Fines, Penalties and Enforcement Under the EU AI Act: A Guide for US Businesses
The EU Artificial Intelligence Act introduces one of the most far-reaching penalty regimes in the history of technology regulation. US businesses that develop, deploy or import AI systems used in the European Union face direct exposure — regardless of where they are incorporated.
REGULATION (EU) 2024/1689 · IN FORCE FROM 1 AUGUST 2024 · LAST REVIEWED APRIL 2026
€35M
Maximum Fine / Prohibited AI
7%
Global Turnover Cap (Tier 1)
27
EU Member States, Each With Enforcement Powers
Aug 2026
Full Enforcement Date for Most Provisions
Contents
- Overview and Significance for US Businesses
- Extraterritorial Scope: Does the EU AI Act Apply to You?
- Implementation Timeline and Enforcement Milestones
- The Three-Tier Penalty Structure (Article 99)
- Fines for General-Purpose AI Model Providers (Article 101)
- The Enforcement Architecture
- Market Surveillance, Complaints and Investigations
- Factors That Affect the Level of a Fine
- SME and Start-Up Provisions
- Practical Implications and Steps for US Businesses
1. Overview and Significance for US Businesses
The EU Artificial Intelligence Act — formally Regulation (EU) 2024/1689 of the European Parliament and of the Council — was published in the Official Journal of the European Union on 12 July 2024 and entered into force on 1 August 2024. It is the world’s first comprehensive, legally binding horizontal framework for the regulation of artificial intelligence, establishing requirements that span transparency, documentation, human oversight, data governance, accuracy, robustness, and cybersecurity, calibrated to the level of risk that an AI system presents. For US businesses, the Act represents the same kind of compliance inflection point that the GDPR did in 2018: a European regulation with deliberate extraterritorial reach, backed by financial penalties that dwarf those seen in most US state-level technology regulation.
The penalty regime is deliberately punitive in design. EU policymakers drew explicit lessons from the early years of GDPR enforcement, during which some supervisory authorities were slow to issue significant fines and companies engaged in prolonged litigation to delay action. The AI Act attempts to concentrate enforcement resources, create clearer investigative pathways, and set maximum fine ceilings well above those that regulators have consistently applied under data protection law. For companies that build and deploy AI at scale — a description that fits a large share of US technology businesses operating internationally — the exposure under the AI Act is not a compliance footnote. It is a material financial and reputational risk.
This article is written for US businesses seeking to understand not simply what the fines are, but how they are structured, which authority can impose them, how investigations will likely proceed, what factors will influence the size of any penalty actually imposed, and what practical steps should be taken now. It covers the penalty tiers set out in Articles 99 and 101, the enforcement architecture established by the Act, the role of the European AI Office and national competent authorities, the factors relevant to fine-setting, and the specific provisions designed to moderate penalties for small and medium-sized enterprises.
Why This Matters Now
Prohibitions on the most dangerous AI practices have been enforceable since 2 February 2025. Rules for general-purpose AI model providers have applied since 2 August 2025. The core obligations for high-risk AI systems — including those most likely to affect US technology companies — become fully enforceable on 2 August 2026. Businesses that have not yet begun compliance planning are already behind.
2. Extraterritorial Scope: Does the EU AI Act Apply to You?
The geographic scope of the AI Act is set out in Article 2 and is deliberately broad. Unlike some product safety frameworks that focus solely on physical placement of goods in a market, the AI Act is written to capture a wide range of actors in the AI value chain, including those located entirely outside the EU. US businesses should not assume that the absence of a European corporate entity, an EU tax registration, or even EU-based employees provides any shelter from the Act’s obligations.
The Act applies to providers who place AI systems or general-purpose AI models on the EU market or put them into service in the EU, irrespective of whether those providers are established in the EU or in a third country. It applies to deployers of AI systems who are located within the EU. Critically, it also applies to providers and deployers located in third countries — including the United States — where the output produced by an AI system is used in the EU. This third category is the most expansive and the one most likely to catch US businesses that do not believe they have a meaningful European presence. If a US company’s AI system produces outputs — recommendations, decisions, generated content, risk assessments, predictions — and those outputs are consumed by individuals or organizations located in the EU, the company is within scope.
