Summary
- A few years ago, working on the DLP of a large healthcare company, we could not tell a patient called Dolores Fuertes apart from a clinical note saying «dolores fuertes», Spanish for “severe pain”. With patterns and dictionaries it was impossible.
- Today’s language models classify each word according to the sentence it appears in. On MEDDOCAN, a benchmark of Spanish clinical cases, the best systems scored an F1 above 0.97. Exact string matching reached 0.43.
- Several DLP vendors now advertise classification with language models. The accuracy figures they publish are their own.
- As of 20 November 2026, Microsoft Purview does not use context to solve this case in Spanish. Its name detection relies on dictionaries. Its medical-terms detection and its healthcare trainable classifiers work in English only. Purview’s generative AI acts after detection: it prioritises alerts and investigates incidents.
- What does solve it in Purview is Exact Data Match, which compares content against the real patient list. That is not AI: it is an exact lookup over hashed values.
Status as of 20 November 2026.
The case: Dolores Fuertes
A few years ago I worked on the DLP of a large healthcare company. What most needed protecting was patient data: a name next to a diagnosis is health data, one of the special categories in Article 9 of the GDPR.
One of the biggest challenges I came across had a proper name. Dolores Fuertes can be a patient. «Dolores fuertes» is also what a doctor writes in any report: «refiere dolores fuertes en la zona lumbar», “reports severe pain in the lower back”. To the DLP engine both strings were the same, and there was no way to separate them. In a healthcare DLP that is critical, because the rule meant to detect patient names fired on the description of a symptom.
Nor was it an isolated case. Many Spanish first names are also ordinary words, and some appear daily in a hospital: Dolores (pains), Consuelo (comfort), Remedios (remedies), Amparo (shelter), Socorro (help). Many surnames are adjectives: Fuertes (strong), Delgado (thin), Calvo (bald), Moreno (dark-haired), Blanco (white). Capital letters do not help: a sentence starts with one, some reports are written entirely in capitals, and text extracted from a scanned document arrives however it arrives.
English has its own versions of the problem, with names such as Grace, Hope or Faith, but in Spanish clinical text the collision is far more frequent.
Why rules could not solve it
A classic DLP detects with patterns. In Microsoft Purview, a sensitive information type combines a primary element (a regular expression, a keyword list or a dictionary) with supporting elements that must appear nearby to raise confidence. It works very well for anything with a shape: a national ID number, an IBAN or a card number with its check digit.
A name has no shape. The only way to look for one with rules is a dictionary of first names and surnames, and the dictionary does not understand the sentence. That leaves two options, both bad:
- A sensitive rule fires on every report that mentions severe pain. Alerts run into the hundreds, nobody reviews them and the policy stays in audit mode indefinitely.
- A strict rule, which excludes the conflicting names or demands much more context, lets through exactly what needed protecting.
What was left was manual tuning: exceptions, thresholds and supporting keywords, revised over and over. A lot of work for a balance that broke with the next type of document.
What has changed: models that read the sentence
Named entity recognition (NER) models based on neural networks, and later transformers and large language models, do not look for a string. They classify each word based on the words around it. To such a model, «Dolores Fuertes, 54 años, ingresa por…» (“Dolores Fuertes, aged 54, admitted with…”) and «refiere dolores fuertes en el hemitórax izquierdo» (“reports severe pain in the left hemithorax”) are different sentences with different structures.
| Sentence (English gloss) | Name dictionary | Model that reads context |
|---|---|---|
| «Dolores Fuertes, 54 años, ingresa por dolor torácico» (Dolores Fuertes, 54, admitted with chest pain) | Name | Person |
| «Refiere dolores fuertes en el hemitórax izquierdo» (reports severe pain in the left hemithorax) | Name | Symptom |
| «Firmado: Dra. Consuelo Blanco» (signed: Dr Consuelo Blanco) | Name | Person (healthcare professional) |
| «Le sirve de poco consuelo» (it is little comfort to her) | Name (if case-insensitive) | Ordinary word |
The table shows what to expect from a well-trained model; no study I know of has tested this particular pair of sentences. What has been measured is name detection in Spanish clinical text. In 2019, the MEDDOCAN shared task (IberLEF) had 18 teams compete at de-identifying 1,000 synthetic Spanish clinical cases. The best system scored an F1 of 0.970 at identifying and classifying entities, and 1.000 on patient names. The exact string-matching baseline, the closest thing to a dictionary, reached 0.427. The winner used recurrent networks. A year later, a general-purpose multilingual BERT model reached 0.967 with no adaptation to the clinical domain.
