Brussels Switches On the AI Act: What 2 August 2026 Actually Changes

The Enforcement Clock: What Binds in 2026 #

Regulation 2024/1689 is not a single switch but a staircase. The text was published in the Official Journal on 12 July 2024 and entered force on 1 August 2024. The first tranche, the Article 5 prohibitions covering social scoring, manipulative subliminal techniques, real-time remote biometric identification in public spaces with narrow exceptions, and the untargeted scraping of facial images for recognition databases, applied from 2 February 2025. Most enterprise legal teams treated that date as preparation rather than activation, because few production systems sat on the wrong side of those lines.

The 2 August 2025 milestone landed harder. Chapter V brought general-purpose AI model providers under transparency, copyright, and technical documentation obligations, with deeper systemic-risk duties for any model trained above 10 to the 25 floating point operations. That threshold currently captures GPT class models from OpenAI, Google's Gemini family, Anthropic's Claude family, Meta's Llama frontier models, and Mistral's Large series, per the AI Office's published methodology in its 2025 implementation guidance. Providers below that compute line still face the baseline transparency duties, which is why Hugging Face and several open-weights labs negotiated a tailored disclosure template through summer 2025.

The 2 August 2026 threshold is the structural one. Annex III high-risk systems, which include AI used in employment decisions, education and exam scoring, credit scoring, critical infrastructure, biometric categorisation, migration and border control, and the administration of justice, become subject to the full conformity assessment regime. CE marking, registration in the EU database, post-market monitoring, human oversight architectures, and fundamental rights impact assessments for public-sector deployers all become legally enforceable. The Commission has confirmed in its February 2026 implementation update that Article 6 implementing acts on conformity assessment procedures will be finalised before the August deadline.

The AI Office and the Code of Practice #

The European AI Office sits inside DG CNECT and is the operational core of the regime. The Commission's recruitment notices through 2025 targeted approximately 140 staff by the end of 2026, drawing technical reviewers, policy officers, legal counsel, and seconded national experts. Roughly half of those positions were filled by Q1 2026, per the AI Office's quarterly staffing update. The Office is the sole enforcer for GPAI obligations, sets the technical standards baseline, and coordinates the European AI Board of national authorities. It does not, however, supervise individual high-risk deployments, which remain a national competency.

The General-Purpose AI Code of Practice was published on 10 July 2025 after a multi-stakeholder drafting process led by thirteen independent chairs and vice-chairs, including Yoshua Bengio on the systemic risk chapter. Initial signatories include OpenAI, Google, Anthropic, Microsoft, Meta, Mistral, and Cohere, with Amazon and IBM signing in the autumn. Meta initially declined to sign the full code, citing concerns about overlap with the DSA and copyright provisions, then reversed in October 2025 after the Commission clarified the safe-harbour effect of signature. Adherence is voluntary but creates a presumption of conformity, which is the only practical route to legal certainty before harmonised standards arrive in 2027.

The Code has three sections: transparency, copyright, and safety and security for systemic-risk models. The safety chapter requires evaluations against a defined taxonomy of catastrophic and offensive cyber risks, model cards with capability and limitation summaries, incident reporting within fifteen days, and serious incident reporting within five. Athena, our in-house regulatory and litigation platform, tracks each signatory's evaluation cadence against the Code's commitments, and the gap between what is published and what is required has narrowed sharply since November 2025.

Penalty Math and the First Wave of Investigations #

The fine architecture is steeper than the GDPR. Article 99 sets three tiers. Prohibited practices under Article 5 carry penalties up to EUR 35 million or 7 percent of worldwide annual turnover, whichever is higher. Most other substantive violations, including the GPAI obligations under Article 53, the high-risk obligations, and breaches by notified bodies, attract up to EUR 15 million or 3 percent. Supplying incorrect, incomplete, or misleading information to authorities is capped at EUR 7.5 million or 1 percent. SMEs and start-ups face the lower of the two figures rather than the higher, which is the only meaningful concession in the penalty design.

The AI Office confirmed in its January 2026 work programme that the first GPAI compliance audits would open in Q1 2026, focused on systemic-risk models. Reuters reported on 11 February 2026 that the Office had sent formal information requests to a small number of frontier providers covering training data summaries, copyright opt-out compliance, and red-team evaluation methodology. These are not yet enforcement proceedings. They are the discovery phase that precedes any decision under Article 75. The Commission has been clear that the goal in 2026 is calibration, not penalty maximisation, and that fines on the upper tier are unlikely before the second half of 2027.

For non-EU providers, the extraterritorial scope under Article 2 is the operative concern. A US foundation-model lab whose model is placed on the EU market, or whose output is used in the EU, falls within scope regardless of physical presence. The mandatory authorised representative requirement under Article 22 means every non-EU provider must designate an EU-established legal entity that holds the technical documentation and answers to the AI Office. Strategos, our geopolitical and regulatory strategy desk, has flagged the authorised representative regime as the practical pinch point for Asian providers entering the EU through 2026.

