The Clock is Ticking: December 2, 2027
The EU AI Act entered into force on August 1, 2024. It applies in stages: the prohibitions since February 2, 2025, the general-purpose AI rules since August 2, 2025, and the Article 50 transparency duties since August 2, 2026. After the Digital Omnibus, the requirements for high-risk AI systems apply from December 2, 2027 for Annex III systems and from August 2, 2028 for Annex I systems. From those dates, national market surveillance authorities can investigate, fine, and require withdrawal of non-compliant AI systems from the EU market.
The penalties are structured in three tiers:
| Violation Type | Fixed Amount | % Global Turnover |
|---|---|---|
| Prohibited AI practices (e.g., social scoring, real-time biometric surveillance) | €35,000,000 | 7% |
| High-risk AI system non-compliance (Article 10, 11, etc.) | €15,000,000 | 3% |
| Misleading or incomplete information to authorities | €7,500,000 | 1% |
The higher of the fixed amount or percentage applies; for SMEs and start-ups, the lower of the two. For a mid-sized company with €500M in global revenue, a high-risk non-compliance finding could mean €15M in fines. For a Fortune 500, it could exceed €1 billion.
Who Is Affected?
The EU AI Act applies to any organization that:
- Places AI systems on the EU market, or
- Puts AI systems into service in the EU, or
- Is a provider, deployer, importer, or distributor of AI systems used by EU residents
Crucially, this applies regardless of where the company is headquartered. US, UK, Canadian, and Asian companies deploying AI to European customers are fully subject to the EU AI Act. The extraterritorial test is market-based, not GDPR’s data-subject test: providers and deployers are in scope, wherever they are established, if they place an AI system on the EU market, put it into service in the EU, or if the system’s output is used in the EU.
High-risk AI systems specifically covered by Annex III include: biometric identification, safety-critical infrastructure management (energy, water, transport), educational and vocational training scoring, employment recruitment and HR screening, credit and insurance risk scoring, law enforcement and predictive policing, migration and asylum processing, and administration of justice.
Article 10 — What It Says About PII
Article 10 is the provision that most directly affects data teams building AI training pipelines. It establishes data governance requirements for the training, validation, and testing datasets used in high-risk AI systems.
“Training, validation and testing data sets shall be subject to data governance and management practices appropriate for the intended purpose … 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(2) and Article 10(3)
Article 10 does not require training data to be free of personal data. Anonymization is a safeguard you can choose, not one the Act prescribes. Article 4a (inserted by Regulation (EU) 2026/1744, which replaced the former Article 10(5)) points the same way: special categories of personal data may be processed for bias detection only where this “cannot be effectively fulfilled by processing other data, including synthetic or anonymised data.” Where personal data remains in a training set, GDPR continues to apply to it.
The 5 PII Actions You Must Take Before the Deadline
Audit Your Training Data Repositories
Before you can anonymize, you need to know what you have. Conduct a full inventory of every training dataset: where it came from, what personal data it contains, and how it was sourced. Use piisafe.eu to scan publicly accessible documentation pages and model cards for exposed PII (the scanner interface is free; PII detection needs an anonym.legal API key from €3/month). Use anonym.legal's REST API to batch-scan file-based datasets.
Document your findings. You will need this audit as part of your Article 11 technical documentation, and regulators will ask for it during conformity assessments.
Anonymize Personal Data in Training Corpora
Apply systematic anonymization to all training data that contains personal information. For NLP training data, this means entity-level anonymization: replacing names, email addresses, phone numbers, national IDs, and other PII with consistent placeholder tokens or synthetic replacements.
The critical requirement is consistency: if “John Smith” appears 47 times across your training corpus, every occurrence must be replaced with the same pseudonym. Inconsistent anonymization corrupts the statistical relationships your model needs to learn.
Document Your Anonymization Methodology
Article 11 requires technical documentation that includes your data governance practices (Article 10) as a component. Your documentation must be specific enough that a national authority could assess your compliance. Vague statements like “we anonymized the data” are insufficient.
A compliant methodology statement might read: “Training data anonymized using anonym.legal (Presidio-based detection, 267 entity types, 48 languages). Entity types targeted: [list]. Processing runs on ISO 27001-certified Hetzner infrastructure in Falkenstein. Text processed in memory, with history turned off.”
