Europe Is Turning the AI Act Into Industrial Policy
Europe’s AI Act was sold as a risk framework. Its latest political repair job looks more like industrial policy.
On May 7, the European Commission welcomed a provisional deal between Parliament and Council on the Digital Omnibus on AI, describing it as part of the EU’s simplification agenda to boost competitiveness and make AI Act implementation easier for businesses while keeping safety and rights protections intact. The important part is not the press-release grammar. It is the sequencing. High-risk rules for standalone systems in areas such as biometrics, critical infrastructure, education, employment, migration, asylum and border control would now apply from December 2, 2027. Rules for systems embedded in products such as lifts and toys would apply from August 2, 2028, according to the Commission’s summary.
That is a major shift from the original August 2026 cliff for many obligations. It is also an admission that rulemaking and implementation infrastructure are different jobs.
The Act still matters. Prohibited practices and AI literacy obligations have already started applying. General-purpose model obligations started in 2025. Transparency rules remain live on the 2026 track. But the heaviest high-risk compliance machinery is being moved closer to the point where standards, guidance, sandboxes and national enforcement bodies may actually exist.
That is not deregulation in the simple sense. It is a bet that a slower, more engineered rollout will make the AI Act more enforceable and more tolerable to European industry.
The Compliance Layer Was Not Ready
The Commission’s own AI Act FAQ is blunt about the problem. High-risk AI systems are supposed to rely on European harmonised standards, because those standards turn broad legal duties into auditable engineering practice. Providers that build to the standards get a presumption of conformity. That is the whole point of the machine.
The standards are late. The Commission says CEN and CENELEC were not able to develop them on the requested August 2025 timeline and that the work is still ongoing. It also says the delayed standards put the smooth entry into application of high-risk rules on August 2, 2026 at risk, according to its Navigating the AI Act FAQ. That is the practical center of the pivot, not a sudden change in European values. A law that requires technical evidence before the evidence template exists is not tough regulation. It is paperwork cosplay with fines attached.
The May deal tries to remove that mismatch. The Council’s account says the agreement reduces recurring administrative costs, improves legal certainty and supports competitiveness. It also postpones the deadline for competent authorities to establish AI regulatory sandboxes at national level until August 2, 2027. That matters because sandboxes are where regulators can learn from live systems before enforcement becomes a guessing contest.
The deal also narrows one of the uglier overlaps in the Act: industrial AI inside already regulated products. Council says co-legislators found a mechanism to resolve cases where sectoral law has similar AI-specific requirements, including medical devices, toys, lifts, machinery and watercraft. In some cases, implementing acts can limit direct AI Act application. Machinery gets a more explicit workaround through delegated acts under the machinery regulation. The goal is clear: stop making a factory automation provider prove the same safety case twice in two slightly different dialects.
That is the competitiveness pivot in concrete form. Europe is not abandoning the risk model. It is trying to attach the risk model to product regulation, standards and administrative capacity.
Relief, Then More Ambiguity
Builders see relief. They also see a half-finished map.
Germany’s Bitkom called the compromise a good signal for industrial AI in Europe, especially AI applications in production and mechanical engineering. It said the deal avoids duplicate testing and documentation processes and lets companies put more resources into innovation. It also welcomed the high-risk deadline shift as a source of legal certainty. But Bitkom’s praise stopped there. The group said the machinery approach should be extended to other heavily regulated sectors, including medical devices and radio equipment, and argued that deadlines for transparency obligations on generative AI deployers also need to move because key interpretive guidance is expected only shortly before the rules apply.
ITI, the global tech trade association, struck the same note in English. It welcomed the delay of key high-risk obligations and the attempt to address duplicative industrial AI regulation, then said the agreement still does not solve the complexity of the EU framework. Its unresolved list is telling: watermarking and labelling deadlines, overlap between horizontal and vertical rules, Digital Omnibus work, AI Act guidelines and codes of practice.
So the industry read is not “Europe blinked.” It is closer to “Europe noticed the build system was failing, but the dependency graph is still messy.”
That distinction matters. The EU’s problem was never only that companies dislike compliance costs. They always do. The deeper problem was that the Act’s implementation calendar assumed a mature standards and enforcement stack. In AI, that stack is being written while the systems are being deployed.
The Commission’s general-purpose AI Code of Practice shows the more workable model. The code is voluntary, built through a multi-stakeholder process, and providers that sign it can use it to show compliance. The Commission says this reduces administrative burden and improves legal certainty compared with proving compliance through alternative means. That is a softer instrument than a statutory duty, but it is also closer to how fast technical markets actually move.
The Real Test Is Whether Europe Can Ship Governance
The AI Act competitiveness pivot should be judged by execution, not ideology.
If the delay buys time for harmonised standards, usable guidance, regulator staffing and sector-specific integration, Europe gets a stronger law. Companies can build compliance into product development instead of hiring lawyers to reverse-engineer duties from late guidance. Regulators can enforce against known patterns instead of improvising. European startups and industrial firms get a clearer path to sell compliant systems into regulated markets. Boring, expensive, useful. The best kind of infrastructure.
If the delay becomes a habit, Europe gets the worst version of both worlds. The AI Act remains a heavy symbolic statute, but the hard parts keep sliding while builders still cannot tell which obligations will survive contact with omnibus politics. That would not improve competitiveness. It would turn regulatory uncertainty into a permanent tax.
The May 7 agreement is therefore not a retreat from the AI Act. It is a conversion attempt. Brussels is trying to turn a rights-first legal framework into a market-operating system for AI deployment. The uncomfortable part is that market-operating systems need releases, documentation, compatibility layers and maintainers. The EU has the rulebook. Now it has to prove it can ship the runtime.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.