# Presidio for European Data: euRedact Compared

> MS Presidio and euRedact measured on the same documents: a published five-language benchmark of names, health data and offences, Presidio's own synthetic set under both definitions of personal data, and 152,300 documents of European identifiers. Where each tool wins, with dates and sources.

Markdown version of https://euredact.dev/compare/presidio, generated from that page at build on 2026-11-19. Every page of the site is listed with its Markdown version at https://euredact.dev/llms.txt; the full SDK reference is https://euredact.dev/llms-full.txt.

*Comparison*

# Presidio for European Data

MS Presidio and euRedact on the same documents, measured three times: a published benchmark of names, health data and offences in five languages, Presidio's own evaluation set under both definitions of personal data, and 152,300 documents of European identifiers. What each tool is for, where Presidio wins, and where configuration cannot help it.

All three measurements are ours, on synthetic documents; two of them run Presidio as shipped as well as configured for Europe. The benchmark figures are read from the same export as [the cloud benchmark page](/benchmarks/cloud) (Presidio measured 2 November 2026, euRedact 14 November). The identifier figures are from [the rules engine’s comparison](https://github.com/euRedact/euRedact/blob/v1.0.0/docs/presidio-comparison.md) of 31 July 2026 for Presidio’s rows, and from euRedact 1.0.0’s corpus results of 14 November 2026 for euRedact’s; the per-type figures are on [the accuracy page](/benchmarks).

## Two Different Jobs

**Presidio** is a general-purpose, extensible PII framework: pattern recognizers for formats that are the same worldwide (IBAN, credit card, e-mail, phone, IP and MAC address), named-entity recognition for names and places through spaCy, and the means to add recognizers of your own. It ships national-identifier recognizers for 7 of the 31 countries euRedact covers (DE, ES, IT, PL, FI, SE and the UK).

**euRedact Core** is a rules engine for the 31 European jurisdictions: 29 entity types, with a checksum wherever the format has one, on your machine, with no dependencies. **euRedact Cloud** adds a model for what has no shape: names, addresses, job titles, and the Article 9 and 10 data (health, religion, political opinion, offences) in Dutch, English, French, German and Polish.

Which tool is the better one depends on which of those two jobs your documents are. The two measurements below cover one each.

## Names, Health Data, Offences

The published benchmark: 27,831 annotated items in 900 synthetic IT, HR and medical documents, five languages. Presidio ran with EU recognizers and term lists, configured in six rounds against 65 development documents from the benchmark’s earlier sets, none of them among the 900. “Found” counts an item when any part of it was redacted; “left readable” is the share with less than 75% of its letters and digits hidden.

*All 900 documents*

*euRedact Cloud*

*MS Presidio*

| --- | --- | --- |
| F2 | 0.921 | 0.656 |
| Found | 94.2% | 63.2% |
| Precision | 84.8% | 77.2% |
| Left readable | 7.3% | 44.6% |
| Art. 9/10 masked | 78.5% | 14.1% |

*Per category*

*Items*

*euRedact found*

*Presidio found*

*euRedact left readable*

*Presidio left readable*

| --- | --- | --- | --- | --- | --- |
| Article 9 and 10 data | 6,812 | 84.0% | 41.6% | 21.5% | 85.9% |
| Names | 7,917 | 98.9% | 90.1% | 1.2% | 10.7% |
| Occupation | 3,244 | 96.9% | 10.2% | 3.6% | 95.2% |
| Places | 2,865 | 98.2% | 99.0% | 1.9% | 1.6% |
| Contact details | 1,310 | 100.0% | 97.6% | 0.1% | 2.5% |
| Government identifiers | 773 | 99.0% | 82.1% | 1.4% | 19.0% |
| Financial | 2,264 | 96.5% | 16.3% | 3.7% | 84.5% |
| Network identifiers | 905 | 100.0% | 99.4% | 0.1% | 1.1% |
| Other identifiers | 1,263 | 84.2% | 68.4% | 16.0% | 32.6% |
| Credentials | 155 | 99.4% | 71.6% | 7.1% | 32.3% |
| Dates | 323 | 100.0% | 93.2% | 0.6% | 6.8% |

**Where Presidio holds its own.** Places, contact details and network identifiers: on those categories the two are within a few points, and on places Presidio finds slightly more. Its person names are spaCy’s and reach 93.6%; with usernames pooled in, 90.1%, against euRedact’s 98.9%.