Protect Inference Pipelines
Article 10 focuses on training data, but GDPR continues to apply to inference-time processing. If your high-risk AI system processes personal data during inference (e.g., a medical AI analyzing patient records, or an HR AI screening CVs), GDPR Article 25 (data protection by design) requires technical measures to minimize personal data processing.
For RAG pipelines that retrieve documents containing PII before passing them to an LLM, consistent pseudonymization ensures the LLM never sees raw personal data. anonymize.solutions supports symmetric de-anonymization: you can pseudonymize before LLM inference and re-identify the response to deliver personalized results to authorized users.
Establish Ongoing Monitoring
EU AI Act Article 72 requires post-market monitoring for high-risk AI systems. Your data governance obligations do not end at deployment. Every new batch of training data added during fine-tuning or retraining must pass through the same anonymization pipeline. Integrate PII scanning into your MLOps pipeline as a quality gate: no training batch should reach your training environment without a clean PII scan report.
How to Audit Your Public AI Documentation
Model cards, README files, data sheets, and technical documentation are frequently overlooked sources of PII exposure. Researchers often include real example outputs containing personal data, or reference real dataset samples in their documentation.
Use piisafe.eu — a website PII scanner with a free interface — to check your publicly accessible AI documentation. PII detection needs an anonym.legal API key from €3/month. Keep the scan results with your technical documentation package.
The Technical Documentation Checklist (Article 11)
Article 11 and Annex IV specify the technical documentation requirements for high-risk AI systems. Before your high-risk obligations apply (December 2, 2027 for Annex III systems), you should have all of the following prepared and available for regulatory review:
- General description of the AI system: intended purpose, version history, deployment context
- Design and architecture: system components, data flow diagrams, training infrastructure
- Training data governance (Article 10): sources, preprocessing steps, PII anonymization methodology, bias assessment
- Validation and testing: accuracy metrics, performance across demographic groups, evaluation datasets
- Risk management (Article 9): identified risks, mitigation measures, residual risk assessment
- Post-market monitoring plan: metrics monitored, incident escalation procedure, update cycle
GDPR + EU AI Act: Double Compliance
Many organizations treating the EU AI Act as a separate compliance project are missing an efficiency opportunity. Anonymization work done for GDPR carries over to the EU AI Act:
- GDPR Recital 26 takes data outside the GDPR once it is rendered anonymous, judged by “all the means reasonably likely to be used” to identify a person.
- EU AI Act Article 4a (formerly Article 10(5)) names “synthetic or anonymised data” as the first choice before special categories of personal data may be processed for bias detection.
GDPR-standard anonymization supports your Article 10 data governance, but it does not fulfil Article 10 on its own: the Article also requires documented governance practices, checks for relevance and representativeness, and examination for errors and bias. Organizations already running GDPR anonymization programs can reuse that work as one part of it.
Penalties Are Not Theoretical
Some compliance teams are treating the EU AI Act as a future concern, expecting enforcement to be soft in the early years. This is a dangerous assumption. EU data protection authorities demonstrated with GDPR that they are prepared to impose large fines, and that the fines grew over time: Google €50 million (CNIL, 2019), Amazon €746 million (Luxembourg CNPD, 2021), Meta €1.2 billion (Irish DPC, 2023).
The EU AI Act creates dedicated national market surveillance authorities in each member state specifically tasked with enforcement. The European AI Office coordinates cross-border enforcement. Unlike early GDPR enforcement which focused on data breaches, EU AI Act enforcement will also be triggered by civil society complaints, whistleblowers, and competitor filings — any of which could put your AI system under scrutiny before you expect it.
Conclusion: Start Now, Not at the Last Minute
With the enforcement deadline approaching, the window for action is narrowing. A full training data audit for a mid-sized ML team takes 2–4 weeks. Anonymization pipeline implementation takes another 1–2 weeks. Documentation preparation takes a week. Technical documentation review by legal counsel adds another week. Added up, that is an estimate of roughly 5–8 weeks; the actual duration depends on data volume, the system’s risk classification and how deep the legal review goes.
Free first step: Use piisafe.eu to scan your public AI documentation today — the interface is free; PII detection needs an anonym.legal API key from €3/month. Then contact our team to discuss implementing an anonymization pipeline for your training data before the deadline.