Article 10 of the EU AI Act regulates training data, not model weights. Three data-architecture choices decide whether you can produce the evidence by 2 December 2027.
By
Billy Allocca
Table of Contents
EU AI Act Compliance Starts at the Data Layer: Architecture Decisions You Cannot Undo
EU AI Act compliance for high-risk systems depends on the data layer because Article 10 regulates training, validation, and testing datasets: their provenance, preparation, representativeness, bias examination, and documented gaps. Those obligations can only be met if the storage format, governance model, and catalog were built to record that evidence before the model was trained.
The AI Act is Regulation (EU) 2024/1689, in force since 1 August 2024 [1]. Its high-risk obligations were due on 2 August 2026, but the AI Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the Annex III deadline to 2 December 2027 and the Annex I deadline to 2 August 2028 [2][3]. The Commission describes the change as simplification "while preserving strong safeguards," and the Article 10 data obligations were carried forward intact [3]. As of September 2026 that leaves 15 months to make the datasets behind credit scoring, insurance pricing, and hiring systems auditable.
The post on the SLM-first sovereign AI architecture and the guide on data sovereignty versus data residency cover the deployment-location questions this article leaves aside.
What Does EU AI Act Article 10 Require From Your Data?
Article 10 is a dataset regulation written as a list of governance practices, and each practice maps to a capability your data layer either has or lacks. Paragraph 3 sets the quality bar: datasets "shall be relevant, sufficiently representative, and to the best extent possible, free of errors and complete in view of the intended purpose," and paragraph 4 adds that they must reflect the geographical, contextual, behavioural, and functional setting of use [4].
A high-risk AI system is one listed in Annex III (biometrics, critical infrastructure, education, employment, essential services including credit scoring and life or health insurance pricing, law enforcement, migration, and justice) or embedded as a safety component in a product covered by Annex I [5].
Article 10(2) requirement, plain language | Technical capability required |
|---|---|
(a) Design choices for the dataset | Versioned dataset definitions in the catalog |
(b) Origin of the data and its original collection purpose | Column-level lineage tied to GDPR purpose tags |
(c) Preparation operations (labelling, cleaning, enrichment, aggregation) | Transformation-level lineage; immutable input and output snapshots |
(d) Assumptions about what the data represents | Semantic layer with definitions attached to columns |
(e) Assessment of availability, quantity, and suitability | Quality metrics stored with the training snapshot |
(f) Examination for biases affecting health, safety, or fundamental rights | Governed, logged access to special-category data for bias testing, masked elsewhere |
(g) Measures to detect, prevent, and mitigate those biases | Reproducible pipelines that write results back to the catalog |
(h) Data gaps and how they will be addressed | Catalog gap register linked to the risk file |
None of these capabilities can be added after training with any credibility, because the evidence has to describe the dataset as it existed when the model learned from it.
What Other EU AI Act Obligations Depend on the Data Layer?
Three further provisions turn Article 10 into a document trail about the past. Technical documentation under Article 11 must exist before the system is placed on the market, be kept up to date, and contain at minimum the Annex IV elements [6]. Annex IV point 2(d) asks for datasheets covering "the training data sets used, including a general description of these data sets, information about their provenance, scope and main characteristics; how the data was obtained and selected; labelling procedures ... data cleaning methodologies" [7]. Record-keeping under Article 12 requires automatic logging of events over the system's lifetime [8], and deployers must retain those logs for at least six months and confirm that input data is relevant and sufficiently representative for the intended purpose [9].
Article 2(7) of the AI Act states that GDPR applies to all personal data processed under the Act [10]. The Omnibus deleted Article 10(5) and re-homed its substance in a new Article 4a, which permits processing special categories of personal data for bias detection and correction only where strictly necessary and under safeguards including pseudonymisation, access controls, and deletion once the correction is done [11].
A working checklist for the data layer:
Every training table can be reconstructed as of its training timestamp.
Every column carries source, original purpose, and classification tags.
Bias-testing queries on special-category columns are logged with the requesting identity.
Access logs are retained for at least six months in a form an auditor can read.
Which Data Architecture Decisions Cannot Be Undone?
Three decisions determine whether that evidence exists, and each is expensive to reverse once models depend on it. The 2021 CIDR paper that defined the lakehouse described data lakes as managing "just a bunch of files," which makes transactions, versioning, and quality enforcement hard to offer [12]. A lakehouse is an architecture that keeps data in open, directly accessible files and adds a transactional metadata layer that provides versioning and auditing [12]. Apache Iceberg is the common implementation: time travel "enables reproducible queries that use exactly the same table snapshot," and schema changes do not resurrect deleted data [13].
