EU AI Act: PII Compliance for AI Builders
Prohibited practices have been banned since February 2, 2025, and Article 50 transparency rules apply since August 2, 2026. High-risk obligations follow on December 2, 2027 (Annex III) and August 2, 2028 (Annex I). Penalties reach €35M for prohibited practices and €15M for high-risk non-compliance. Here is what you need to do with your training data, documentation, and inference pipelines.
Who Must Comply
The EU AI Act applies to any organisation that places AI systems on the EU market or puts them into service in the EU — regardless of where the company is headquartered. US, UK, and Asian companies deploying AI to European users are fully subject to its requirements.
High-Risk AI Systems
- Biometric identification and categorisation
- Safety-critical infrastructure management
- Educational or vocational training scoring
- Employment recruitment and screening
- Credit, insurance, and financial scoring
- Law enforcement and predictive policing
- Migration and asylum processing
- Administration of justice
Limited Risk Systems
- Chatbots and conversational AI
- Deepfake content generation
Transparency obligations apply: users must be informed they are interacting with AI.
Minimal Risk Systems
- Spam filters
- AI-powered games
- Inventory management
- Basic process automation
No specific EU AI Act obligations, though voluntary codes of conduct are encouraged.
Penalty Structure
The EU AI Act penalty regime is tiered by severity. The higher of the fixed amount or the percentage of global annual turnover applies. For a company with €1 billion in revenue, a prohibited AI violation could reach €70 million.
| Violation Type | Fixed Amount | % Global Turnover | Example (€1B company) |
|---|---|---|---|
| Prohibited AI practices | €35,000,000 | 7% | Up to €70,000,000 |
| High-risk AI non-compliance | €15,000,000 | 3% | Up to €30,000,000 |
| Misinformation to authorities | €7,500,000 | 1% | Up to €10,000,000 |
Note: The higher of the fixed amount or the percentage applies. For SMEs and start-ups, the reverse applies: each fine is capped at whichever is lower, the fixed amount or the percentage (Article 99(6)).
Article 10 — Data Governance Requirements
High-risk AI systems must use training data that is subject to data governance practices, including examination for possible biases. Article 10 does not require data sets to be free of personal data; Article 4a (formerly Article 10(5), amended by Regulation (EU) 2026/1744) even permits limited processing of special categories of personal data for bias correction, under strict safeguards.
“Training, validation and testing data sets shall be relevant, sufficiently representative, and to the best extent possible, free of errors and complete in view of the intended purpose.”
— EU AI Act, Article 10(3)
Audit Training Datasets for PII Exposure
Scan all training data repositories for exposed personal information. Use piisafe.eu for web-based documentation or the REST API for file-based batch scanning. Generate and retain audit reports as evidence.
Anonymize or Pseudonymize Personal Data
Apply pseudonymization to training corpora, for example with anonym.legal's API. For NLP training, the same real name should map to the same pseudonym throughout the corpus — that preserves semantic coherence. Pseudonymized data is still personal data under the GDPR.
Document Anonymization Methodology
Record the exact anonymization method used: detection engine version, entity types targeted, replacement strategy (replace/mask/hash/encrypt), and infrastructure details. The replace, redact, mask and hash methods produce deterministic results, enabling reproducible methodology documentation.
Implement Ongoing Pipeline Monitoring
Integrate PII scanning into your data ingestion pipeline. Every new training data batch should be automatically scanned before it reaches your training environment. Use the REST API with CI/CD integration for automated enforcement.
Establish Re-identification Risk Assessment
Document your re-identification risk assessment process. For each anonymized dataset, record: the anonymization technique applied, residual risk level, quasi-identifier analysis, and the qualified assessment that risk is sufficiently low under GDPR Recital 26 and EU AI Act Article 10 standards.
Article 11 — Technical Documentation Checklist
High-risk AI providers must maintain technical documentation before placing the system on the market. This documentation must be kept up to date throughout the system’s lifecycle.
- General description of the AI system and its intended purpose
- Description of the system’s design, development methodology, and architecture
- Training data governance practices (Article 10) — including PII anonymization methodology
- Validation and testing procedures with performance metrics
- Risk management system and mitigation measures (Article 9)
- Post-market monitoring plan and incident reporting procedure
How anonymize.solutions Helps
Four components directly address EU AI Act Article 10 and Article 11 requirements, providing both the anonymization capability and the audit trail documentation that regulators will request.
REST API for Training Data Pipelines
Integrate PII anonymization directly into your data ingestion pipeline. Batch process training datasets via API, generate audit logs, and produce compliance-ready reports for Article 11 documentation.
piisafe.eu for Documentation Scanning
Scan publicly accessible AI documentation, model cards, README files, and data sheets for exposed PII. The interface is free; PII detection requires an anonym.legal API key (from €3/month incl. VAT). Export results for DPIA and Article 11 annexes.
