Under the European AI Act, high-risk AI systems are subject to extensive obligations regarding technical documentation, risk management, transparency, human oversight, and conformity assessment. These requirements apply to both providers and organizations that deploy such systems. In “AI-Driven Intelligence,” you can read about how organizations can use AI responsibly. In this article, we answer the most frequently asked questions about what exactly high-risk AI entails and what obligations come with it.
Which AI systems exactly fall into the high-risk category?
High-risk AI systems are systems used in contexts where the outcomes can have significant consequences for people’s health, safety, or fundamental rights. The AI Act defines two main categories: systems used as safety components in products already covered by existing EU legislation, and systems deployed in specific application areas designated by law.
The designated high-risk application areas include, among others:
- Critical infrastructure such as energy, water, and transportation
- Education and vocational training, such as systems that determine who is admitted to a program
- Employment and Human Resources Management, such as automated application screening
- Access to essential services such as credit, social benefits, and health insurance
- Law Enforcement and Border Control
- The Administration of Justice and Democratic Processes
- Biometric Identification and Categorization of Individuals
It is important to note that not every AI system used in one of these sectors is automatically classified as high-risk. The key factor is whether the system plays a decisive role in processes that have direct consequences for people. A simple scheduling system in healthcare generally falls outside the high-risk category; a system that makes medical diagnoses or assesses patient admissions, however, does fall under it.
What technical requirements apply to high-risk AI systems?
High-risk AI systems must meet a set of technical requirements focused on reliability, safety, and verifiability. At the core of these requirements are robustness, accuracy, and the minimization of risks throughout the system’s entire lifecycle.
The main technical requirements are:
- Risk Management System: An ongoing process that identifies, evaluates, and manages risks throughout the system’s entire lifecycle.
- Data Quality: Training, validation, and test data must be relevant, representative, and free of errors. Bias in the data must be actively minimized.
- Robustness and accuracy: The system must perform consistently and reliably, even in the event of unexpected input or cyberattacks.
- Cybersecurity: Technical measures to protect the system against tampering and misuse.
- Logging and recording: Automatic recording of events throughout the lifecycle, so that the system’s performance can be reviewed afterward.
- Human oversight: The system must be designed so that people can effectively monitor and understand it, and halt or correct it if necessary.
These technical requirements do not apply only once, at the time of the system’s launch. They remain in effect throughout the system’s entire operational lifespan and require periodic evaluation and adjustment.
What should be included in the technical documentation for a high-risk AI system?
The technical documentation for a high-risk AI system must demonstrate that the system complies with all applicable requirements of the AI Act. This documentation is prepared by the provider before the system is placed on the market and must be retained for at least ten years.
The documentation must include at least the following:
- A general description of the system, including its intended purpose and the technology used
- A detailed description of the system’s architecture and logic
- Information about training dates, validation methods, and test results
- A description of the risk management system and the control measures implemented
- Performance metrics, including accuracy thresholds and known limitations
- Measures for human oversight and the ways in which users can monitor the system
- A description of cybersecurity measures
- The declaration of conformity and, where applicable, the CE marking
In addition to the technical documentation, suppliers must also prepare user instructions. These are intended for the organizations that use the system and contain practical information on proper commissioning, monitoring, and identifying risks.
How does the conformity assessment for high-risk AI systems work?
A conformity assessment is the process by which a provider demonstrates that a high-risk AI system complies with the requirements of the AI Act. For most high-risk systems, the provider can conduct this assessment itself through an internal procedure. For specific categories, such as biometric identification systems, the involvement of a notified body is mandatory.
Self-Assessment vs. External Assessment
In the self-assessment process, the provider follows a structured internal procedure based on the harmonized standards published in support of the AI Act. The provider prepares the technical documentation, implements the risk management system, and records the results. Upon successful completion, a declaration of conformity is issued and, where applicable, a CE marking is affixed.
Registration in the EU database
High-risk AI systems must be registered in a central EU database before being placed on the market. This database is publicly accessible and promotes transparency and oversight. The registration includes key information about the system, the provider, and the intended use. Providers outside the EU are required to designate an authorized representative within the EU who will fulfill these obligations on their behalf.
What are the obligations for organizations that use high-risk AI?
Organizations that deploy high-risk AI systems—known as “deployers”—have specific obligations under the AI Act. They are not responsible for the technical development of these systems, but they are responsible for ensuring their responsible use in practice.
The main obligations for organizations that use high-risk AI are:
- Follow the instructions for use: The system may only be used in the manner specified by the provider.
- Organizing human oversight: Concrete measures must be in place to ensure human oversight, including the designation of responsible staff members.
- Fundamental Rights Impact Assessment: Government agencies and certain other organizations must assess in advance the potential impact on the fundamental rights of those affected.
- Retention of Logs: Automatically generated log files must be retained for the period specified by the provider, and for at least six months.
- Reporting Incidents: Serious incidents or malfunctions with potentially dangerous consequences must be reported to the competent national authority.
- Transparency toward those affected: When the system is used to make decisions about individuals, those individuals must be informed of this.
Please note: An organization that modifies a high-risk AI system, brands it with its own name, or uses the system for a purpose for which it was not intended becomes, from a legal standpoint, the provider itself. In that case, all obligations that apply to the original provider also apply to the organization.
What are the consequences of failing to comply with the high-risk AI requirements?