There are limited carve-outs. AI systems developed and deployed exclusively for military, national security, and defense purposes by or on behalf of a member state government fall outside the regulation’s scope, as do AI systems used solely for scientific research and development in certain circumstances. However, these exemptions are narrow and will not assist the vast majority of commercial AI deployments. AI systems released under open-source licenses receive some tailored treatment under the Act, but open-source release does not provide a blanket exemption from obligations, particularly where a provider profits from the deployment of such systems or where systems are classified as high-risk or fall within the prohibited categories.
Key Compliance Question for US Businesses
The central question is not “do we have a European subsidiary?” but rather “are our AI systems’ outputs being used in the EU?” For most US technology companies offering AI-powered products and services globally, the answer will be yes — triggering obligations under the Act and, upon non-compliance, exposure to its penalty regime.
Where a provider established outside the EU places an AI system on the EU market, the Act requires that provider to designate an authorized representative established in the EU before placing the system on the EU market. This representative has a defined legal role — they can be held responsible by national authorities for compliance failures — but the provider itself remains primarily liable. The obligation to appoint an EU representative is itself enforceable, and failure to do so can independently constitute a violation subject to sanction.
3. Implementation Timeline and Enforcement Milestones
The EU AI Act does not apply all at once. It is structured around a phased implementation schedule that began running from the date of entry into force on 1 August 2024. Understanding this timeline is essential for US businesses attempting to prioritize compliance activity and to understand when enforcement exposure actually arises for each category of obligation.
1 August 2024
Regulation entered into force. The six-month and twelve-month clocks began running. Member States began the process of designating national competent authorities.
2 February 2025
Prohibited AI practices (Article 5) became enforceable. AI systems falling within the absolute prohibitions — including certain biometric categorization systems, social scoring, and AI that exploits psychological vulnerabilities — became unlawful to place on the market, put into service, or use. Penalties for violations of Article 5 are at the highest tier.
2 August 2025
General-purpose AI (GPAI) model obligations became applicable. Providers of GPAI models — including foundation models and large language models — must comply with transparency, documentation, copyright policy, and (for models presenting systemic risk) adversarial testing and incident reporting obligations. The European AI Office assumed primary enforcement responsibility for GPAI models from this date.
2 August 2026
Full application of the Act for most provisions. Obligations for high-risk AI systems listed in Annex III — covering areas such as employment and HR, access to education, credit scoring, law enforcement, and critical infrastructure — become fully enforceable. This is the date by which most US businesses deploying commercially significant AI systems must have their compliance programs in place.
2 August 2027
Obligations for high-risk AI systems embedded in regulated products covered by Annex I (such as medical devices, machinery, and vehicles) become applicable. These systems are subject to existing product safety conformity assessment processes and the AI Act’s additional requirements are layered on top.
4. The Three-Tier Penalty Structure (Article 99)
Article 99 of the EU AI Act establishes a three-tier penalty structure, calibrated to the seriousness of the non-compliance. The structure closely mirrors the two-tier approach familiar from Article 83 of the GDPR, but adds a third tier and reaches maximum figures significantly higher than those under data protection law. In each case, the applicable maximum is expressed as the higher of an absolute euro amount or a percentage of the offending undertaking’s total worldwide annual turnover — a mechanism that ensures the penalty scales to the size of the business and cannot be trivially absorbed as a cost of doing business by large multinational corporations.
Tier 1 — Highest
€35,000,000 or 7% of global annual turnover
Applies to violations of Article 5 — the absolute prohibitions on certain AI practices. Also applies to violations of the obligations on providers of GPAI models with systemic risk where the violation is of the most serious kind.
Tier 2 — Mid-Range
€15,000,000 or 3% of global annual turnover
Applies to violations of most other obligations under the Act, including obligations on providers of high-risk AI systems (conformity assessments, technical documentation, human oversight measures, transparency requirements) and obligations applicable to deployers, importers, distributors, and authorized representatives.
Tier 3 — Base Level
€7,500,000 or 1.5% of global annual turnover
Applies to the supply of incorrect, incomplete, or misleading information to notified bodies and national competent authorities in the context of a notification, certification, or investigation. This tier targets procedural non-compliance and obstructive behavior during enforcement processes.