Federated computational governance is the data mesh principle under which global policies are "baked computationally into the platform" while domains keep authority over their own models [14]. Column-level lineage records which input columns produced each output column and through what transformation type, as the OpenLineage facet specifies [15].
Decision | Choice that produces Article 10 evidence | Choice that cannot | Cost of reversing later |
|---|---|---|---|
Data lake vs. lakehouse | Iceberg or an equivalent open table format with snapshot retention | Files in object storage without table metadata, or a proprietary warehouse format | Training sets rebuilt from pipeline code and backups; snapshots never taken cannot be recovered |
Centralized vs. federated governance | One policy engine enforced at query time across every engine, domains owning definitions | Per-tool permissions (warehouse grants, IAM buckets, notebook ACLs) | Re-permissioning thousands of objects; the historical access record stays fragmented |
Catalog depth | Column-level and transformation-level lineage with classification and purpose tags | Table-level inventory only | Re-parsing every job that wrote a training table; original collection purpose often unrecoverable |
In each row, reversal means recreating history the original architecture never captured, and for data collected under GDPR purpose limitation the original purpose is frequently unknown to anyone still at the company.
What Is the EU AI Act Compliance Timeline for a CDO?
The dates have been fixed in law since 27 July 2026, and the Annex III deadline sits 15 months away [16].
Date | Obligation | Status on 10 September 2026 |
|---|---|---|
1 August 2024 | Regulation (EU) 2024/1689 enters into force [1] | Done |
2 February 2025 | Prohibited practices (Article 5) and AI literacy apply [16] | In force |
2 August 2025 | General-purpose AI model obligations apply [16] | In force |
27 July 2026 | AI Omnibus, Regulation (EU) 2026/1744, enters into force [2][3] | In force |
2 August 2026 | Article 50 transparency obligations apply [16] | In force |
2 December 2027 | Chapter III obligations, including Articles 10 to 12, apply to Annex III systems [3][16] | 15 months out |
2 August 2028 | Same obligations apply to Annex I embedded systems [3][16] | 23 months out |
Penalties under Article 99 scale with the breach. Prohibited practices carry fines of up to EUR 35 million or 7 percent of worldwide annual turnover, whichever is higher; breaches of operator obligations, including Article 26 deployer duties, carry up to EUR 15 million or 3 percent; incorrect information to authorities carries up to EUR 7.5 million or 1 percent [17]. For a bank with EUR 20 billion in revenue, the 3 percent tier is EUR 600 million.
Why Not Wait Until Closer to December 2027 to Change the Data Architecture?
Waiting works for policy documents and fails for datasets, because the models you will defend in 2028 are being trained on today's tables. A credit-scoring model retrained in Q1 2027 draws on transaction history collected from 2024 onward, and if those tables live in a lake with no snapshots and a table-level catalog, the Annex IV datasheet will have to describe provenance, selection, and cleaning that nobody recorded [7]. Article 11 requires the documentation to exist before the system is placed on the market, so the real deadline is the next model launch [6].
There is a legitimate version of the wait argument. CEN-CENELEC targeted Q4 2026 for the harmonised standards that define presumption of conformity [18], and the Commission has promised simplified documentation forms for SMEs and small mid-caps [11]. Those documents will set the format of the evidence, and the evidence remains which data, from where, prepared how, tested for which biases.
NexusOne Makes Article 10 Evidence a Property of the Data Layer
NexusOne is an AI-native data layer that sits horizontally across the estate, and its governance model produces the Article 10 record as a by-product of operation. Identity from Active Directory, Okta, LDAP, or SAML is federated into one Keycloak layer, so every query and agent request carries one principal. A policy defined once in Apache Ranger is enforced simultaneously in Trino, Apache Spark, Apache Kyuubi, and per object in S3 storage, so an auditor sees one access record.
Tag a dataset in DataHub as special-category, map roles to tags, and Ranger policies generate automatically at bucket, schema, and table level, so the bias-testing access Article 4a permits is granted to a named role and masked for everyone else. Data lives in Apache Iceberg, so training snapshots can be tagged and retained. The AI & Data Control Plane applies the same identity and policy model to agent and MCP requests at the boundary, with PII redaction before any external model call, per-role budgets, and every request logged. The layer runs identically on-prem, in your cloud, or air-gapped. The one governance model for data and AI post covers enforcement; the guide on data lineage for AI models covers provenance. To review your Annex III inventory, book a working session with a NexusOne architect.
