The EU AI Act establishes a risk-based framework that places the heaviest compliance obligations on AI systems most likely to affect people’s health, safety, or fundamental rights. For businesses building or deploying AI in Europe, understanding where high-risk AI systems sit within that framework is not optional. The EU AI regulation has real teeth: enforcement powers are now active, fines are substantial, and the rules apply to companies regardless of where they are headquartered.
This breakdown covers everything you need to know about high-risk AI under the EU AI Act: how systems get classified, which use cases are covered, what obligations apply, what penalties look like, and where the compliance timeline currently stands after the Digital Omnibus reforms of 2026.
How the EU AI Act classifies high-risk systems
The EU AI Act uses a four-tier risk hierarchy, and classification determines every downstream compliance obligation. At the top sit prohibited AI practices. Below that are high-risk AI systems, which face the most extensive requirements. Limited-risk and minimal-risk systems face lighter or no mandatory obligations.
Article 6 of the AI Act defines two independent routes into the high-risk category. The first route applies when an AI system functions as a safety component of a product covered by EU harmonisation legislation listed in Annex I, such as medical devices, machinery, or toys, and that product requires third-party conformity assessment. The second route applies when a system falls into one of the eight use-case domains listed in Annex III, regardless of whether it is embedded in a physical product.
The Article 6(3) exception and its limits
An Annex III system can escape the high-risk label if it only performs a narrow procedural task, improves a previously completed human activity, or carries out preparatory work without influencing the actual decision. However, this exception disappears entirely if the system profiles natural persons. As the Article 6 classification analysis from Close IT makes clear, profiling is the single most common place teams get the classification wrong.
The intended purpose of a system, as defined by the provider in its instructions for use, promotional materials, and technical documentation, plays a central role in determining classification. Providers who believe their Annex III system is not high-risk must document that self-assessment before placing it on the market and register it in the EU database. The burden of proof sits with the provider, not the regulator.
The European Commission published draft guidelines on high-risk classification in May 2026. Those guidelines adopt an expansive interpretation of the conformity assessment test, which means more AI systems than a straightforward reading of the Act suggests may fall within the high-risk regime. A public consultation closed in July 2026, and final guidelines are expected to follow.
Complete list of high-risk AI use cases
Annex III of the EU AI Act lists eight domains where AI systems are presumed high-risk. Each domain covers specific applications, and the boundaries matter because misclassification has direct compliance consequences.
- Biometrics: AI that identifies, categorises, or tracks individuals. Standard biometric authentication, such as a phone’s face unlock, is not covered because it verifies a known identity rather than identifying someone within a broader population.
- Critical infrastructure: AI that serves as a safety component in energy, water, transport, or digital infrastructure. Systems that only optimise energy usage without safety implications fall outside this category.
- Education and vocational training: AI that determines access or admission to educational institutions, or evaluates learner trajectories. The Commission’s draft guidelines indicate that learner-trajectory systems typically involve profiling and cannot use the Article 6(3) exception.
- Employment and worker management: AI used in recruitment, candidate selection, performance monitoring, and access to self-employment. These systems are high-risk across the board.
- Access to essential services and benefits: AI used for credit scoring, creditworthiness assessment, insurance risk pricing, and access to public services. Draft guidelines expanded the insurance category to include long-term care, personal pensions, and credit life insurance.
- Law enforcement: AI used by police or judicial authorities to assess risk, detect offences, or analyse evidence. AI used by lawyers for case research is generally not high-risk unless it directly influences judicial decisions.
- Migration, asylum, and border control: AI that assesses risks, verifies documents, or assists in asylum decisions.
- Administration of justice and democratic processes: AI that assists judges in sentencing or case research, or that influences electoral processes.
Two new prohibited practices were added to Article 5 by the Digital Omnibus, adopted by the European Parliament in June 2026: AI systems that generate non-consensual intimate imagery and AI systems that generate child sexual abuse material. These prohibitions take effect on December 2, 2026.
