EU AI Act Omnibus Agreement: What Changed

Enforcement delayed to December 2, 2027

High-risk AI systems in the following categories received a 16-month extension:

• Biometric identification and categorization (Annex III, category 1)
• Critical infrastructure management (category 2)
• Education and vocational training (category 3)
• Employment, worker management, access to self-employment (category 4)
• Access to essential private and public services (category 5)
• Law enforcement (category 6)
• Migration, asylum, and border control (category 7)
• Administration of justice and democratic processes (category 8)

Practical impact: Providers of standalone high-risk AI systems now have until December 2, 2027 to complete conformity assessments, register in the EU database, implement risk management systems (Art. 9), and establish post-market monitoring (Art. 72).

Enforcement delayed to August 2, 2028

AI systems embedded in products covered by existing EU product safety legislation received a 24-month extension. Affected product categories include:

• Machinery and lifts (Machinery Regulation)
• Toys (Toy Safety Directive)
• Radio equipment (Radio Equipment Directive)
• Medical devices and in-vitro diagnostics
• Civil aviation safety
• Motor vehicles and their components
• Marine equipment

Rationale: The extended timeline aligns AI Act enforcement with ongoing revisions to EU product safety regulations, preventing duplication between sectoral and AI-specific requirements. The Omnibus also clarified the interplay between the AI Act and the Machinery Regulation.

AI nudification and CSAM generation banned

The Omnibus adds a new prohibited practice: AI systems that generate non-consensual sexually explicit or intimate content, including child sexual abuse material. This prohibition takes effect upon formal adoption of the Omnibus (expected Q3-Q4 2026).

SWT3 relevance: AI-GRD.1 and AI-GRD.2 witness guardrail presence and content safety filter activation. Organizations deploying generative AI should ensure content safety filters are active and witnessed.

Simplified requirements extended to small mid-cap companies

Regulatory relief previously available only to small and medium-sized enterprises (SMEs) now extends to small mid-cap companies. This includes reduced technical documentation requirements and expanded access to regulatory sandboxes.

Important: Substantive compliance requirements for deployed high-risk systems are not reduced. The relief applies to documentation burden and sandbox access, not to the safety and transparency requirements themselves.

Centralized oversight for GPAI models

The Omnibus reinforces the European AI Office's authority over GPAI model oversight. The AI Office can request documentation directly from GPAI providers and coordinate enforcement across Member States. This reduces governance fragmentation -- providers deal with one central authority for GPAI compliance rather than 27 national authorities.

SWT3 relevance: The export_evidence() method and the auditor portal support on-demand documentation requests. When the AI Office requests evidence, the same export pipeline used for Notified Body assessments applies.

New EU-level sandbox and expanded access

The Omnibus creates a new EU-level regulatory sandbox in addition to existing national sandboxes. More innovators gain participation rights, allowing them to test AI solutions in real-world conditions with structured compliance guidance from regulators.

SWT3 relevance: Clearing level 0 (Analytics) is designed for sandbox and pre-deployment environments where full transparency is acceptable. Organizations can adopt SWT3 in sandbox mode with zero cost (Open tier) and graduate to production clearing levels when deploying.

Prohibited AI Practices (Article 5)

The following remain banned and enforceable: social scoring systems, real-time remote biometric identification in public spaces (with narrow law enforcement exceptions), emotion recognition in workplace and education contexts, untargeted facial image scraping for database creation, and subliminal manipulation techniques.

SWT3 procedures: AI-FAIR.1, AI-FAIR.2, AI-FAIR.3 witness bias assessment and fairness metrics.

General GPAI Model Obligations (Article 53)

GPAI model providers must maintain and provide: model card documentation, training data summaries and copyright compliance evidence, and technical documentation per Annex XI. These obligations are active now.

SWT3 procedures: AI-MDL.1, AI-MDL.2, AI-DATA.1, AI-DATA.2, AI-DATA.3, AI-DATA.4.

GPAI Transparency Obligations (Articles 50, 53)

AI-generated content marking (Art. 50), Code of Practice compliance (Art. 53), and systemic risk provisions (Art. 55) become enforceable. The GPAI Code of Practice final draft is expected June 2026.

SWT3 procedures: AI-INF.1, AI-INF.2, AI-MARK.1, AI-MDL.5, AI-MDL.6, AI-MDL.7, AI-SEC.1, AI-GRD.1, AI-GRD.2, AI-GRD.3.

See GPAI Code of Practice Mapping for detailed procedure-to-obligation analysis.

More Time Does Not Mean Start Later

The 16-24 month extension for high-risk systems benefits organizations that start now, not organizations that wait. SWT3 continuous witnessing builds a compliance history over time. Every inference witnessed today becomes part of the evidence record for a future conformity assessment.

An organization that begins witnessing in June 2026 will have 18 months of continuous evidence by the December 2027 deadline. An organization that begins witnessing in November 2027 will have 30 days. Notified Bodies and the AI Office will notice the difference.

GPAI First, High-Risk Second

Organizations with both GPAI and high-risk AI obligations should sequence their compliance work:

1. Now through August 2026: GPAI transparency procedures -- AI-INF.1, AI-INF.2, AI-MDL.1, AI-GRD.1, AI-MARK.1
2. August 2026 through December 2027: High-risk system procedures -- AI-HITL.1, AI-HITL.2, AI-FAIR.1, AI-EXPL.1, AI-SEC.1, AI-DATA.1
3. Ongoing: Chain witnessing (AI-CHAIN.1, AI-CHAIN.2) for multi-system deployments

SWT3 supports incremental procedure adoption. Start with the immediate GPAI obligations and expand coverage as high-risk deadlines approach.

Embedded AI Gets Separate Treatment

Product manufacturers embedding AI have until August 2, 2028. However, if the AI component also operates as a standalone system (e.g., a medical AI that also provides general diagnostic recommendations), the December 2, 2027 deadline may apply to the standalone function.

SWT3 chain witnessing (AI-CHAIN.1, AI-CHAIN.2) proves compliance across product integration boundaries, documenting which system component triggered which inference and under which regulatory regime.