Erster Auftrag und der Grund für eine GmbH: Mein Start in die Selbstständigkeit | Umair Zaffar
Umair Zaffar October 21, 2026 0 Comments

EU AI Act Article 10: Data and Data Governance, Implemented in Practice

EU AI Act Article 10 requires high-risk systems to be built on training, validation and testing data with documented provenance, checked quality and examined biases. Our answer: every dataset gets a versioned record with provenance, zone and known gaps, and no model sees data before it has been pseudonymised and has passed the quality gate.

Article 10 is the part of the AI Act where most production models in banks and insurers would fail today if someone asked. Not because the models are bad, but because nobody can show where the training data came from, who set the labels, and whether the test set contains the same people as the training set. The article addresses providers of high-risk AI systems; Art. 16(a) makes the requirements of Section 2 their obligation. An insurer that trains or fine-tunes a model on its own claims data is usually a provider itself, not just a deployer.

The EU AI Act has been in force since 1 August 2024. The prohibitions and AI literacy duties have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 10 belongs to those high-risk obligations. The Omnibus moved deadlines; it removed no obligation. And provenance cannot be invented retroactively in 2027: whoever needs the evidence then has to write it at the next training run. I describe the overall frame in EU AI Act in practice.

What EU AI Act Article 10 requires

The article sets no thresholds. It requires practices and properties, and it requires that you can prove them. Simplified, by paragraph:

  • Art. 10(1): High-risk systems that involve training models with data must be developed on the basis of training, validation and testing data sets that meet the quality criteria of paragraphs 2 to 4. Where special categories of personal data are used, Art. 4a(1) applies in addition.
  • Art. 10(2): Those data sets need data governance and management practices appropriate for the intended purpose. They must cover eight points: (a) the relevant design choices, (b) data collection and origin, and for personal data the original purpose of collection, (c) preparation, meaning annotation, labelling, cleaning, updating, enrichment and aggregation, (d) the assumptions about what the data is meant to measure and represent, (e) an assessment of availability, quantity and suitability, (f) examination for biases that could affect health and safety or fundamental rights or lead to prohibited discrimination, (g) measures to detect, prevent and mitigate those biases, and (h) identifying and addressing data gaps and shortcomings that stand in the way of compliance.
  • Art. 10(3): The data sets must be relevant, sufficiently representative and, to the best extent possible, free of errors and complete in view of the intended purpose. They need appropriate statistical properties, including with regard to the persons and groups the system is meant to be used on. This may be met at the level of individual data sets or a combination of them.
  • Art. 10(4): The data sets must take into account the geographical, contextual, behavioural or functional setting in which the system is meant to be used. A model trained on data from one country is not automatically suitable for another.
  • Art. 10(5) (former) and Art. 4a: Until the Digital Omnibus this is where the exception lived that allows providers to exceptionally process special categories of personal data for bias detection and correction. The rule now sits in Article 4a, with the same six conditions: it cannot be done with other data, including synthetic or anonymised data; technical limits on reuse and pseudonymisation; strict access control and documented access; no transfer to other parties; deletion once the bias is corrected or the retention period ends; and an entry in the record of processing activities explaining why there was no other way. Art. 4a(2) opens the route to deployers and to providers of other systems as well, but obliges nobody to carry out bias detection.
  • Art. 10(6): For high-risk systems that do not train models, paragraphs 2 to 5 apply only to the testing data sets. If you only use a finished model, you still have to treat at least your test data this way.

Article 10 does not stand alone. Annex IV point 2(d) requires data sheets on the training data in the technical documentation under Art. 11: provenance, scope, main characteristics, selection, labelling, cleaning. Art. 17(1)(f) requires data management procedures in the quality management system. And Art. 26(4) obliges deployers to keep their input data relevant and sufficiently representative to the extent they control it. In short: you must be able to account for every dataset a model was built or tested with, and you must have looked for biases before somebody else finds them.

The technical example: claims classification at an insurer

An insurer builds a model that pre-sorts incoming claims: automatic settlement, case handling, in-depth review on suspicion of fraud, or a query back to the customer. The model does not decide on the benefit, but it decides how fast and how critically a human looks at the case. It is trained on historical claims: structured fields from the claims system, free text from notifications and expert reports, scanned documents, and as the label the actual handling decision of recent years.

