Data classification is the act of deciding how sensitive each piece of information is, writing that decision down, and attaching rules to it. It sounds bureaucratic and becomes useless when it is: a scheme nobody follows is worse than no scheme, because it creates the illusion that data is controlled. This guide is the minimum version that actually works.
Why bother
- Incident scope. When something leaks, the first question is "what was in there?" Classification answers it in minutes instead of weeks.
- Cost. Not everything needs the same encryption, backup tier or access review. Classification stops you paying enterprise prices for public marketing copy.
- Obligation. Privacy law, contracts and industry standards all ask you to know what personal and confidential data you hold and where it lives.
Four levels are enough
| Level | Test | Typical examples |
|---|---|---|
| Public | Could be published on the website with no harm. | Marketing copy, public documentation, published prices. |
| Internal | Harmless outside the company in the short term, but not intended for release. | Meeting notes, internal wiki, staff directory, non-sensitive project plans. |
| Confidential | Would cause real damage to people, customers or the business if disclosed. | Customer contracts, source code, financial reports before publication, personal data of employees. |
| Restricted | Regulated, contractual, or catastrophic if disclosed. Access is need-to-know and logged. | Health records, payment data, credentials and keys, security incident details, merger documents. |
More levels than four is usually a taxonomy problem disguised as a policy. If two levels are always treated the same way in practice, they are one level.
The criteria behind the decision
Classify by the harm of disclosure first, then check the other three:
- Harm. Who is hurt, and how badly, if this is public tomorrow?
- Legal or contractual duty. Does a regulation, standard or customer contract name this data type? GDPR personal data, PCI card data and HIPAA records are effectively pre-classified.
- Aggregation. Individually harmless records become confidential in bulk. A single employee's salary is confidential; the whole payroll file is restricted.
- Time. Most classifications expire. Financial results are restricted until published, then public. Build the expiry into the label.
Make the label mean something
A label with no handling rule is decoration. For each level, define storage, sharing, retention and deletion:
| Storage | Sharing | Access review | |
|---|---|---|---|
| Public | Any | Open | None |
| Internal | Company systems | Staff with a work reason | Annual |
| Confidential | Approved systems, encrypted | Named individuals, no external forwarding | Quarterly |
| Restricted | Approved systems, encrypted, access logged | Named individuals, approved per case | Monthly |
Common mistakes
- Classifying everything restricted. It is the safe answer and it destroys the point of classifying. If everything is critical, triage becomes impossible.
- Classifying the container, not the content. A "public" shared drive containing one contract is not public.
- No owner. Every classified dataset needs one named person who decides who gets access.
- No tooling. If applying a label takes three manual steps, nobody will apply it. Labels must attach where the data is created — the document editor, the database column, the bucket policy.
- No review. Data ages. A project folder from 2019 still marked confidential is probably archive material, and treating it as live costs attention.
A rollout that survives contact with reality
- Pick the four levels and their handling rules. One page, not a policy document.
- Inventory where data actually lives: systems, shared drives, SaaS tools, databases.
- Pilot on one dataset with a real owner — payroll, customer support tickets or the code repository.
- Label it, apply the rules, and measure what breaks. Something will.
- Publish the scheme, then extend dataset by dataset. Do not wait for a perfect inventory.
- Review quarterly, and delete the classification of anything nobody has opened in a year.
GDPR and data classification
GDPR does not publish an official classification scheme — but it pre-classifies parts of your inventory for you, and your labels have to agree:
- Personal data (anything identifying a living person) starts at Confidential in most schemes: employee records, customer contacts, support tickets with names in them.
- Special categories (Article 9: health, biometrics, politics, religion, sexuality, trade-union membership) and criminal-offence data (Article 10) belong in Restricted, full stop. Processing them needs an Article 9 condition, not just a label.
- Pseudonymized data is still personal data. The label can drop one level only if re-identification keys are held separately and controlled as Restricted.
- Truly anonymous data leaves GDPR’s scope — and reaching "truly anonymous" is harder than stripping names: free text, IP addresses and small cohorts re-identify people.
Two documents follow from the labels: the records of processing (Article 30) become easy when every entry carries a classification, and a breach notification decision (72 hours, Article 33) becomes a lookup instead of a debate — if a Restricted dataset leaked, you notify.
Frequently asked
What are the four data classification levels?
Public, Internal, Confidential, Restricted. Public is safe to publish; Internal is routine business information; Confidential causes real damage if disclosed; Restricted is regulated or catastrophic and gets need-to-know access with logging. The handling rules matter more than the names.
How does GDPR affect data classification?
GDPR pre-classifies: personal data is effectively Confidential by default, and special-category data (health, biometrics, beliefs) is Restricted. Your scheme must also survive the two GDPR tests — can you list where each class lives (Article 30), and can you decide within 72 hours whether a leak is notifiable (Article 33).
How do you classify sensitive data?
Start from the harm of disclosure, then check legal duty, aggregation and time. Sensitive data (credentials, health records, payment data, security incident details) is Restricted: encrypted at rest, access named and logged, reviewed monthly, and deleted on schedule.
Related: zero trust architecture explains why the label has to follow the data into every request, and after a data breach covers what to do when the classification was wrong.