**Where it does not.** The Article 9 and 10 data: term lists in five languages, written for the benchmark, still leave 85.9% of those items readable, because a diagnosis or an offence is rarely a term from a list. That is the gap a model closes and a pattern cannot, and it is where euRedact Cloud’s lead is widest. Occupations and financial amounts entered the benchmark’s guideline after Presidio’s set-up was frozen, so it was never asked for them; its 10.2% and 16.3% there measure that, not its term lists.

Every type, set and language, and the other three systems measured, are on [the cloud benchmark page](/benchmarks/cloud); the method, the 60-document sample and the harness are in [the benchmark repository](https://github.com/euredact/benchmark-2026-11-14).

## On Presidio’s Own Set

Presidio’s synthetic evaluation set: 1,500 English sentences generated from presidio-research’s templates, with US identifiers. Scored twice on 16 November 2026: against Presidio’s own labels (2,863 spans of 17 types: organisations, every date and time, ages, countries and nationalities included), and against the same texts regolded under euRedact’s guideline (2,402 spans: what identifies or describes a natural person). A small set beside the five-language benchmark: one short sentence per document and 2,863 spans, against 27,831 annotated items in 900 full documents there. Read it as a second opinion on definitions, not as the main measurement.

*System*

*Presidio’s labels*

*euRedact’s guideline*

| --- | --- | --- |

*F2*

*Left readable*

*F2*

*Left readable*

| --- | --- | --- | --- |
| MS Presidio 2.2.366, as shipped | 0.792 | 32.0% | 0.606 | 46.8% |
| MS Presidio 2.2.366, our EU configuration | 0.740 | 38.3% | 0.631 | 46.0% |
| euRedact 1.0.0 Cloud | 0.848 | 20.0% | 0.915 | 10.3% |

**Home ground.** The set was generated with Presidio’s tooling for Presidio’s types, and its recognizers were written for them. euRedact Cloud still leaves less readable on Presidio’s own labels (20.0% against 32.0%), by a smaller margin than on the European sets.

**Different definitions.** 776 of Presidio’s 2,863 spans (27%) are not scored under euRedact’s guideline. 508 of them are nothing it tags: countries and regions, years and timestamps, ages, honorifics, company websites; a country is too coarse to point to a person. euRedact Cloud leaves those alone by design and loses recall for each of them on Presidio’s labels. The other 268 are organisations, which the guideline marks but does not count as personal data; euRedact Cloud does redact them, and found 76% of them here to Presidio’s 80%. On the regold, where only personal data counts, euRedact’s lead widens (0.915 to 0.631).

**Our EU configuration costs Presidio here** (0.740 against 0.792 on its own labels): it was built for European documents, with the US recognizers swapped out, and the set is American. Under euRedact’s guideline it helps a little, from the term lists for Article 9 and 10 data.

**Conditions.** One run per system on a small sample, 16 November 2026: presidio-analyzer 2.2.366 with its predefined recognizers and spaCy’s large English model, and the same analyzer under the configuration the five-language benchmark uses; the public SDK 1.0.0 against the production service. The regold was annotated under euRedact’s guideline by language models in two blind passes with every disagreement adjudicated, so on that gold euRedact’s definitions are the yardstick, as Presidio’s are on its labels. The set is [public](https://github.com/data-privacy-stack/presidio-research) (data/synth\_dataset\_v2.json at commit c46963a).