Mandatory obligations for high-risk AI providers
Providers of high-risk AI systems carry the heaviest obligations under the EU AI Act. Article 16 lists the full set, and the requirements span the entire lifecycle of the system, not just the moment it goes to market.
Technical and data requirements
Article 10 requires training, validation, and testing datasets to meet specific quality standards. Providers must assess datasets for potential biases that could affect health, safety, or fundamental rights, and put measures in place to detect, prevent, and mitigate those biases. Article 11 requires complete technical documentation demonstrating compliance before the system is placed on the market. Article 12 mandates automatic event logging so that risks can be identified and post-market monitoring can function effectively.
Risk management and human oversight
Article 9 requires providers to establish a risk management system as a continuous, iterative process throughout the system’s entire lifecycle, updated as post-market monitoring data comes in. Article 14 requires systems to be designed so that human operators can understand, monitor, and, where necessary, override the system’s outputs. These are not one-time setup tasks. They are ongoing obligations.
Conformity assessment and registration
Two conformity assessment paths exist: internal self-assessment by the provider, or third-party assessment through a notified body. The path depends on the system’s use case and risk profile. A quality management system aligned with Article 17 is required regardless of which path applies. Once conformity is confirmed, providers must draw up an EU declaration of conformity, affix the CE marking, and register the system in the EU database under Article 49 before placing it on the market.
Post-market monitoring and incident reporting
Article 72 requires providers to run a continuous post-market monitoring system that actively collects and analyses performance data throughout the system’s lifetime. Under Article 73, serious incidents must be reported to national market surveillance authorities within defined timeframes: 15 days as a standard, 2 days for widespread or severe incidents, and 10 days when a death is involved.
Deployer obligations
Deployers carry significant obligations too, and these are frequently overlooked. Under Article 26, deployers must use systems according to the provider’s instructions, assign human oversight personnel with appropriate training, ensure input data is relevant, and immediately report serious incidents. Article 27 requires certain deployers, including public bodies, private entities providing public services, and deployers of credit scoring or insurance pricing AI, to conduct a Fundamental Rights Impact Assessment (FRIA) before deployment. A GDPR Data Protection Impact Assessment does not substitute for a FRIA. The FRIA covers a broader range of fundamental rights beyond data protection. Under Article 25, a deployer who substantially modifies a high-risk AI system or places their own name or trademark on it becomes a provider and inherits the full set of provider obligations.
Penalties and enforcement for non-compliance
The EU AI Act uses a three-tier penalty structure, and the fines are larger than those under the GDPR for the most serious violations.
- Tier 1 (prohibited practices): Up to €35 million or 7% of total worldwide annual turnover, whichever is higher.
- Tier 2 (high-risk system non-compliance): Up to €15 million or 3% of global annual turnover, whichever is higher.
- Tier 3 (misleading information to authorities): Up to €7.5 million or 1% of global annual turnover, whichever is higher.
For SMEs, the calculation works differently. The upper bound is set by whichever amount is lower, not higher, between the fixed absolute amount and the turnover percentage. This is a meaningful protection for smaller businesses. Regulatory sandboxes are also available with priority access for SMEs and startups, providing a structured environment to test and develop AI systems with reduced regulatory exposure.
Enforcement is decentralised. National market surveillance authorities handle high-risk AI system enforcement at Member State level. The AI Office, operating within the European Commission, holds enforcement powers over general-purpose AI model providers, with those powers formally activating on August 2, 2026. As the Wilson Sonsini enforcement alert from August 2026 notes, the AI Office has described “technical compliance dialogues” as its preferred initial enforcement tool, though formal sanctions are available for flagrant violations.
The Act has extraterritorial reach under Article 2. It applies to any provider placing an AI system on the EU market, regardless of where the provider is based, and to any deployer located within the EU. Non-EU companies whose AI system outputs are used within the EU are also in scope.
Key compliance deadlines and transition timeline
The compliance timeline for high-risk AI systems has shifted significantly following the Digital Omnibus, adopted by the European Parliament in June 2026 and by the Council on June 29, 2026. The Omnibus extended two specific deadlines while leaving all others unchanged.