Whether such a classifier is high-risk is a matter of interpretation. Annex III point 5(c) names risk assessment and pricing in life and health insurance, not claims pre-sorting. We treat the system as if it were high-risk anyway: it prepares decisions about people, the same data pipeline later feeds models that do fall under Annex III, and the data work is the same. Article 10 bites in four places:

  • The labels are old decisions. If handlers used to send certain postcodes or names to fraud review more often, the model learns exactly that. Art. 10(2)(f) requires us to examine this before we train.
  • The free text is full of people. Names, addresses, health details and account numbers sit in claims notifications. Art. 10(2)(b) asks for the original purpose of collection, and that was claims settlement, not model training.
  • The data comes from three systems with three truths. Duplicates, renamed claim types, a process change after which the labels are no longer comparable. That is Art. 10(2)(c) and (h).
  • Representativeness is not given, it has to be produced. Regions, sales channels, claim types and languages are distributed differently in the data than in future operations. That is Art. 10(3) and (4).

How we implement it

At SIFAMO we treat data like models: registered, with an owner, with an expiry date. The building blocks come from our governance layer; the rest is discipline.

Three data zones and IDs instead of names

Every source gets a zone. Red is raw data with personal reference: customer master data, unredacted claims notifications, health details. Red data never reaches a training environment, and the policy gateway rejects calls with a red payload no matter who sends them. Gray is pseudonymised data with a documented purpose, access control and logged access. This is where training and evaluation happen. Green is data without personal reference: aggregates, metrics, data sheets, synthetic cases. Only green artefacts leave the project.

The way from red to gray is pseudonymisation. Identifiers are replaced by stable IDs, free text is scrubbed of names, addresses, account numbers and health details, and the mapping table lives in a separate system that neither the training environment nor the model can reach. The IDs are stable per person, because that is the only way we can later verify that nobody sits in both the training and the test set. We check the pseudonymisation by sampling the data, not by reading the policy.

One record per dataset: provenance, version, gaps

Just as every model has a registry record, every training, validation and testing dataset has a dataset record. It is the data sheet Annex IV asks for, and it is versioned because the dataset changes. Simplified:

{
  "dataset_id": "claims-train-2026-v7",
  "purpose": "Training claims classifier (routing, not a benefit decision)",
  "owner": "Head of Claims, Motor",
  "zone": "gray (pseudonymised, access logged)",
  "sources": ["Claims system, extract 2026-09-30", "Document archive, OCR text", "Handler decisions 2023-01 to 2026-06"],
  "legal_basis": "Performance of contract; purpose of collection: claims settlement",
  "transformations": ["Pseudonymisation v3", "Free-text scrubbing v5", "Label rule LR-2026-02", "Dedup by claim ID"],
  "split": "by person, in time: test = last 6 months",
  "rows": {"train": 0, "validation": 0, "test": 0},
  "quality_gate": "passed 2026-10-05, report QG-0412",
  "bias_report": "BR-0412, group metrics and counterfactual test",
  "known_gaps": ["Postal claims before 2024 without OCR", "Travel line under-represented", "Label rule differs before 2024-03"],
  "hash": "sha256:...",
  "expires": "2027-04-05"
}

Three fields carry the article: sources and transformations are the origin and preparation from Art. 10(2)(b) and (c), step by step and with a version for every rule. known_gaps is Art. 10(2)(h): gaps we know about and have not closed, written down openly instead of hidden. The model registry record points to the dataset version it was trained and tested with. That makes it possible to walk back from any decision in the evidence log, through the model version, to the datasets.

Data quality as a gate, not a report

Before every training run the quality checks execute, and a dataset that fails them is not trained on. The thresholds are set by the business owner, not the data science team, because the owner has to explain later why they are sufficient. We check:

  • Schema and completeness. Mandatory fields, value ranges, share of missing values per field and per source.
  • Duplicates and leakage. The same claim from two systems, the same person in training and test, labels that only came into existence after the decision.
  • Label quality. A sample is re-labelled by two handlers. The disagreement rate is documented, not argued away. “Free of errors to the best extent possible” means for us: measured and disclosed.
  • Representativeness. Distribution of claim types, regions, channels and periods in the dataset against the distribution in live operations, with tolerances the owner signs. That is our reading of “sufficiently representative”, because the regulation gives no number.
  • Breaks in time. Process or rule changes after which labels mean something else. Old cases are relabelled or excluded, never silently mixed in.