Those figures need reading with care: the corpus is synthetic, and the organisers themselves warn that results are not directly comparable to real clinical records. Even so, the gap between 0.43 and 0.97 explains why what was a huge challenge a few years ago now has a technical solution.
How the tools apply it
- Presidio, the open-source project created by Microsoft and now run by an independent community, combines regular expressions, NER (with spaCy, Stanza or transformer models) and context words that raise a finding’s score. It supports Spanish models and can return why it flagged something: which recogniser detected it and which context word raised the confidence.
- Google Cloud Sensitive Data Protection has types for person names (PERSON_NAME) and medical terms (MEDICAL_TERM), and hotword rules that raise or lower a finding’s likelihood when certain words appear nearby. It does not document whether those two types work in Spanish.
- Palo Alto Networks (Enterprise DLP) uses patterns augmented with machine learning and document classifiers built on deep neural networks. In 2024 it stated that the language models it uses to train them had cut potential false positives by more than 90%, without publishing the method.
- Netskope classifies files with language models. For those rules it sends part of the extracted text to its classification service. According to its documentation it does not store it or use it for training, but it does process it.
- Forcepoint puts a small language model that runs on CPUs at the centre of its classification (AI Mesh), combined with neural classifiers and rules.
None of the three vendors publishes accuracy figures for Spanish.
What about Microsoft Purview?
I checked the official documentation on 20 November 2026. This is what is available for a case like Dolores Fuertes:
- Classic sensitive information types. Still patterns with supporting elements inside a proximity window, plus confidence levels. Microsoft calls them pattern-based classifiers.
- Named entities. “All full names” uses a dictionary of first names, one of surnames and patterns of how names are formed, different for each country, and Spanish is among the supported languages. It would detect Dolores Fuertes, but the documentation does not say it uses context to rule out severe pain. Microsoft describes named entities as complex dictionary and pattern-based classifiers.
- Medical terms. “All medical terms and conditions” is dictionary-based and, according to its own documentation, detects English terms only. It would not recognise «dolores fuertes» as a symptom.
- Trainable classifiers. These are the ones that do use machine learning. The built-in healthcare ones, “Healthcare” and “Health/Medical forms”, support English only. So do custom classifiers.
- Generative AI. Data Security Triage Agent, which runs on Security Copilot and consumes paid Security Compute Units (SCUs), prioritises DLP alerts that have already been raised. Data Security Investigations analyses content after an incident, and Data Security Posture Agent, in preview, helps discover data. They are useful tools, but they come in once the rule has fired: if it fires on every report that mentions severe pain, the agent helps sort the alert queue and the queue is still there. The documentation does not say which languages they support either.
- What has been announced. At Ignite, on 17 November 2026, Microsoft announced enhanced built-in sensitive information types from January 2027, starting with the hundred-plus most used, promising precise detection with less manual tuning. It did not explain the method or mention languages, and the examples it gave were card numbers and US Social Security numbers. It will need testing when it arrives.
Today, in Spanish, Purview does not tell Dolores Fuertes from severe pain by reading the sentence.
What does work in Purview: Exact Data Match
Exact Data Match (EDM) tackles the problem from another angle: instead of guessing whether a word is a name, it checks whether it belongs to a person who is actually in your records.
- You prepare a table with the real data, for example medical record number, first name, surnames and date of birth, and hash it with a salt. Only the hashed values are uploaded to the service. The table can hold up to 100 million rows and can be refreshed daily.
- The rule first needs a primary element detected with a pattern, such as the medical record number, and then looks nearby for the supporting fields, such as the name. Matching is done by comparing hashes.
- That way, «dolores fuertes» in a report only counts if it appears next to the record number of a patient with that name.
Microsoft recommends EDM precisely to reduce false positives with personal and health data of known individuals. It has an operating cost: the table must be kept in sync with the patient master index, and Purview’s advanced classification capabilities come with E5-tier plans. You can check what yours covers in the Purview licensing simulator.
