Find where patient-identifiable data actually lives, classify it automatically, and apply protection that travels with the document.
You cannot protect data you cannot locate. This guide covers the discovery exercise that establishes where PHI sits across your estate, and the labelling scheme that keeps it protected afterwards. Expect the discovery findings to be uncomfortable and useful in roughly equal measure.
Sensitive information type selection, auto-labelling policy configuration in simulation mode, label taxonomy design, and the sequencing that avoids breaking clinical workflows on day one.
Healthcare estates accumulate patient-identifiable data in places nobody planned. Extracts pulled for a project, spreadsheets built for an audit, correspondence attached to a case — each was created for a reason and none was catalogued.
The operational risk is the obvious one. The immediate one, for most providers, is that AI assistance across the estate is now a board-level topic, and it cannot proceed responsibly until you know what is where and who can reach it.
Check these before beginning. Most stalled implementations stall on one of them.
Microsoft 365 E5, or E3 with the Compliance or Information Protection add-on. Auto-labelling for SharePoint and OneDrive requires the E5 tier.
Compliance Administrator or Information Protection Administrator in Purview, plus a privacy officer who can approve the classification scheme.
An agreed data classification policy. If one does not exist, write it first — a labelling scheme without a policy behind it is arbitrary.
A decision about which workloads are in scope initially. We recommend SharePoint and OneDrive before Exchange and endpoints.
Configuration is straightforward once these are clear. Skipping them is why most first attempts produce something that works and cannot be maintained.
Purview ships with built-in types for health record numbers, insurance identifiers, national identity numbers and similar patterns. These detect content; they do not protect it. Detection accuracy depends on confidence level and instance count, both of which you tune.
A label can apply encryption, restrict access, add visual markings and control whether content can leave the tenant. Crucially, the protection persists with the file, so a labelled document forwarded outside your organization remains protected.
Auto-labelling policies run in simulation before enforcement. This tells you what would have been labelled without labelling anything, which is how you discover that a well-intentioned rule would have encrypted every clinical letter in the organization.
Settings shown are the ones that matter, not every field on the form. Values are starting points to validate against your own environment.
Four or five labels. More than that and misclassification becomes routine, which is worse than no classification because it produces false confidence.
A workable healthcare scheme is General, Internal, Confidential, and Confidential — PHI, with sublabels only where protection genuinely differs.
Content Explorer and Activity Explorer show what already matches your sensitive information types across the estate. Do this before creating labels, because the results usually change the taxonomy you were about to build.
Pay particular attention to volume by location. Most PHI concentrates in a small number of sites, and knowing which ones lets you sequence remediation.
Build the rule, scope it to a subset of sites rather than the whole tenant, and run it in simulation for at least a week. Review the matches manually — a sample of a few hundred is enough to see the pattern.
Expect the first simulation to be wrong. That is what it is for.
Labelling protected content that is already shared with everyone in the organization does not un-share it. Run the oversharing review in parallel — broad-access links, broken inheritance, sites without owners — and fix the highest-risk findings first.
This is the step that most often extends the timeline, and it is the one that actually reduces risk.
Turn the policy on for the piloted sites. Watch the support queue for a fortnight. Then extend site by site rather than tenant-wide.
Add Exchange and endpoint DLP afterwards, as separate exercises with their own simulation periods.
Every one of these is avoidable, and every one of them is common enough that we check for it by default.
Every additional label multiplies misclassification. A fifteen-label scheme designed by committee will be applied incorrectly and will produce confident, wrong protection decisions.
Encryption is the label capability that breaks things. Clinical systems, PDF processors and third-party viewers all handle protected files differently. Test with your actual application set before enforcing.
It does not. A protected document shared organization-wide is still reachable by everyone in the organization. These are two separate remediation exercises that need to run together.
Content keeps arriving. Set a recurring review cadence, or the position you established will drift within a year.
We run scoped discovery against a single workload as a fixed-price engagement, and the findings report is yours whether or not you take the remediation work further with us.
Describe the situation in your own words.