Understanding the “Higher Of” Mechanism
The structure of each tier requires the imposing authority to calculate both the absolute euro ceiling and the turnover-based ceiling, and then apply whichever is higher. For large US technology companies — many of which have global revenues measured in the tens or hundreds of billions of dollars — the turnover-based figure will in virtually every case produce a substantially higher ceiling than the nominal euro cap. A business with $50 billion in global annual revenue facing a Tier 1 violation has a theoretical maximum exposure of approximately $3.5 billion. Even accepting that actual fines will typically be imposed well below their statutory ceilings, these figures place EU AI Act enforcement firmly in the category of enterprise-level financial risk.
For smaller businesses, the absolute euro figures cap the exposure, which provides some protection for start-ups and scale-ups. However, even €7.5 million at the base tier represents a material penalty for a company of modest size, particularly when reputational damage, litigation costs, and remediation expenses are factored in alongside the fine itself.
Prohibited AI Practices — Tier 1 Exposure Since February 2025
The most severe penalties attach to the Article 5 prohibitions, which have been enforceable since 2 February 2025. These include the deployment of AI systems that use subliminal techniques to manipulate behavior in ways that cause harm; systems that exploit vulnerabilities arising from age, disability, or social or economic situation; AI-based social scoring by public authorities; real-time remote biometric identification in public spaces for law enforcement (subject to narrow exceptions); retrospective remote biometric identification systems; emotion recognition in workplaces and educational institutions; and biometric categorization systems that infer sensitive attributes such as race, political opinion, trade union membership, or sexual orientation. US businesses offering or planning to offer any functionality that might fall within these categories must treat compliance as an immediate and urgent matter, not a future obligation.
Relationship to Other Regulatory Frameworks
US businesses already subject to the GDPR should note that AI Act violations and data protection violations can co-exist and be independently sanctioned. An AI system that processes biometric data unlawfully may generate simultaneous liability under both the AI Act and the GDPR. The Act expressly contemplates coordination between national competent authorities under the AI Act and data protection supervisory authorities. Where a single set of acts constitutes a violation of both frameworks, the principle of coordination will need to be worked through in practice, but businesses should not assume that liability under one framework reduces or eliminates liability under the other. The cumulative penalty exposure can therefore be very significant indeed.
5. Fines for General-Purpose AI Model Providers (Article 101)
Article 101 of the AI Act creates a separate penalty regime specifically for providers of general-purpose AI (GPAI) models. This category includes the large foundation models and large language models that underpin much of the current wave of AI development — models trained on vast datasets with the capability to perform a wide range of tasks. US companies that develop and make available such models within the EU are squarely within the scope of Article 101, and given the concentration of GPAI model development in the United States, this provision is of particular relevance to US technology businesses.
Under Article 101, the European Commission — acting through the AI Office, which sits within the Commission — has the power to impose fines on GPAI model providers for non-compliance with their obligations under Chapter V of the Act. The applicable ceiling for most GPAI model violations is €15,000,000 or 3% of total worldwide annual turnover in the preceding financial year, whichever is the higher. Where a GPAI model with systemic risk presents the most serious violations — for example, failures to conduct adequate adversarial testing, failures to implement serious incident reporting, or failures to take meaningful corrective action following a systemic risk assessment — the Commission may apply the higher Tier 1 ceiling of €35,000,000 or 7% of global turnover.
GPAI model providers are also subject to a specific regime under Article 101(3) allowing the Commission to impose periodic penalty payments where a GPAI model provider fails to comply with a request for information or a decision requiring access to a model. These periodic penalties can be imposed at up to 3% of the average daily worldwide turnover for each day of non-compliance, providing the Commission with a powerful tool to compel cooperation during investigations and ensure that enforcement proceedings are not frustrated by non-responsive subjects.
GPAI Model Obligations Now in Effect
From 2 August 2025, GPAI model providers must maintain and keep up to date technical documentation; make available a summary of training data used; have a copyright policy that complies with EU law; and — for models presenting systemic risk — conduct model evaluations including adversarial testing, report serious incidents to the AI Office, implement cybersecurity protections, and report their energy consumption. Failure to comply with any of these obligations is now enforceable. US companies releasing frontier AI models in the EU market should already be in compliance with these requirements or have a documented roadmap to achieve it.
