Your Microsoft 365 Sensitivity Labels Aren't Doing Anything. Here's Why.

Open the Microsoft Purview compliance portal in any tenant that's been around more than a year. Click into Information Protection. You'll see sensitivity labels. You'll see policies. You'll see dashboards. Then look at your actual SharePoint sites, OneDrive folders, and email. Almost nothing is labeled.

This is the most common state of data classification in Microsoft 365: configured, published, and not classifying anything. The problem isn't the technology. The problem is what happens between "we have labels" and "the labels mean something."

Why this is suddenly a real problem

Three things changed at once, and most companies are still catching up.

First, cyber liability insurance carriers started auditing classification posture before renewal. "We use Microsoft 365" is no longer a checkbox. They want to see which data is labeled, what the labels enforce, and whether the policies match what's in the policy document. Unlabeled sensitive data on a renewal questionnaire is a premium increase or a non-renewal.

Second, AI assistants now read everything they can reach. Copilot, third-party agents, and the indexing layers underneath them surface data based on permissions, not sensitivity. A SharePoint folder marked "internal" with no label is fair game. The classification gap that used to mean "we should fix this" now means "the AI is going to summarize our acquisition memo into someone's chat window."

Third, the shift to remote and hybrid work scattered the documents that used to live on a file server with one set of NTFS permissions. They live in OneDrive, Teams, and external sharing links now. Without labels, you can't tell what should never have left the building.

The Microsoft 365 stack already solves this. It just doesn't solve itself.

How the layers actually stack

Data classification in Microsoft 365 isn't one feature. It's four layers that have to be configured together to do anything useful.

1. Sensitivity labels are the taxonomy. Public, Internal, Confidential, Restricted -- whatever your organization actually means by those words, written into Purview as labels with a defined order of escalation. Labels are metadata that travels with the file. They survive being downloaded, emailed, copied to a thumb drive, and renamed. Without labels, the rest of the stack has nothing to act on.

2. Auto-labeling policies are how labels get applied at scale. You define conditions ("contains a credit card number," "matches the pattern of an internal employee ID," "originates in the Legal SharePoint site") and Purview applies the corresponding label automatically. Manual labeling exists too, and matters for edge cases, but a deployment that depends on users remembering to click a label is a deployment that ships unlabeled data.

3. Data Loss Prevention (DLP) policies are the enforcement layer. They watch for sensitive content moving through email, Teams, OneDrive, and SharePoint, then block, warn, or audit based on conditions. A DLP policy can target a specific label ("don't let anything labeled Confidential leave the tenant via email") or content patterns directly. The label-aware version is dramatically more powerful, because the label encodes intent that pattern matching can't.

4. Retention labels and policies decide how long content lives and what happens when it expires. They're a separate label type from sensitivity labels, but they share the same architectural idea: metadata applied at scale, enforced by the platform. A document labeled for seven-year retention can't be deleted by an end user even if they want to.

These four layers are independent. You can configure sensitivity labels without auto-labeling. You can run DLP without ever touching sensitivity labels. You can have retention without classification. Most tenants do exactly that, in some incoherent combination, and then wonder why the audit found unlabeled tax documents in someone's personal OneDrive.

Where deployments break

Three failure modes show up in nearly every tenant I've audited.

The taxonomy is borrowed. Someone copied a generic three-tier scheme from a Microsoft Learn article and never adapted it. Public, Internal, Confidential. Nobody at the company knows the difference between Internal and Confidential. Files get labeled inconsistently, or not at all, because the labels don't map to anything the business actually does. The fix is upstream of Purview: write the data classification policy first, then build the labels to enforce it.

Auto-labeling is off. Manual labeling shipped, end users got a one-time training email, and three months later 4% of files have a label. Auto-labeling requires E5 (or the Information Protection and Governance add-on for E3 tenants), and it's the difference between classification working and classification existing. Without it, your label coverage will never improve, because asking users to classify their own data correctly at scale is asking them to do a job nobody hired them for.

No DLP. Or DLP without labels. Labels with no enforcement are a filing system. DLP that doesn't reference labels is pattern matching with all the precision and false-positive rate of a 2008 spam filter. The combination is what matters: labels carry intent, DLP enforces against intent. Either alone is a dashboard.

There's a special circle of IT hell reserved for tenants that configure all four layers and then exempt the executive team from every policy. The labels exist. The enforcement exists. The data the auditor cares about most lives in the one mailbox the rules don't apply to. Don't do this.

The trade-offs that matter

Most of the configuration choices in Purview are genuine trade-offs, not best practices. Pick the one that matches your operating model.

Manual vs. auto-labeling. Auto-labeling is more accurate at scale and survives turnover. Manual labeling preserves user judgment for edge cases and works without the E5 license. Most healthy deployments run both: auto for the high-volume patterns (PII, financial data, regulated identifiers) and manual for the long tail. Pure manual is a coverage problem. Pure auto is a false-positive problem.

Restrictive vs. permissive defaults. A new SharePoint site can default to "Internal" or "no label." Defaulting to Internal gets you coverage immediately but mislabels public-facing content, which then trips DLP rules and creates noise. Defaulting to "no label" keeps the noise down but leaves a coverage gap that has to be closed by auto-labeling or it never closes. Pick one and instrument it.

E5 vs. E3 plus add-on. Full Information Protection is included in E5/A5/G5. E3 tenants can buy the Information Protection and Governance add-on, which gets you most of the same capability at lower per-user cost. Run the math against your actual headcount. The break-even moves around, but for tenants under about 200 users the add-on usually wins. Above that, E5 starts pulling ahead because of the things bundled around it (Defender, Audit Premium, Insider Risk Management).

Skai would point out that the answer to most of these trade-offs is "ask the people who own the data what protection actually means to them, then make the technology match." She's usually right.

What to check today

Open Purview, go to Information Protection, click Labels, and look at the Activity Explorer. If the chart is flat or shows fewer than 100 label applications a week in a tenant of 50+ users, the deployment isn't classifying anything. That's the test. Ten minutes, no consultants required.

Everything downstream -- DLP rules, retention enforcement, audit posture, the conversation with your insurance underwriter at renewal -- depends on labels actually getting applied. If they aren't, you have configured tools, not protection. The two look identical until the day they don't.