## European Identifiers

The rules engine’s own corpus: 152,300 synthetic documents generated per country from real formats, 667,268 labelled identifiers across 31 countries and 24 languages, both engines scored by the same rule. Presidio was measured on it on 31 July 2026, as shipped and again with the recognizers for individual countries it ships but does not load (21 of them), phone regions for all 31 countries, credit cards in three more languages and per-language analysis. euRedact’s rows are release 1.0.0 of 14 November 2026 on the same corpus, under the scoring rule in force since 0.7.0, which no longer counts a correct detection of a repeated value as a false positive: recall is comparable across the two dates, precision favours the newer run slightly. Names and places are not labelled in this corpus, so they count for neither side.

*Engine*

*Measured*

*Precision*

*Recall*

*Identifiers missed*

| --- | --- | --- | --- | --- |
| Presidio 2.2.364, as shipped | 31 July 2026 | 99.91% | 57.68% | 282,311 |
| Presidio 2.2.364, configured for the EU | 31 July 2026 | 93.15% | 68.29% | 211,552 |
| euRedact 1.0.0, no country declared | 14 November 2026 | 99.8% | 99.7% | 2,277 |
| euRedact 1.0.0, with countries declared | 14 November 2026 | 99.9% | 99.8% | 1,352 |

**Out of the box, Presidio is the more precise engine:** 99.91% to euRedact 1.0.0’s 99.9%. What it claims, it is almost always right about; it simply does not claim most European identifiers. Once its own recognizers are registered, on the formats that are the same worldwide (IBAN, credit card, e-mail, phone, IP and MAC address, health IDs) it is at or near 100%, as a specialist engine is; as shipped, credit cards are at 18.2% and phone numbers at 68.4%.

**The gap is structural, not a matter of tuning.** Registering what Presidio already ships brings credit cards from 18.2% to all 6,415 in the corpus and phone numbers from 68.4% to 97.9%. It cannot recover national identifiers, postal codes or licence plates where it has no recognizer, which is 24 of the 31 countries, and it ships none for VINs, UUIDs, secrets, BICs, IMEIs, social handles, passports, GPS coordinates or generic tax identifiers. Presidio is a framework: a team that wrote recognizers for the missing 24 would close most of this gap, and that is the intended way to use it. This measures day one.

**What the comparison does not say.** The corpus is euRedact’s own, built to exercise European formats, which favours euRedact’s absolute numbers. Names and places were not tested, and on text where the personal data is names rather than identifiers this comparison predicts nothing. Presidio detects cryptocurrency addresses; euRedact does not. The full tables per type, per country and per language, the configuration and the limitations are in [the document](https://github.com/euRedact/euRedact/blob/v1.0.0/docs/presidio-comparison.md).

## Which to Use

*Presidio, when*

- the personal data is mostly names and places in a language spaCy has a model for and euRedact Cloud does not (on the five it covers, euRedact found 98.9% of names to Presidio’s 90.1%), or you run the rules engine alone, which does not attempt names;
- you want a framework to extend with your own recognizers;
- your identifiers are worldwide formats, or from the seven countries it covers;
- you need cryptocurrency addresses.

*euRedact, when*

- the documents carry member-state identifiers from any of the 31 countries, with checksums;
- the documents carry health, religion, political opinion or offences in the five languages: euRedact masks 78% of the Article 9 and 10 items on the benchmark, Presidio 14%;
- the rules must run on your machine with no dependencies, and the model only in the EU;
- you want the measurements, their weak spots included, published with the release.

The two can also run in sequence: Presidio’s NER for names where its languages reach, euRedact Core for the identifiers. With the euRedact SDK, the identifiers the rules find are masked on your machine before anything is sent anywhere.

## Measure It on Your Documents

The benchmark’s harness runs Presidio and euRedact on the same documents; the published sample runs on the free plan with an API key. Or see a redaction first in the live demo, on a sample or text you made up, no account needed.

[Try the live demo](/demo)[The sample and harness on GitHub →](https://github.com/euredact/benchmark-2026-11-14)