Bias examination and the route via Article 4a

The bias examination runs in two stages. First counterfactually: otherwise identical cases in which only the first name, postcode, date of birth or language of the free text varies. If the assignment to fraud review flips, the model stops here. Second with group metrics on the attributes the insurer lawfully holds, such as age and region: how often each group lands in in-depth review, and how often wrongly, measured against the later outcome.

For attributes such as ethnic origin or religion an insurer usually holds no data, and we do not collect them. Where such an examination cannot be done without special categories and the data is lawfully present, we use the route via Article 4a, the former Art. 10(5), with all six conditions: the justification for why synthetic or anonymised data is not sufficient goes into the record of processing activities before the first access. The attributes sit in a sealed analysis environment with a named access list and logged access. They are used exclusively for measuring and never become training features. They do not leave the environment, and they carry a deletion date recorded in the dataset register. I say this deliberately clearly: the exception is a permission to measure, not a permission to collect.

What the auditor sees

An auditor gets no slides. They get the dataset record with version, sources and known gaps, the quality report with the thresholds and who set them, the bias report with both stages, the sampled proof of pseudonymisation, and the model registry record that points at exactly these versions. Plus the three sign-off signatures: business for purpose and data, risk for classification and bias result, IT for zones and access. That is AI governance you can read out of a system instead of narrating it.

Typical mistakes

  • Pseudonymisation in the prompt, raw data in training. The gateway protects the runtime path, but the model was fine-tuned on unredacted claims notifications. The names are then in the weights, and the model version is retired, not patched.
  • Labels as truth. Historical decisions are adopted without asking who made them and with which patterns. The model learns yesterday’s practice and gets a seal of approval for it.
  • Random split. Training and test are separated at random, people appear on both sides, and the test score is too good to be true.
  • “We hold no protected attributes, so there is no bias.” First name, postcode and language are proxies. If you do not test them, you have not examined anything.
  • The vendor writes the data sheet afterwards. It then describes what one would have liked to do. Provenance is created while building or not at all.

Article 10 checklist

  • Every source has a data zone, and red data technically cannot reach a training environment.
  • Every training, validation and testing dataset has a versioned record with owner, sources, legal basis and original purpose of collection.
  • Every preparation rule (pseudonymisation, cleaning, labelling) is versioned and referenced in the record.
  • The split is separated by person and by time, and the leakage check is part of the gate.
  • Representativeness is measured against the live distribution, with tolerances the business has signed.
  • The bias examination includes counterfactual tests on proxy attributes and group metrics, with a documented result.
  • Every use of special categories for bias measurement meets the six conditions of Art. 4a and has a deletion date.
  • Known gaps are in the record, and the model registry record points to the dataset versions.

Open questions

  • What does “sufficiently representative” mean in numbers? The regulation gives none, and until harmonised standards on data quality are available and take hold, every provider sets its own tolerances. Our reading is the comparison with the live distribution. Whether a supervisor sees it the same way, we do not know yet.
  • How far does Article 10 reach into a foundation model? Whoever builds a classifier on top of a purchased language model does not control its pre-training data. Art. 10(6) helps where no training happens at all. For fine-tuning it remains open which evidence beyond one’s own datasets will be expected.
  • How do you prove that synthetic data would not have been enough? Art. 4a demands exactly that as its first condition. A robust method for showing the limits of synthetic data for bias measurement is a research question, not settled practice.

Frequently asked questions

Does Article 10 also apply to us as a deployer?

Directly, Article 10 addresses providers. As a deployer you are covered by Art. 26(4): keep input data relevant and sufficiently representative to the extent you control it. But as soon as you train, fine-tune or substantially modify a high-risk system on your own data, you can become a provider yourself under Art. 25, and then Article 10 applies in full. More on this under all articles at a glance.

Does training data have to be error-free?

No. Art. 10(3) says free of errors and complete “to the best extent possible”, in view of the intended purpose. What you need is a measured error rate, a justification for why it is sufficient for the purpose, and a list of the gaps you know about. A data sheet without known gaps is a warning sign to me, not a mark of quality.

May we collect ethnic origin or religion to measure bias?

Collect is the wrong word. Article 4a exceptionally allows processing special categories for bias detection and correction when there is no other way and all six conditions are met. It creates no licence to gather new attributes and no obligation to do so. Where the data is lawfully present, we use it in a sealed environment exclusively for measuring and delete it afterwards.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment