You've probably heard this during a processor call or contract negotiation: "We strip the names and assign a code, so the data is anonymous." It sounds reasonable, but it's the belief that just cost IQVIA Solutions Italy €7 million.
Misunderstanding Anonymity
Many compliance teams view de-identification as a spectrum where removing direct identifiers moves data closer to being "out of scope." By replacing names with codes and dropping tax IDs, while keeping clinical details, they assume the dataset is no longer personal data.
This belief has some basis. Recital 26 of the GDPR states the regulation doesn't apply to anonymous information and frames identifiability around "means reasonably likely to be used." If re-identification is difficult, the thinking goes, the data is effectively anonymous.
Why This View Falls Short
The problem is stopping there. Recital 26 also clarifies that pseudonymised data, which could be linked to a person with additional information, remains personal data. Article 4(5) defines pseudonymisation as a safeguard, not an escape from regulation.
In the IQVIA case, the Garante focused on what the codes enabled. Each patient had a persistent code linking records over time, alongside birth year, sex, diagnoses, symptoms, prescriptions, medical tests, vaccinations, and location data. The authority concluded that this combination allowed patient re-identification using reasonably available means. Swapping identifiers for codes didn't change that.
A persistent code isn't a barrier to identification; it's a key that builds a richer profile with each new record, making re-identification easier. The feature that aids research also undermines anonymity claims.
Lessons from the Case
The case offers several lessons beyond the headline finding.
Anonymity Wasn't Verified Against Actual Data. The database held direct identifiers for over 3,300 patients, including names and tax codes, next to health data. This arose from free-text fields transferred from doctors' systems. Assuming structured fields are clean misses identifiers typed into notes. An unverified anonymity claim is just an assumption.
Controllership Begins at Collection. The Garante held IQVIA responsible as a controller from the point of data collection from about 800 general practitioners. You can't argue that anonymisation happens "later" and that obligations start only once the data is clean. If you're processing data to make it identifiable or not, you need a lawful basis, and for health data, an Article 9(2) condition.
Other Obligations Followed the Anonymity Claim. Once the data was personal, other failures were evident. The authority found no lawful basis for processing health data, inadequate patient information, undefined retention periods, no data protection impact assessment, and inadequate technical and organizational measures. The anonymity claim was foundational, and everything else depended on it.
Steps to Take
Treat "anonymous" as a conclusion to prove, not a label to apply.
Assume Coded Data Is Personal Data. Run compliance on this basis unless you can demonstrate otherwise. Your DPO should sign off on anonymity claims, not the team benefiting from it.
Test for Singling Out, Linkability, and Inference. Examine your dataset like a motivated outsider. Can a patient be isolated by birth year, sex, diagnosis, and location? Can records be linked over time through a persistent identifier? If so, you're holding pseudonymised data.
Scan Free Text First. Search narrative fields for names, tax codes, addresses, and phone numbers. Don't assume integration layers filter them.
Complete a DPIA Before Processing. Large-scale health data processing falls under Article 35(3)(b). The DPIA documents re-identification analysis and retention logic. Skipping it removes essential controls.
Establish Lawful Basis and Article 9 Condition Up Front, and incorporate retention periods into the design.
Consider Anonymisation at the Source. The Garante suggested doctors anonymise information before supplying it. Moving the operation upstream changes your risk profile. Be cautious: if you specify and receive the output, you may still determine purposes and means.
When Conventional Wisdom Holds
Pseudonymisation isn't worthless. Article 25 and Article 32 name it as an appropriate measure. It reduces risk and supports compatibility arguments. Using codes for internal analytics, with the key held separately under strict access controls, is sound practice.
Genuine anonymisation is achievable. Properly aggregated outputs, with sufficient cell sizes and no persistent identifiers, can fall outside the GDPR. The pharmaceutical customers in this case received aggregated research reports, and aggregation at the output stage is legitimate.
The error isn't using these tools; it's mislabeling the input dataset as anonymous because the output is. Aggregated reports can be anonymous while the database behind them is not.
If your research pipeline relies on an anonymity assumption, test it this quarter, not after an inspection.