The systemic risk designation under the AI Act is triggered by reference to the computational power used in training, expressed as floating-point operations (FLOP). Models trained using more than 10^25 FLOP are presumed to present systemic risk and are therefore subject to the enhanced tier of obligations — and the enhanced fine ceiling. As training compute has grown rapidly in recent years, and given that the threshold was set at 10^25 FLOP (meaning many leading large language models already meet or approach it), this is not a theoretical concern. US businesses developing frontier AI models should assume that their most capable models are or will shortly be within the systemic risk category, and plan their compliance and enforcement exposure accordingly.
6. The Enforcement Architecture
One of the AI Act’s most consequential design choices for US businesses is its bifurcated enforcement architecture. Enforcement is divided between the European AI Office at the Commission level — with jurisdiction over GPAI models — and national competent authorities in each of the 27 EU Member States, responsible for enforcement in relation to AI systems. Understanding which body has authority over which obligations, and how they interact, is essential for predicting how enforcement will develop and how US businesses should structure their engagement with regulators.
European AI Office (European Commission)
Established within the European Commission, the AI Office is the central EU-level body responsible for overseeing the implementation and enforcement of the AI Act with respect to GPAI models. It holds exclusive competence to investigate and fine providers of GPAI models for violations of Chapter V obligations. It also plays a coordinating and guidance role for national authorities, maintains the register of AI systems, and oversees the development of codes of practice for AI developers. For US businesses developing frontier AI, the AI Office is the primary enforcement counterpart.
National Competent Authorities (27 Member States)
Each EU Member State is required to designate at least one national competent authority, which encompasses both a notifying authority (responsible for third-party conformity assessment bodies) and a market surveillance authority. Market surveillance authorities are the primary enforcement bodies for AI systems other than GPAI models. They have powers to request information, conduct audits, access source code, order suspension of AI systems, and recommend fines. Each member state retains discretion in how it structures these authorities within its existing regulatory infrastructure.
European Artificial Intelligence Board (AI Board)
The AI Board is composed of representatives of national competent authorities from all 27 member states, with the European Data Protection Supervisor as an observer. It is an advisory and coordinating body, not an enforcement body in its own right. It issues opinions, recommendations, and guidance; promotes consistent application of the Act across member states; and acts as a forum for resolving jurisdictional questions between national authorities. Its role is analogous to that of the European Data Protection Board under the GDPR.
Scientific Panel of Independent Experts
The Act establishes a Scientific Panel of independent experts who support the AI Office and the AI Board with technical and scientific advice. The Panel has a specific role in flagging GPAI models that may present systemic risk and in supporting the evaluation of model capabilities during investigations. Its assessments are non-binding but are likely to carry significant weight in enforcement proceedings, particularly where highly technical questions of model capability and risk are at issue.
The One-Stop-Shop Question
US businesses familiar with the GDPR’s lead supervisory authority mechanism — under which a company typically has a primary data protection regulator in the member state where it has its EU establishment — will naturally ask whether an equivalent principle applies under the AI Act. The answer is nuanced. For AI systems (as opposed to GPAI models), jurisdiction is allocated to national competent authorities in each member state where the system is placed on the market or put into service. There is no single lead supervisory authority mechanism for AI systems of the kind that exists under the GDPR. In principle, therefore, a US business could face parallel investigations by multiple national authorities for the same AI system.
However, the Act does include coordination mechanisms. Where a market surveillance authority in one member state takes action or initiates proceedings, it must notify other relevant authorities through the EU’s market surveillance information and communication system. National authorities are required to cooperate and share information. The AI Board can provide opinions to resolve inconsistencies and facilitate coordinated enforcement. In practice, the most significant AI Act cases against major US providers are likely to attract coordinated involvement from multiple national authorities — creating a dynamic more like a joint investigation than a series of parallel independent proceedings. Finland was among the first member states to formally designate its market surveillance authority and confirm enforcement powers under the Act, and other member states have been following suit through 2025.