Member-State Divergence: NCAs and Sandboxes #

The Act requires each member state to designate at least one notifying authority and at least one market surveillance authority, with the option to consolidate them. By the Article 70 deadline of 2 August 2025, fifteen member states had formally designated their national competent authorities. By April 2026 the count is at twenty-three, with three still in advanced legislative process and one in political deadlock. Spain moved earliest, establishing AESIA in Coruna in 2023, well before the regulation entered force. France routes supervision through CNIL with technical support from PEReN. Germany has assigned market surveillance to the BNetzA but kept fundamental rights oversight at the Bundesbeauftragte fur Datenschutz, which is a coordination structure that legal observers expect to be tested early.

Article 57 requires each member state to operate at least one regulatory sandbox by 2 August 2026. Sandboxes provide a controlled environment for testing high-risk systems before market placement, with regulatory guidance and a degree of legal certainty. As of April 2026, eighteen member states have functioning or beta sandboxes, six have announced launch dates before August, and three remain in procurement. The quality varies. Spain and Denmark run technically rigorous programmes with dedicated staff. Several Eastern member states have published frameworks but have not yet seated technical reviewers, which means a paper sandbox without an operational one. The Argus monitor we maintain on EU regulatory infrastructure tracks sandbox staffing and live cohort numbers each quarter.

Convergence is the explicit Commission goal but not the realistic 2026 outcome. Coordination through the AI Board is improving, yet the texture of supervision will diverge for at least three years. Forum-shopping risks are mostly theoretical, because high-risk systems must be assessed where they are placed on the market, but the cost of compliance varies materially by jurisdiction.

The Draghi Tension: Competitiveness Versus Precaution #

Mario Draghi's report on the future of European competitiveness, published on 9 September 2024, treated the AI Act as a partial answer to a problem and a partial cause of another. The report's central finding, that the EU faces a productivity gap of roughly 30 percent against the United States that is driven by tech sector underperformance, framed regulation not as a discrete policy choice but as a binding capacity constraint. Draghi did not call for the AI Act to be reopened. He called for its implementation to avoid layering compliance costs on top of GDPR, the DSA, the DMA, NIS2, the Data Act, and forthcoming product liability changes, where the cumulative burden, not any single rule, becomes the impediment.

The Commission has internalised the message at the rhetorical level. The Competitiveness Compass published in January 2025 and the AI Continent Action Plan published in April 2025 both promise simplification, faster sandbox throughput, and a single point of contact for SMEs through the AI Office. Whether implementation matches the promise will be visible in two metrics by year-end: the number of high-risk systems that secure CE marking before December 2026, and the average elapsed time from sandbox application to operational testing. Both metrics are tracked in the Commission's quarterly AI Act dashboard.

The political risk is a divergence between the soft-law messaging of simplification and the hard-law architecture of penalties. If the AI Office opens proceedings against European providers more visibly than against US labs, the perception of asymmetric burden will harden, and the Council will face pressure to amend the regulation in 2027.

Implications for Labs, Deployers, and EU Corporates #

For foundation-model labs, signing the Code of Practice is now table stakes for EU market access. The presumption of conformity is the only path to legal certainty until harmonised standards under Article 40 are published, which the JTC 21 work programme places in 2027 at the earliest. Non-signatories face a default position of having to demonstrate compliance directly to the AI Office, with no procedural shortcut. The compute threshold of 10 to the 25 FLOPs is binding even for models trained outside the EU, and the systemic-risk designation triggers the full safety and security chapter, not just transparency.

For high-risk system deployers, the operational burden lands in 2026 in three places. First, the fundamental rights impact assessment under Article 27 applies to public bodies and to private deployers in specific sectors such as banking and insurance, and it must be completed before deployment. Second, the human oversight architecture under Article 14 must be documented, not merely asserted, with clear protocols for the natural persons assigned to oversee the system. Third, post-market monitoring under Article 72 requires a documented plan and incident logs that are auditable on request. Most enterprise AI programmes we see are still in policy-drafting mode rather than operational mode.

For EU corporates that procure rather than build, the practical question is contractual flow-down. The provider must hand over technical documentation sufficient for the deployer to discharge its own obligations, which means master service agreements signed in 2024 or earlier almost certainly need amendment. Procurement teams should treat 2 August 2026 as a hard contractual deadline, not a negotiation horizon.

Sources #

• EUR-Lex, Regulation (EU) 2024/1689 of the European Parliament and of the Council
• European Commission, AI Office
• European Commission, General-Purpose AI Code of Practice
• Mario Draghi, The Future of European Competitiveness, September 2024
• European Commission, AI Continent Action Plan, April 2025
• European Commission, Competitiveness Compass, January 2025
• Politico Europe, AI Act Tracker
• Reuters, EU launches first AI Act compliance probes, February 2026
• Bird and Bird, EU AI Act Tracker
• Wilson Sonsini, EU AI Act Resource Center
• AESIA, Spanish Agency for the Supervision of Artificial Intelligence
• CEN-CENELEC JTC 21, AI standardisation work programme