Failure to comply with the requirements for high-risk AI systems can result in substantial fines. The AI Act establishes a tiered fine structure that took effect on August 2, 2025, with the amount of the fine depending on the nature of the violation.
The three stages of penance are:
- Prohibited AI Practices (Article 5): A maximum of 35 million euros or 7% of global annual revenue, whichever is higher.
- Other non-compliance, including high-risk obligations: Up to 15 million euros or 3% of global annual revenue.
- Inaccurate or misleading information provided to regulators: Up to 7.5 million euros or 1% of global annual revenue.
For small and medium-sized enterprises, the lower of the fixed amount or the percentage applies in each case, which offers some protection. Oversight of high-risk AI rests with national market surveillance authorities. In January 2026, Finland became the first Member State to grant enforcement powers under the AI Act to its national authority. Other Member States are expected to follow suit soon, leading to further tightening of enforcement throughout 2026.
In addition to financial penalties, authorities may also require that a system be withdrawn from the market or that its use be suspended until the noncompliance has been rectified. The reputational damage associated with public enforcement actions can be just as severe for many organizations as the fines themselves.
How Pegamento Helps with High-Risk AI Compliance
Navigating the requirements surrounding high-risk AI requires technical knowledge, legal insight, and practical implementation experience. We help organizations deploy AI responsibly and in compliance with regulations, without compromising speed or innovation.
What we can do for you:
- Assessment and Risk Classification: We identify which AI applications within your organization fall into the high-risk category and determine the resulting obligations.
- Agentic AI within compliance frameworks: Our Agentic AI for customer service is designed with human oversight as a core principle. Whereas traditional RPA bots operate based on fixed instructions, our self-thinking Agentic AI assistants take the initiative independently and act proactively—but always within the parameters you set.
- Technical Documentation and Risk Management: We provide support in preparing the required documentation and establishing an ongoing risk management system.
- Everything under one roof: From consulting and implementation to management and support, you have a single point of contact for the entire process, without the need for complex supplier coordination.
- Certified Process: Our approach is based on ISO 27001 (information security), ISO 9001 (quality management), and ISO 26000 (corporate social responsibility), which aligns with the compliance requirements of the AI Act.
Would you like to know where your organization stands right now and what steps are needed to become compliant? Contact us, and together we’ll determine the best approach for your situation.
Frequently Asked Questions
How do I know if my existing AI system is classified as high-risk under the AI Act?
Start by conducting a thorough inventory of all AI applications within your organization and assess them against the use cases that the AI Act designates as high-risk. When doing so, don’t just look at the sector, but specifically at the role the system plays: does it have a direct impact on decisions that affect people in the areas of employment, healthcare, credit, or safety? If you’re unsure, it’s wise to seek legal or technical advice, as an incorrect classification can lead to non-compliance and associated fines.
What is the difference between a provider and a deployer, and why is that distinction so important?
A provider is the party that develops and markets a high-risk AI system; a deployer is the organization that puts the system into practice. The distinction is crucial because both parties have their own, non-transferable obligations under the AI Act. Important note: as soon as a deployer modifies the system, puts its own name on it, or uses it for a purpose other than its intended one, that party legally becomes a provider itself, and all associated obligations apply in full.
How do I start setting up a risk management system for a high-risk AI application?
A good risk management system starts with systematically identifying all possible risks throughout the system’s entire lifecycle, from development to decommissioning. Next, you assess the likelihood and impact of each risk and establish control measures, such as technical restrictions, user instructions, and monitoring procedures. It’s important to note that this isn’t a one-time exercise: the system requires periodic review, especially when the model is updated, the data changes, or its use is expanded.
As a deployer, do I also need to take any action if my employees use the AI system on a daily basis?
Yes, human oversight is one of the core obligations for deployers and goes beyond just a technical setting. You must specifically define who within the organization is responsible for monitoring the system, how employees are trained to critically evaluate the output, and how they can stop or correct the system if necessary. Also, ensure that employees understand what the system can and cannot do, so they do not blindly rely on the results.
What is a Fundamental Rights Impact Assessment, and when is it required?
A fundamental rights impact assessment (FRIA) is a preliminary analysis of the potential effects of a high-risk AI system on the fundamental rights of data subjects, such as the right to privacy, non-discrimination, and due process. This assessment is mandatory for government agencies and private organizations that use high-risk AI in publicly accessible services, such as lending or insurance. The results must be documented and may lead to additional control measures before the system is put into use.
What happens if my high-risk AI system causes a serious incident?
In the event of a serious incident—such as a system failure with dangerous consequences or a significant violation of fundamental rights—you, as the deployer, are required to report this to the competent national market surveillance authority. Do this as soon as possible and document the incident in detail, including the cause, the consequences, and the corrective measures taken. Also, retain the automatically generated log files, as they are essential for the investigation and can demonstrate that your system and processes were properly configured.
On average, how long does it take to make a high-risk AI system fully compliant?
The turnaround time depends heavily on the complexity of the system, the quality of the existing documentation, and the maturity of the internal processes. For organizations already working with frameworks such as ISO 27001 or ISO 9001, the step toward AI Act compliance is considerably smaller, as many building blocks are already in place. In practice, a full compliance process—including risk assessment, documentation, and the implementation of oversight procedures—ranges from a few months for simpler systems to more than a year for complex, large-scale applications.