Coordination with Data Protection Authorities
National AI market surveillance authorities are required to cooperate with national data protection supervisory authorities where an AI system processes personal data, given the obvious overlap between the AI Act’s obligations and those of the GDPR. Data protection authorities may themselves investigate aspects of AI deployments under the GDPR, and referrals between authorities are contemplated. US businesses that have established relationships with EU data protection authorities should not assume that their engagement with those authorities will replace any need to engage with AI market surveillance authorities. The two frameworks have distinct legal bases, distinct supervisory structures, and distinct enforcement tools.
7. Market Surveillance, Complaints and Investigations
The AI Act grants national market surveillance authorities extensive investigative powers. These are drawn largely from the EU’s broader market surveillance framework under Regulation (EU) 2019/1020 but are adapted to the specific characteristics of AI systems, including their intangibility, the complexity of their technical operation, and the difficulty of tracing harms to a specific AI system’s behavior. For US businesses, understanding these powers is important both for assessing the likelihood and pace of enforcement and for planning how to manage any regulatory engagement that arises.
Market surveillance authorities may request access to the technical documentation that providers of high-risk AI systems are required to maintain. This documentation must include a general description of the AI system, its intended purpose, the data it has been trained on, the components it uses, its performance metrics, and the risk management process the provider has applied. Authorities may request that providers make this documentation available at short notice, and refusal or delay is itself a sanctionable failure. Beyond documentation, authorities have the power to conduct on-site inspections, require access to training and testing datasets, and — in specific circumstances — require access to source code. The power to compel access to source code is significant for US businesses that regard their model architecture and training methodologies as core intellectual property: the Act does not exempt proprietary materials from disclosure where necessary for a legitimate enforcement inquiry.
Article 85 of the Act provides individuals and organizations — including civil society organizations — with the right to submit complaints to national competent authorities about AI systems they believe to be non-compliant. This complaint mechanism, combined with the mandatory reporting obligations imposed on deployers who use high-risk AI systems and have reason to believe those systems present a risk, creates multiple pathways through which enforcement investigations may be initiated. US businesses should anticipate that, as the Act beds in and regulators and the public become more familiar with its requirements, complaints will become an increasingly important trigger for enforcement activity — as has been the case under the GDPR.
Interim Measures and Suspension Powers
Where a national competent authority has sufficient grounds to believe that an AI system presents a serious risk and that urgent action is necessary to prevent harm, it may order interim measures — including the temporary suspension of an AI system from the EU market — pending the conclusion of a full investigation. This power is closely analogous to the urgent binding decision powers available to EU data protection authorities under Article 66 of the GDPR. The ability to obtain rapid suspension of an AI system’s operation is a powerful enforcement tool, and US businesses should be aware that regulatory intervention can effectively take a product or service offline in EU markets while proceedings are ongoing.
Where a market surveillance authority initiates a formal investigation, it will typically begin by requesting information and documentation from the provider or its EU authorized representative. The authority will set deadlines for responses, and the Act is clear that providers and their representatives must cooperate with authorities in good faith. The manner in which a company engages during an investigation — its transparency, the quality of its documentation, the promptness of its responses, and whether it proactively identifies and remedies issues — will all be factors in any subsequent enforcement decision. Companies that are experienced in regulatory engagement from their GDPR compliance work will be at an advantage in navigating AI Act investigations, but should note that the technical complexity of AI systems typically requires more intensive preparation and specialist legal support than a data protection investigation.
8. Factors That Affect the Level of a Fine
The penalty ceilings set out in Article 99 are maxima, not prescriptions. In practice, the fine actually imposed in any given case will be determined by applying a range of factors specific to the circumstances of the violation, the company, and the harm or risk of harm caused. The Act sets out a non-exhaustive list of factors to be considered, closely modeled on the factors applied by data protection authorities under Article 83(2) of the GDPR and informed by established practice in EU competition law enforcement. For US businesses, understanding these factors is important both for assessing enforcement risk and for building the internal compliance frameworks that will be their best defense if a regulatory investigation occurs.
- Nature, gravity, and duration of the infringement — Violations of the absolute prohibitions in Article 5 will attract the most serious treatment. Violations that persist over a long period after a company had reason to be aware of a compliance issue, or that are repeated, will be treated more severely than isolated, short-lived failures.
- The degree of responsibility of the operator — Whether the violation results from deliberate decisions by senior management, or whether it reflects systemic negligence throughout an organization, or whether it was an isolated failure by a junior employee acting outside authorization will all be relevant. For large US enterprises, the expectation will be that the company had sufficient resources to have established proper compliance processes.
- The nature and number of persons affected and the harm or potential harm caused — Violations that affect large numbers of vulnerable individuals — for example, AI systems used in employment decisions, credit scoring, or law enforcement affecting populations across multiple member states — will be treated as significantly more serious than violations confined to a small and limited context.
- Whether the provider took measures to mitigate harm — Companies that, upon becoming aware of a compliance failure, promptly take remedial action, notify the relevant authority, withdraw or suspend the non-compliant system, and cooperate openly with the investigation will be treated substantially more favorably than those that delay, minimize, or seek to conceal the issue.
- Degree of cooperation with the supervisory authority — Active and genuine cooperation — providing documentation promptly, making relevant personnel available, proactively sharing information relevant to the investigation — will be a mitigating factor. Obstruction or significant delay will be treated as an aggravating factor and may independently trigger Tier 3 penalties.
- Financial benefit gained or loss avoided — Where a company has gained a competitive advantage by avoiding the costs of compliance, or has profited from a non-compliant AI system’s deployment, the fine will be calibrated to ensure that it exceeds the financial benefit of the violation. This prevents companies from treating fines as a predictable cost of doing business.
- Prior violations — Where a company has previously been found in violation of the AI Act or of related EU regulatory requirements — including the GDPR — repeat violations will attract enhanced penalties.
- Any relevant technical and organizational measures implemented — The presence of a serious, well-resourced, and genuinely operational AI compliance program — including risk assessments, technical documentation, conformity assessments, internal auditing, and staff training — will be a significant mitigating factor, even where it did not ultimately prevent the violation in question.
Member states have the authority to set their own enforcement guidelines and to determine the specific process for fine calculation within the ceilings set by Article 99. As with the GDPR, it is likely that enforcement practice will vary somewhat between member states in the early years, with some authorities adopting more structured and transparent methodologies and others proceeding on a more case-by-case basis. US businesses with the most significant EU exposure should monitor enforcement decisions from the national authorities in member states where they have the greatest commercial presence or user base.
9. SME and Start-Up Provisions
The EU AI Act contains specific provisions designed to reduce the compliance burden and moderate the potential penalty exposure for small and medium-sized enterprises (SMEs) and start-ups. These provisions reflect a conscious policy decision to avoid imposing requirements that, while appropriate for large and well-resourced technology businesses, could have a disproportionate chilling effect on innovation by smaller companies.
For SMEs — defined by reference to the EU’s standard definition, which encompasses companies with fewer than 250 employees and either annual turnover not exceeding €50 million or a balance sheet total not exceeding €43 million — and for natural persons who are providers of AI systems, Article 99 specifies that the applicable fine ceiling is the lower of the percentages and the absolute euro amounts set out for each tier. This is the reverse of the approach for large enterprises, where the higher amount applies. In practical terms, this means that SME providers face an absolute upper limit per tier rather than a percentage-based ceiling that could dramatically magnify the nominal figure.
Beyond the fine-setting provisions, the Act requires member states to take specific measures to support SMEs and start-ups in understanding and complying with their obligations. National competent authorities are expected to provide guidance, establish regulatory sandboxes — controlled environments in which new AI systems can be tested in real-world conditions with regulatory oversight — and make compliance documentation and templates available. AI regulatory sandboxes are mandatory under the Act, and member states are required to establish them as part of their national implementation. For US start-ups seeking to enter the EU market with innovative AI products, these sandboxes offer a structured pathway to develop compliant systems while engaging directly with regulators, which can significantly de-risk market entry.
A Note on US Corporate Structure
US businesses operating through large parent corporations or groups should note that for the purpose of calculating turnover-based fine ceilings, the relevant turnover is typically that of the “undertaking” — a concept drawn from EU competition law that captures the economic entity as a whole, including subsidiaries and related companies under common control. A US start-up that is a subsidiary of a large technology group is unlikely to benefit from the SME fine ceiling even if its own revenues are modest. Corporate structure will not allow a large business to reduce its fine exposure by routing EU AI activities through a smaller subsidiary.
