EU AI Act enters high risk enforcement phase, putting enterprise AI under immediate scrutiny this week
Enzo Reno ·
European regulators are moving into active oversight of high risk AI systems this week across the European Union, a shift tied to the EU AI Act provisions that became applicable 24 months after entry into force. Enterprises in banking, healthcare, and the public sector face requests for technical documentation and registration in the new EU database as national authorities and the European Commission’s AI Office coordinate first round checks.
This matters now because failure to meet conformity requirements, including risk management, data governance, logging, and human oversight, carries penalties of up to 35 million euros or 7 percent of global annual turnover, according to the Regulation published in the Official Journal of the European Union in 2024.
What changes now
High risk categories include credit scoring and access to essential services, recruitment and worker management, medical devices, critical infrastructure, and certain public sector uses. Providers must complete conformity assessments, apply CE marking, and register systems in the Commission’s database before placing them on the EU market, per the Commission’s implementation guidance and the final text.
Procurement and compliance leaders will need updated assurances from cloud and model vendors. Microsoft says Copilot for Microsoft 365 adheres to the EU Data Boundary alongside Azure AI services, expanding commitments first laid out for Microsoft Cloud in January 2023, according to Microsoft documentation. Google states Vertex AI provides audit logs, model cards, and data residency options in Europe, and AWS highlights Bedrock Guardrails and regional isolation for customer content in the EU, per company materials.
Model selection will be scrutinized. Enterprises building on closed and open models must document capabilities and safeguards, then map them to AI Act controls. Meta released Llama 3 in April 2024 under an open license with published model cards, Google introduced Gemini 1.5 with long context in February 2024, and Anthropic announced the Claude 3 family in March 2024, each with varying transparency and evaluation artifacts cited in provider reports.
What to watch this week
Expect early spot checks on documentation completeness and registration status, plus sector regulators coordinating with national market surveillance authorities on priority use cases. Gartner has forecast that by 2026 more than 80 percent of enterprises will use generative AI APIs or models in production, up from less than 5 percent in 2023, raising the volume of deployments that fall into or near high risk categories.
For the broader technology landscape, this week signals a budget shift from pilots to assurance. Tooling for model evaluations, lineage, and policy controls is set to gain urgency, a space already drawing capital as seen with Arthur’s 42 million dollar Series B in 2021 and Protect AI’s 35 million dollar round in 2023, reported by TechCrunch and the companies, as enterprises recalibrate AI roadmaps around regulatory reality.