Seven percent of global turnover. Read that as a company director would. Under the European Union's Artificial Intelligence Act, the top tier of fines reaches €35 million or 7 percent of worldwide annual revenue for the most serious violations — and the percentage is calculated on the whole group's books, not on whatever the European arm happened to earn. The Act became broadly enforceable on August 2, 2026. On that date, a subject that had spent years circulating around conference panels stopped being a conversation. Governance became a compliance line item, with a budget, a deadline, and a number attached.
Europe did not act alone, and the Act is not the whole story either. Two other instruments matured alongside it. ISO/IEC 42001 — described as the first management-system standard for AI that can be certified — has now been formally adopted in Europe as well. Across the Atlantic, the U.S. National Institute of Standards and Technology stretched its AI Risk Management Framework into new territory, adding sector profiles for cybersecurity and critical infrastructure. A law, a certificate, a framework. Between them, they now set the terms for how serious organizations run AI.
A law written in risk tiers
The Act had a long runway, which is part of why it caught people out. The European Commission proposed it in 2021. A political deal arrived in late 2023, and the text was published as Regulation (EU) 2024/1689 in March 2024 — the first comprehensive AI law anywhere. Nothing about the timetable was rushed. Companies had years to prepare.
Its design is built on risk. A small set of practices classed as "unacceptable risk" faces an outright ban. "High-risk" systems — biometrics, critical infrastructure, employment, education, law enforcement and the like — carry strict obligations. For certain systems — chatbots and deepfakes included — users are owed transparency. Separate duties also fall on providers of general-purpose AI models. The obligations scale with the danger.
The switching-on happened in stages: February 2025 brought the bans on unacceptable-risk practices, August 2025 the duties for general-purpose models, and August 2, 2026 the enforcement powers and transparency rules. The heaviest load lands later: high-risk obligations — documented risk management, data governance, technical documentation, human oversight, robustness, post-market monitoring — come with key deadlines in December 2027 and August 2028, after the EU's 2026 Omnibus package pushed them back to give providers more time.
The fines deserve a second look, because the structure is the message. The €35 million or 7 percent ceiling is reserved for prohibited practices, with lower tiers beneath it. That ceiling sits above the headline maximum available under the GDPR, which held the record for European regulatory pain until now. So any company using AI in hiring, credit scoring, healthcare triage or public services inside the EU needs a data-protection-grade compliance program — and then some, because the exposure is priced off global turnover.
And the reach is wider than the map suggests. The law binds any company putting AI systems on the European market, wherever that company is headquartered. A provider on another continent with European customers is inside the perimeter, and its European deployers carry duties of their own. That is how a Brussels regulation reaches boardrooms that never did business in Europe directly — through customers, subsidiaries and supply chains.
The certificate that proves less than buyers think
ISO/IEC 42001 arrived in December 2023, published jointly by ISO and the IEC. Think of ISO 27001 or ISO 9001 reworked from the ground up for AI. It spells out how an organization should set up, operate, maintain and continually improve the way it governs AI systems. Annex A sets out the AI-specific controls: AI policy, resources, impact analysis, model lifecycle, data governance, stakeholder communication. It is, in other words, a management system — machinery for making governance a habit rather than an event.
Its spread has a simple explanation: certification is possible. Audits are carried out by independent bodies, which then issue the certificates. Buyers have noticed: "ISO 42001 certified" now turns up in vendor questionnaires the way ISO 27001 used to. On March 18, 2026, CEN and CENELEC adopted it as a European standard, EN ISO/IEC 42001:2026. A formal milestone — and where a widespread misunderstanding begins.
Certification to ISO/IEC 42001 does not give presumption of conformity with the AI Act. In this context, presumption of conformity is a point of law, not praise. Conform to a harmonised standard cited by the European Commission in the Official Journal under Article 40, and you are presumed to meet the matching legal requirements. As of September 2026, the Official Journal still carried no citation of a harmonised AI standard. The European standard actually being drafted for the Act's quality-management requirements is prEN 18286, targeted for late 2026; EN 18286 was made available in July 2026, and it is the document to watch. Until a citation happens, the certificate and the legal presumption live in different rooms.
None of that makes the certificate useless. The discipline it builds — documented processes, named accountability, internal audit, management review, corrective action — is what Article 17 will demand in some form. Article 9 aligns closely with its risk and impact assessment process. It also leaves a paper trail, the kind regulators tend to value. That evidence can certainly be reused when doing AI Act conformity work. Do not present the certificate as if it settled the legal question, though. It doesn't.
Washington chose a framework, then leaned on it
The United States went the other way: no single AI law, but a voluntary framework whose influence keeps growing. January 2023 saw NIST release the AI Risk Management Framework 1.0 (NIST.AI.100-1). It is organized around four functions — Govern, Map, Measure, Manage — with Govern running through the other three. NIST states plainly that the framework is not a checklist. It is a process, meant for building trustworthy AI across the whole lifecycle.
The framework has kept growing since. July 2024 brought a Generative AI Profile (NIST AI 600-1), which named risks left unspecified in the 2023 document: hallucination, data leakage, dangerous capabilities. A Cyber AI Profile arrived in February 2026 and tied cybersecurity governance directly to AI risk. Next came a draft profile covering trustworthy AI in critical infrastructure, plus an April 2026 concept note spanning energy, water, healthcare, finance, transportation and other sectors.
"Voluntary," though, is doing heavy work in that sentence. The FTC, the Consumer Financial Protection Bureau, the FDA, the SEC and the Equal Employment Opportunity Commission all cite its principles in enforcement guidance — the same FTC that has moved from guidance to action with its first probe into rogue AI agents. Showing governance aligned with NIST is increasingly expected of federal contractors. Then, on February 19, 2026, the U.S. Treasury Department published a Financial Services AI Risk Management Framework built directly on NIST's structure, developed with the Cyber Risk Institute, mapping 230 control objectives across the AI lifecycle. For multinationals, the AI RMF has quietly become the working companion to EU AI Act compliance.
Three instruments, one stack
So why is this biting now? Because deployment stopped being experimental. Multi-agent setups are leaving pilots behind for production, and agents are reaching into finance, where autonomous systems already trade at high frequency. Inference itself is spreading onto open, decentralized networks, run by people the big cloud providers don't employ. With each step, the area governance must cover gets wider.
A global company in 2026 faces a stack with three parts. The management system in the stack is ISO/IEC 42001: the machinery covering policies, roles, audits and reviews. Risk methodology comes from the NIST framework: how risks are governed, mapped, measured and managed within that machinery. Then there is the law itself, the EU AI Act: what in-scope systems must demonstrably achieve, under threat of fines. The pieces were built to interlock: a Microsoft-authored crosswalk between the NIST framework and ISO 42001 already sits on the NIST AI Resource Center.
Reading all this, the temptation is to choose a single instrument and declare the job done. Rely on the law alone and you are left with duties but no method. Rely on the certificate and you get process, but no legal cover. Rely on the framework and you get sound risk practice, yet no European regulator has to accept it.
The gaps are real, though. Neither a certificate nor a framework self-assessment can discharge the conformity duties set by the Act. Work is still under way on the harmonised standards that would create those legal presumptions. Because high-risk duties phase in across 2027 and 2028, programs built today have to aim at requirements whose final technical form remains unsettled. And the big three aren't the whole map: the Bletchley Park AI Safety Summit of November 2023 and its successors opened an international conversation on frontier-model safety.
So where does a builder start? With an inventory: list every AI system in use and classify each one against the Act's risk tiers. You can't govern what you haven't mapped, and the 2027–2028 deadlines reward whoever starts early. An inventory sounds dull. It is also where programs quietly fail, because unlisted systems surface later — in an audit, or in an incident. Whether or not a certificate follows, build those management habits now: the documentation discipline is what auditors and market-surveillance authorities will want to see. Run the NIST cycle — Govern, Map, Measure, Manage — as the working method, and add the Generative AI Profile if large language models are in play. Procurement belongs inside the perimeter as well: ask vendors for governance evidence, since deployer duties under the Act extend to systems you did not build. And budget honestly: a 42001 certificate is an asset, not a shield.
The quiet shift is over. Governance is infrastructure now — load-bearing, audited, and as consequential as any other part of the stack.
Standards on paper are now meeting enforcement in practice. In Washington, the FTC has opened its first probe into rogue AI agents, an early test of how far existing consumer-protection law stretches over autonomous software. And when governance fails at the state level, the cost is concrete: Denmark's national population register lost 8.8 million records, a reminder that data-protection rules matter most before the breach, not after it.
Sources
- European Commission — Regulatory framework on AI (AI Act overview): digital-strategy.ec.europa.eu
- Medium / Luis Ospina — "ISO/IEC 42001: What That AI Governance Certificate Won't Tell You" (EN ISO/IEC 42001:2026 adoption, March 18, 2026): medium.com
- Certify Consulting — "ISO 42001 vs NIST AI RMF vs EU AI Act: Which Do You Need?" (high-risk obligations timeline, harmonised standards status): certify.consulting
- CertCrush — "ISO 42001 Explained: What Certification Actually Requires" (presumption of conformity, EN 18286): certcrush.app
- Mitratech — "NIST AI Risk Management Framework (RMF)" (AI RMF 1.0, Generative AI Profile, Treasury Financial Services framework Feb 2026): mitratech.com
- SC Media — "NIST releases draft cybersecurity framework for AI adoption" (Cyber AI Profile, December 2025 draft): scworld.com
