2.4.1. Sensitive Information Types, Sensitivity Labels, and Label Policies
💡 First Principle: You can't protect what you can't recognize. Sensitive information types recognize patterns (a card number), trainable classifiers recognize kinds of document (a contract), and sensitivity labels record a decision about how the item must be handled — so classification feeds labels, and labels feed protection.
Recognizing sensitive data:
| Classifier | How it recognizes data | Example |
|---|---|---|
| Sensitive information type (SIT) | A pattern — a primary element (regex or keyword list), supporting elements, a confidence level (high, medium, low) and proximity (how close supporting evidence must be) | Credit card number, passport number, a custom employee-ID format |
| Trainable classifier | Machine learning from example documents — pretrained or custom-trained | Resumes, source code, contracts |
| Sensitivity label | Not a detector — the result: a label applied manually, by default, or automatically | Confidential \ All Employees |
Sensitivity labels are created in the Microsoft Purview portal (Information Protection > Sensitivity labels). A label can apply encryption (who can do what with the item), content markings (headers, footers, watermarks), and — for groups and sites — container settings such as privacy and external sharing. Label priority (the order in the list) defines which labels are "higher," which matters for downgrade justification. Sublabels share their parent's priority. Label scope controls where a label can be used (files, emails, meetings, groups and sites).
Label policies publish labels to users and set behavior:
- A default label for unlabeled documents, email and meetings.
- Mandatory labeling — users must label before saving or sending. A default label takes priority over mandatory labeling.
- Justification to remove or downgrade a label.
A label does nothing until it's published in a label policy. Microsoft's default label policy publishes labels to all users and sets a General label as the default.
Auto-labeling comes in two forms:
| Client-side (label setting) | Service-side (auto-labeling policy) | |
|---|---|---|
| Runs in | Office apps, while the user works | SharePoint, OneDrive and Exchange — at rest and in transit |
| User interaction | Can recommend or apply; user can override | None |
| Simulation mode | No | Yes — a policy must complete a simulation before you turn it on |
| Scale | — | Up to 500,000 files per tenant per day; 100 policies per tenant |
Both auto-labeling forms need E5-level licensing (Microsoft 365 E5, E5 Compliance or E5 Information Protection & Governance).
Encryption and Copilot. If a label applies encryption, Copilot can return content only when the user has the EXTRACT usage right ("copy and extract content") as well as VIEW. With VIEW but not EXTRACT, Copilot won't summarize the item, though it can link to it. EXTRACT is included in the Co-Author and Co-Owner (full control) permission levels and in Outlook's Encrypt-Only option, but not in Do Not Forward.
⚠️ Exam Trap: A label that's created but never published in a label policy is invisible to users and to auto-labeling in Office. And "apply a label with View-only rights" is a valid way to stop Copilot summarizing a document — because it removes EXTRACT — whereas a label with markings only doesn't affect Copilot at all.
Reflection Question: Legal wants board papers to be readable by directors but never summarized by Copilot. Which usage right decides that, and which permission level would you assign to directors in the label's encryption settings?