- February 2, 2025: Article 5 prohibited AI practices became enforceable. AI literacy obligations under Article 4 took effect. Both are already active and actionable.
- August 2, 2025: General-purpose AI model obligations took effect. The AI Office and AI Board became operational.
- August 2, 2026: Article 50 transparency obligations for AI-generated content apply. The AI Office’s formal enforcement powers over GPAI providers activated. This deadline was not deferred by the Digital Omnibus.
- December 2, 2026: Article 50(2) transparency requirements apply to legacy AI systems already on the market. Two new Article 5 prohibitions (non-consensual intimate imagery and CSAM) take effect.
- December 2, 2027 (revised): Full high-risk obligations under Articles 9 to 17 and Article 26 apply to standalone Annex III high-risk AI systems. This is a 16-month extension from the original August 2, 2026 deadline, as confirmed by the Gibson Dunn Digital Omnibus analysis.
- August 2, 2028 (revised): Full high-risk obligations apply to AI systems embedded in regulated products under Annex I, such as medical devices and machinery. This is a 12-month extension from the original August 2027 deadline.
The absence of harmonised technical standards was cited as a primary reason for the extensions. The Digital Omnibus added a new Article 40(2) paragraph requiring the Commission to ask European standards bodies to create unified technical standards covering both the AI Act and existing harmonisation laws simultaneously.
Common compliance pitfalls to avoid
Several patterns of error appear consistently across organisations preparing for AI Act compliance. Understanding them early saves significant remediation effort later.
Misclassification and the profiling trap
Misclassification is the most consequential compliance error. Many organisations treat recruitment tools, credit scoring systems, or diagnostic support as general-purpose software. Under the AI Act, these are high-risk use cases with mandatory obligations. The Article 6(3) exception is also frequently misapplied. Any system that profiles natural persons is always high-risk, regardless of how narrow or preparatory the task appears. Classification is also not a one-time exercise. When a system changes, gains new use cases, processes new data, or is deployed in a new context, the classification must be reassessed.
Confusing organisational roles
A single company can simultaneously be a provider, a deployer, a distributor, and an importer, each with different obligations. Organisations that develop parts of an AI system, integrate third-party components, deploy internally, and distribute externally need to map each activity to the correct role. Deployer obligations in particular are frequently underestimated. A business deploying a third-party AI credit scoring system without assigning human oversight personnel, retaining logs, or informing applicants faces Tier 2 penalties, even though it did not build the system.
Treating the deadline extension as a pause
The Digital Omnibus extended the enforcement deadline for standalone Annex III systems to December 2027. The obligations themselves have not changed. Conformity assessments, technical documentation, risk management systems, and human oversight design all still need to be completed. Using the extension as a reason to delay preparation is a strategic error, not a compliance strategy.
Incomplete AI system inventory
Compliance begins with knowing which AI systems are in use. Industry experience shows that a significant proportion of SMEs underestimate the number of AI systems operating within their organisation. Employee use of tools like ChatGPT or Microsoft Copilot, often without management awareness, is the most common blind spot. An accurate inventory is the foundation of every other compliance step.
Documentation built for inspections, not for compliance
Assembling compliance documentation quickly before an inspection, rather than building a continuous audit trail over time, is flagged by regulators as a red flag. A version-controlled audit trail that demonstrates consistent updates signals a functioning compliance process. Documentation assembled after the fact does not.
For businesses managing AI content workflows at scale, tools that generate and publish content automatically need to be assessed against the Act’s transparency and classification requirements. WP SEO AI’s content scaling service is built with human specialist oversight at every stage, which aligns directly with the human oversight obligations the Act imposes on deployers of AI systems that influence decisions or communications.
The EU AI Act is now an operational compliance framework, not a future regulation to monitor. For any organisation building, deploying, or distributing AI systems that interact with EU markets, the classification analysis, obligation mapping, and documentation work need to start now, well ahead of the December 2027 deadline for Annex III systems.
This content was generated with the help of AI and it may contain mistakes
