Blog · Comparison & AI · · 2 min read

Presidio or Safe-Doc: why 'almost' is not enough

Presidio is a good project. Let us say it plainly: Microsoft's open-source tool does honest work, and plenty of technical teams lean on it without ever regretting the choice. But anonymization hides a trap. Detecting "almost everything" is not protecting. One missed value, and the person becomes identifiable again. And Presidio, used raw, always lets at least one slip through, exactly as the real case reproduced just below makes clear.

Raw Presidioleaks it
Company : NovaTech
SIREN : 552 100 554
SIRET : 552 100 554 00012
Docusign ID : 6A9B815A-581E-42D2…
Capital : €2,500,000
Safe-Docmasked
Company : [ORGANISATION_1]
SIREN : [SIREN_1]
SIRET : [SIRET_1]
Docusign ID : [REFERENCE_1]
Capital : [AMOUNT_1]

Real case from our comparator, on company bylaws. Presidio even detects more entities overall (34 vs 31), but leaves the company name, the SIREN, the SIRET, the Docusign ID and the capital. Yet the SIREN or the Docusign ID alone is enough to trace back to the company and the document. Detecting more is not protecting better.

Detecting more is not protecting better

Look at the numbers. In this real example, Presidio even finds more entities than Safe-Doc (34 vs 31). Impressive? Not really. Because it leaves the SIREN, the SIRET and the Docusign ID, three markers that on their own are enough to trace back to the company and to the exact document. Anonymization is all or nothing. The total means nothing. What matters is the value that slips through. And one is enough.

Why Safe-Doc goes further: months of R&D

Safe-Doc is not a bundle of rules. It is an engine. An engine built over months, tuned on French and English corpora, on dozens of data types (IBAN, national IDs, internal references, compound names, domain abbreviations), and above all measured continuously by benchmarks that hunt down exactly the cases generic tools let slip. That work stays invisible. Yet it is what separates "almost" from "reliable".

The comparison

CriterionPresidio (raw)Safe-Doc
FR and EN detectionGeneric, to configureTuned FR + EN engine
Hard cases (IBAN in a table, compound names, abbreviations)Often missedCovered (dedicated R&D)
Re-identification (key kept)To buildBuilt in
PDF, Word, Excel, scansOn youHandled (OCR)
Quality measurementTest it yourselfContinuous benchmarks
Ready to useNo (dev project)Yes

To be fair: Presidio is free, open-source, flexible. For a technical team that wants to build and configure everything itself, line by line, it is a genuinely useful brick. But raw, on real documents, "almost" gets expensive. One leak is enough.

In short

Presidio is a starting point. A good one, for developers ready to build. Safe-Doc is the choice for those who cannot afford a single leak, in French and in English, and without standing up any engineering project at all. Read next: the ChatGPT and GDPR at work guide.

Zero leak, not "almost". Grab a real document. Test Safe-Doc on it, now.

Part of the guide : Comparisons ↗