian | First version. Facts checked against Microsoft Learn on 2026-10-15. |