Key Takeaways
Article 10 of the EU AI Act regulates training, validation, and testing datasets, so compliance evidence has to be captured by the data layer at training time.
The AI Omnibus (Regulation 2026/1744) moved the Annex III high-risk deadline to 2 December 2027 without changing the Article 10 data obligations.
Articles 11 and 12 and Annex IV require datasheets and logs that describe datasets as they existed in the past, which only snapshot-capable storage can support.
Table format, governance topology, and catalog depth are the three decisions that cannot be reversed without reconstructing history that was never recorded.
Special-category data may be processed for bias detection under the new Article 4a only with pseudonymisation, access controls, and deletion, which requires tag-driven policy enforcement.
FAQ
What Are the EU AI Act Data Architecture Requirements?
For high-risk systems, Article 10 requires documented governance practices covering data design, origin and original collection purpose, preparation steps, representational assumptions, suitability assessment, bias examination and mitigation, and known gaps. Annex IV requires datasheets describing provenance, selection, labelling, and cleaning, and Article 12 requires automatic logging over the system's lifetime. Architecturally, this means versioned open table storage, column-level lineage with classification tags, and a single access log across engines.
How Do You Comply With the EU AI Act as an Enterprise?
Inventory models against Annex III and Annex I to find the in-scope systems. For each, confirm that training tables can be reconstructed as of training time, that every column has source and purpose lineage, and that bias-testing access to special-category data is governed and logged. Draft the Annex IV documentation before the next retraining goes live, since Article 11 requires it to exist before the system is placed on the market. The Annex III deadline is 2 December 2027.
What Does EU AI Act Article 10 Say About Data Governance?
Article 10 requires that training, validation, and testing datasets for high-risk systems follow governance practices appropriate to the system's purpose, listing eight areas from design choices through data-gap identification. Paragraph 3 sets the quality standard: relevant, sufficiently representative, and as far as possible free of errors and complete. Paragraph 4 requires datasets to reflect the setting in which the system will be used. The former paragraph 5 on special-category data was moved by the Omnibus to a new Article 4a.
What Are a CDO's Obligations Under the EU AI Act?
A CDO typically owns the evidence behind both provider and deployer duties. As a provider, that means Article 10 dataset governance, Article 11 technical documentation, and Article 12 logging. As a deployer, Article 26 requires input data that is relevant and sufficiently representative, log retention of at least six months, and use of the provider's information to complete a GDPR data protection impact assessment. In practice the CDO decides whether the storage format, governance model, and catalog can produce those records.
References
Regulation (EU) 2024/1689 (Artificial Intelligence Act), EUR-Lex. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
EU AI Omnibus enters into force, amending the AI Act, White & Case, July 2026. https://www.whitecase.com/insight-alert/eu-ai-omnibus-enters-force-amending-ai-act
AI Omnibus enters into force, European Commission, Shaping Europe's digital future, 27 July 2026. https://digital-strategy.ec.europa.eu/en/news/ai-omnibus-enters-force
Article 10: Data and Data Governance, EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/article/10/
Annex III: High-Risk AI Systems Referred to in Article 6(2), EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/annex/3/
Article 11: Technical Documentation, EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/article/11/
Annex IV: Technical Documentation Referred to in Article 11(1), EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/annex/4/
Article 12: Record-Keeping, EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/article/12/
Article 26: Obligations of Deployers of High-Risk AI Systems, EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/article/26/
Article 2: Scope, EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/article/2/
Digital Omnibus on AI, EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/ai-act-explorer/digital-omnibus/
Lakehouse: A New Generation of Open Platforms that Unify Data Warehousing and Advanced Analytics, Armbrust, Ghodsi, Xin, Zaharia, CIDR 2021. https://www.cidrdb.org/cidr2021/papers/cidr2021_paper17.pdf
Apache Iceberg, project home page (schema evolution, time travel and rollback). https://iceberg.apache.org/
Data Mesh Principles and Logical Architecture, Zhamak Dehghani, martinfowler.com, 3 December 2020. https://martinfowler.com/articles/data-mesh-principles.html
Column Lineage Dataset Facet, OpenLineage specification. https://openlineage.io/docs/spec/facets/dataset-facets/column_lineage_facet
Implementation Timeline, EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/implementation-timeline/
Article 99: Penalties, EU Artificial Intelligence Act explorer. https://artificialintelligenceact.eu/article/99/
Update on CEN and CENELEC's Decision to Accelerate the Development of Standards for Artificial Intelligence, CEN-CENELEC, 23 October 2025. https://www.cencenelec.eu/news-events/news/2025/brief-news/2025-10-23-ai-standardization/