Deterministic Audit Trail
The replace, redact, mask and hash methods are deterministic: the same input and settings produce the same output. Encrypt uses a random initialization vector, so its ciphertext differs on each run. This enables reproducible documentation of your anonymization methodology.
One Building Block for GDPR and the EU AI Act
GDPR Article 25 (data protection by design) and EU AI Act Article 10 share data-minimization goals. Anonymization supports the data governance requirements of both — one technical building block, not a guarantee of compliance with either framework.
EU AI Act FAQ
Answers to the most common questions about EU AI Act compliance, penalties, and PII requirements.
In stages. Prohibited AI practices have been banned since February 2, 2025. General-purpose AI model (GPAI) requirements became effective August 2, 2025, and Article 50 transparency obligations apply since August 2, 2026. Following the 2026 Digital Omnibus amendment, high-risk AI system requirements apply from December 2, 2027 (Annex III) and August 2, 2028 (Annex I).
High-risk systems include: biometric identification and categorisation, safety-critical infrastructure management, educational or vocational training scoring, employment recruitment and screening, credit and insurance scoring, law enforcement risk assessment, migration and asylum processing, and administration of justice. The full list is in Annex III of the EU AI Act.
Three tiers apply: Prohibited AI practices — up to €35M or 7% of global annual turnover (higher applies). High-risk AI non-compliance — up to €15M or 3%. Providing false or misleading information to authorities — up to €7.5M or 1%. For SMEs and start-ups, the lower of the fixed amount and the percentage applies (Article 99(6)).
Article 10 requires data governance practices for training, validation and testing data: bias examination, and data sets that are relevant, sufficiently representative and, to the best extent possible, free of errors and complete. It does not require data free of personal information; personal data in these sets remains subject to GDPR, which is where anonymization helps. Failure to implement adequate data governance is a direct Article 10 violation.
Yes. The EU AI Act has extraterritorial scope: any company placing AI systems on the EU market or putting them into service in the EU must comply, regardless of where the company is headquartered. US, UK, Asian, and other non-EU companies are fully subject to the regulation if their AI systems are used by EU residents or organisations.
Partially. GDPR Article 25 (data protection by design) and Article 5 (data minimization) support EU AI Act Article 10 compliance. However, the EU AI Act adds additional requirements: technical documentation (Article 11), conformity assessment (Article 43), and post-market monitoring (Article 72) that go beyond GDPR requirements. GDPR-grade anonymization supports the data governance requirements of both frameworks, but does not by itself establish compliance with either.
Use piisafe.eu to scan documentation repositories and web-based training data (free interface; PII detection requires an anonym.legal API key). Use anonym.legal's REST API to scan file batches programmatically. Both generate audit trail exports (HTML, JSON, CSV) suitable for Article 11 technical documentation packages.
De-identified data that meets the “no reasonable means of re-identification” standard (GDPR Recital 26) falls outside personal data definitions and thus outside GDPR and EU AI Act personal data restrictions. Article 10’s data governance requirements (relevance, representativeness, bias examination) still apply to anonymized training data. Document your anonymization methodology to demonstrate compliance.
Yes. RAG (Retrieval-Augmented Generation) systems that retrieve and process personal data in inference pipelines may require PII protection measures, particularly for high-risk AI applications. GDPR obligations for processing personal data apply to RAG retrieval, and the EU AI Act adds risk management requirements.
The EU AI Act entered into force on August 1, 2024. The phased timeline: Prohibited AI practices — February 2, 2025. GPAI model obligations — August 2, 2025. Article 50 transparency obligations — August 2, 2026. Following the 2026 Digital Omnibus amendment: high-risk AI system requirements (Annex III) — December 2, 2027; high-risk AI in Annex I (product safety legislation) — August 2, 2028.
Yes. General-purpose AI models (GPAIs), including LLMs, are subject to transparency obligations from August 2025: publishing training data summaries, copyright compliance policy, and energy consumption information. Models with systemic risk (defined as trained with more than 10²⁵ FLOPs) face additional adversarial testing and incident reporting requirements. Application developers deploying GPAIs in high-risk use cases must comply with high-risk provisions.
The replace, redact, mask and hash methods are deterministic — the same input and settings produce the same output; encrypt uses a random initialization vector, so its ciphertext differs on each run. This enables precise methodology statements such as: “Training data anonymized with anonym.legal (hybrid NLP + pattern engine, 273 entity types across 48 languages), running on ISO 27001-certified Hetzner infrastructure in Falkenstein, Germany.”
Yes — piisafe.eu provides a website scanner with a free interface that checks publicly accessible documentation, README files, model cards, and data sheets for exposed PII. PII detection requires an anonym.legal API key (from €3/month incl. VAT). Select the “EU AI Act Article 10” compliance preset for targeted scanning relevant to Article 10 data governance requirements.
Does this apply to your organisation?
The guide is general; your obligation is not. Tell us the sector and where the data is allowed to go, and we will tell you what applies.