10. Practical Implications and Steps for US Businesses
The EU AI Act creates a genuinely new category of regulatory risk for US businesses. It is not simply an extension of the GDPR, nor is it reducible to a product safety regulation of the familiar kind. It is a comprehensive horizontal framework that imposes obligations throughout the AI lifecycle — at the design, training, testing, deployment, monitoring, and decommissioning stages — and backs those obligations with a penalty regime calibrated to deter non-compliance by even the largest technology companies in the world. The question for US businesses is not whether to engage with the Act, but how to do so in a way that manages risk proportionately and effectively.
The first and most fundamental step is to conduct a thorough inventory of the AI systems and models that your business develops, deploys, distributes, or imports, and to assess which of those systems fall within the scope of the EU AI Act by reason of their use or intended use in the EU. This scoping exercise must be done with legal rigor. Many US businesses will find that the scope of the Act extends considerably further than their initial assumptions suggest, particularly given the breadth of the extraterritorial provisions and the width of the definition of “AI system” adopted in Article 3 of the Act. Legal counsel experienced in both EU AI regulation and the specific technical characteristics of AI systems is essential for this exercise.
Once the scope is established, businesses should classify each in-scope AI system by risk category. The distinction between prohibited systems, high-risk systems listed in Annex III, systems subject to transparency obligations only, and minimal risk systems is critical because the obligations — and the penalty exposure — vary dramatically between categories. For each high-risk AI system, the business must then ensure that it has in place the technical and organizational measures required by the Act: a risk management system; technical documentation; data governance and data management practices; record-keeping and logging capabilities; transparency and information provision to deployers and users; human oversight mechanisms; and where required, conformity assessment procedures.
For providers of GPAI models, the compliance agenda is somewhat different and in some respects more urgent, given that GPAI model obligations have been in effect since August 2025. The Act’s requirements for technical documentation of GPAI models, training data transparency, copyright compliance, and — for systemic risk models — adversarial testing and incident reporting are demanding, and US businesses that develop frontier AI should already have mapped their existing practices against these requirements and identified gaps. The European AI Office has published codes of practice for GPAI model providers, and engagement with these codes — even where compliance is achieved by alternative means — is likely to be viewed favorably by the AI Office in any subsequent enforcement context.
The designation of an EU authorized representative is not merely a compliance formality. It creates a defined legal presence in the EU for enforcement purposes, and the choice of representative — whether an EU subsidiary, an external consultant, or a purpose-built compliance entity — should be made with care. The representative must be genuinely empowered to represent the provider in dealings with national authorities, to provide access to required documentation, and to take instructions on behalf of the provider in the event of an investigation. Regulators will test whether the representative function is substantive or purely nominal.
Finally, US businesses should invest in cross-functional governance structures that bring together legal, technical, product, and compliance teams around AI Act compliance. The obligations of the Act cannot be discharged by a legal team alone. Technical documentation requires engineers; data governance requires data scientists; human oversight mechanisms require product designers; conformity assessments require both technical and legal expertise. Building the governance infrastructure to manage AI Act compliance on an ongoing basis — including processes for monitoring regulatory developments, updating documentation as systems change, reporting incidents, and responding to regulatory inquiries — is an investment that will pay dividends both in reduced penalty risk and in the increasingly important commercial context of demonstrating trustworthy AI practices to EU business partners and customers.
Do Not Wait for August 2026
While the full complement of high-risk AI system obligations does not apply until 2 August 2026, compliance programs cannot be built in days or weeks. Achieving the technical documentation, risk management processes, conformity assessments, and governance structures required by the Act typically takes twelve to eighteen months for organizations of any complexity. For US businesses that have not yet begun this work, the time to act is now — not as a matter of theoretical forward planning, but as an immediate operational priority.
Speak with Our EU AI Act Practice
Our data protection and technology lawyers advise US businesses on EU AI Act compliance, risk assessment, authorized representative obligations, and regulatory engagement. We help clients understand their exposure and build compliance programs that are proportionate, practical, and defensible.