To fine-tune the remaining rules, Purview lets you mark in Content explorer and in alerts whether a finding was correct, and it shows a summary of the context it appeared in. This is not available for named entities or EDM, and the marking does not retrain anything: it helps whoever administers the tenant adjust thresholds, confidence levels and exclusions.
What AI does not solve
- There is still a threshold. A model returns a score and someone decides at what point the rule fires. More sensitivity brings more false positives, and less sensitivity lets more through.
- Where classification happens. If the model runs on an external service, part of the content leaves your environment. With health data you need to know where it is processed, under which data processing agreement and with what retention.
- Small models fail more. A study published in iScience in October 2025 on 3,650 UK clinical records measured an F1 of 0.939 for a dedicated de-identification service and 0.898 for GPT-4 with ten examples. Small models collapsed: Phi-3 reached 0.448 and the 8-billion-parameter Llama 3, 0.198. The authors observed over-redaction and invented content.
- Latency. In a DLP that acts in real time, in email or on the user’s device, every millisecond counts. Google advises against its name detector in latency-sensitive operations.
- Explaining the decision. If a rule blocks a message, someone will ask why. Presidio returns its reasoning; most commercial tools document only a score.
- Language. Something that works in English guarantees nothing in Spanish, least of all with clinical vocabulary.
How I would approach it today
- I would start with the information. What health data exists, where it lives, who can reach it and through which channels it leaves. The tool comes later.
- I would separate what has a shape from what does not. National ID, social security number or medical record number, with patterns. Patient names, with EDM against the patient master index.
- I would combine conditions. Record number plus name plus clinical term says far more than a name on its own.
- I would test any AI-based tool with our own documents, in our language. A test set with cases like Dolores Fuertes, measuring false positives and false negatives before and after.
- I would ask where the model runs and which part of the content is sent.
- I would deploy in audit mode first and record every adjustment before moving to blocking.
- I would use triage AI for what it is good at: sorting and explaining alerts. A badly designed rule is not fixed by putting an agent behind it.
That is how the Data Loss Prevention service works. If you also use Microsoft 365 Copilot or agents, the guide to DLP for Copilot and agents explains how permissions, labels and policies fit together.
Frequently asked questions
Does AI eliminate DLP false positives?
No. It reduces the ones that depend on context, such as a name that is also an ordinary word, but any model returns a score and someone has to set the threshold. Raising it lets more through; lowering it generates more alerts.
Can Microsoft Purview tell a name from a symptom in Spanish?
As of 20 November 2026, not by context. “All full names” supports Spanish but works with dictionaries and patterns; “All medical terms and conditions” and the healthcare trainable classifiers work in English only. For data on known patients, Microsoft recommends Exact Data Match, which compares against the real, hashed patient list.
Can I send health data to an AI model for classification?
It depends on where it is processed. Health data is a special category under Article 9 of the GDPR. If the model sits outside your environment, review the data processing agreement, the processing location and the retention, and prefer options that classify inside your own environment.
Which Microsoft licence do I need?
Basic sensitive information types are available in mid-tier plans. Named entities and the other advanced classifiers require E5-tier plans or add-ons. The Purview licensing simulator tells you what your plan covers, and it is worth confirming against the current Product Terms.
Any questions, I am here. No fluff.
Microsoft PurviewSensitive information types · All full names · All medical terms and conditions · Named entities · Trainable classifiers · Classifier definitions · Exact Data Match · Reduce false positives · Increase classification accuracy · Data Security Triage Agent · Ignite announcements (17 November 2026)
ResearchMarimon et al., MEDDOCAN overview (IberLEF 2019) · García-Pablos et al., BERT on MEDDOCAN (LREC 2020) · Clinical de-identification with language models (iScience, October 2025)
Other toolsPresidio · context · Presidio · project transition · Google · hotword rules · Google · infoType reference · Palo Alto Networks (1 October 2024) · Netskope · file classifiers · Forcepoint · AI Mesh
RegulationRegulation (EU) 2016/679 (GDPR), Article 9
Content last reviewed on 20 November 2026, using the sources consulted that day. Microsoft Purview’s status, languages and licensing change often: check them on Microsoft Learn before deciding. The Palo Alto Networks, Netskope and Forcepoint figures come from the vendors themselves. This is a courtesy translation; the Spanish version is the reference text.