Navigating High-Risk AI Classification Across Jurisdictions
Global manufacturer's approach to Article 6 high-risk AI classification across 8 EU member states
Overview
A EUR 2.3B multinational manufacturer deployed predictive maintenance AI across European facilities. The EU AI Act's December 2024 enforcement created classification uncertainty: Is predictive maintenance 'high-risk' under Article 6? Which requirements apply across German, Spanish, Polish, and 5 other jurisdictions? This case explores the classification methodology and implementation strategy.
Background
The manufacturer operated production facilities across Germany, Poland, Spain, and distributed products to 12 EU countries. Predictive maintenance AI systems reduced unplanned downtime by 22% and maintenance costs by EUR 18M annually. However, the EU AI Act created regulatory ambiguity: the Act lists 'safety-critical production planning' as potentially high-risk, but predictive maintenance's classification was unclear. The organization faced a regulatory timeline: Article 6 high-risk systems require technical documentation (Article 10), transparency information (Article 13), and human oversight protocols (Article 14) by December 2024. Delaying compliance risked operational shutdown of AI systems. Yet implementing overly strict controls might waste resources on unnecessary requirements.
The Challenge
The classification challenge cascaded across multiple dimensions: 1. Article 6 Interpretation: Is predictive maintenance 'high-risk'? The regulation lists "substantial impact on safety and legal rights" but doesn't explicitly cover manufacturing systems. Different member states interpreted Article 6 differently. 2. Scope Ambiguity: The organization ran 12 AI systems (forecasting, quality control, supply chain, maintenance). Which of these triggered Article 6? Were they all high-risk, or only some? 3. Multi-Jurisdiction Fragmentation: Polish regulators were conservative; Spanish were pragmatic. Germany had additional factory-level safety requirements (ISO 26262). A one-size-fits-all compliance framework wouldn't survive scrutiny. 4. Technical Documentation Burden: Article 10 requires exhaustive technical dossiers for high-risk systems. Creating these for even 3 systems would require 6-8 months of engineering effort. 5. Human Oversight Design: Article 14 requires human review of high-risk AI decisions. For predictive maintenance, what should human operators actually review? When? How often? The organization had 9 months before Article 6 enforcement. Starting with correct classification was critical—misclassification could invalidate months of work.
Approach
A methodical approach to Article 6 classification and tiered compliance: Workstream 1: Systematic Classification (Months 1-2) Mapped 12 AI systems against explicit Article 6 criteria. The analysis determined: - Predictive maintenance = High-risk (safety-critical production planning) - Quality control = Medium-risk (affects product safety but not production planning) - Forecasting = Low-risk (commercial decision, not safety-critical) - Other systems: Variable based on output impact Result: 3 systems required full Article 10-14 compliance; 9 systems required simplified requirements. Workstream 2: Technical Documentation (Months 2-5) For the 3 high-risk systems, created Article 10 compliance dossiers covering: - Training data lineage (what data trained the model?) - Model validation methodology (how was it tested?) - Performance metrics (what accuracy/reliability was achieved?) - Known limitations (what scenarios might it fail in?) - Update protocols (how is the model maintained?) Integrated documentation with existing ISO 9001 quality management systems rather than creating parallel processes. Workstream 3: Operator-Centric Transparency (Months 5-7) Article 13 requires technical information disclosure to operators. The organization created: - Plain-language explanations (not technical documentation) - Transparency dashboard showing model confidence scores - Clear explanation of recommendations vs. mandatory decisions - Maintenance protocols when model confidence is low Workstream 4: Multi-Jurisdiction Harmonization (Months 7-9) Defined compliance standards that exceeded all 8 member state requirements. Pre-coordinated with German BaFin and Spanish AEPD regulators before full deployment.
Outcomes
Full Article 6-14 compliance
Zero violations in audit
EUR 400M contract enabled
Regulatory Achievements: ✓ Full Article 6 classification clarity across 8 jurisdictions ✓ Zero compliance violations in regulatory audit (June 2024) ✓ 12 manufacturing AI systems deployed without legal holds ✓ All Article 10, 13, 14 requirements implemented Operational Metrics: ✓ 98.3% operator comprehension on Article 13 transparency information (vs. typical 60-70%) ✓ Human override response time: <2 minutes (exceeding Article 14 standards) ✓ Documentation completeness: 100% of Article 10 technical requirements ✓ Post-deployment compliance audits: All passed Business Impact: ✓ Zero operational disruption to manufacturing AI systems ✓ Enabled EUR 400M+ new contract (Middle East buyer cited EU AI Act compliance as selection factor) ✓ Established marketplace positioning as "EU AI Act compliant manufacturer"
Impact
The classification and compliance approach enabled: - Confident market expansion across 8 EU jurisdictions - Competitive advantage in government procurement (compliance demonstrated) - Template for applying Article 6 classification logic to other high-risk AI systems - Foundation for operating in multiple regulatory regimes simultaneously
Key Insights
1. Article 6 Classification Drives Everything: Get classification wrong, and downstream documentation/transparency/oversight requirements cascade incorrectly. The early classification work (Months 1-2) was 20% of effort but determined 80% of compliance burden. 2. Technical Documentation Must Integrate with Existing Systems: Organizations can't maintain two parallel quality management systems. Integrating Article 10 documentation with ISO 9001 reduced duplication and improved compliance rigor. 3. Transparency Requires Audience Adaptation: Regulators, operators, and product managers need different information. Article 13 technical documentation must serve factory operators (non-technical), not regulators. 4. Human Oversight Needs Infrastructure: Article 14 human review requirements aren't policy documents—they require decision tools, operator training, and documented response protocols. Policy alone fails in practice.
Sectors
- Manufacturing
- Industrial Technology
- Automotive & Machinery
Techniques
- Article 6 Risk Classification
- Technical Documentation (Article 10)
- Transparency Protocols (Article 13)
- Human Oversight Design (Article 14)
Related
- EU AI Act Risk Assessment
- Technical Documentation
- Regulatory Mapping
- Compliance Auditing
Explore More
Discover how we approach AI governance, compliance, and risk across global jurisdictions.