AI Law, Risk & Governance

High-Risk AI under the AI Act: What Companies Need to Know

Marin Niehues24 February 202511 min read

Two women filling out forms

Under the AI Act, high-risk AI comprises systems used in sensitive fields such as recruitment or credit scoring, as well as safety components of regulated products. For these systems, Regulation (EU) 2024/1689 prescribes a fixed set of obligations, ranging from risk management and documentation to human oversight. For companies, early classification is decisive, because it determines their own role and the scale of the delivery task. This article provides an overview and does not constitute legal advice.

The article AI management: what executive leadership must steer describes the steering of the AI portfolio as a whole. The article on the EU AI Act gives an overview of the regulation in its entirety, including prohibited practices.

Which AI systems are high-risk?

AI systems are high-risk when they serve as a safety component of a regulated product or are used in one of the eight areas listed in Annex III. The regulation sorts AI systems into risk tiers. At the top are prohibited practices, followed by high-risk systems, then systems with transparency obligations and finally those with minimal risk. Most business applications fall into the lower tiers. The high-risk class covers a clearly defined segment, which nonetheless means a substantial delivery task for the companies concerned.

Two routes lead into this class. The first concerns AI as a safety component of a product that falls under the EU product legislation listed in Annex I, for example machinery or medical devices. The second route runs through the areas in Annex III of the regulation:

  • Remote biometric identification, biometric categorization and emotion recognition (pure verification, for example at building access, is excluded)
  • Operation of critical infrastructure
  • Education and vocational training
  • Employment and workforce management, for example candidate selection or performance evaluation
  • Access to essential services, for example creditworthiness assessment of natural persons or risk assessment and pricing in life and health insurance
  • Law enforcement
  • Migration, asylum and border control
  • Administration of justice and democratic processes

For mid-sized companies, the areas of employment and access to services are the ones most likely to apply to standard applications. Software that pre-screens job applications falls into this category, as does a credit scoring procedure for private customers. Detection of financial fraud is expressly excluded from the credit assessment category.

Article 6(3) exempts systems that pose no significant risk because they perform a narrow procedural task, improve the result of a previously completed human activity, detect decision-making patterns without replacing human assessment, or merely prepare an assessment. Anyone relying on this exemption documents the assessment and registers the system. Systems that profile natural persons always remain high-risk in the Annex III areas.

Classification in four steps

Classification works as a structured process that an organization sets up once and repeats for every new system. Four steps have proven themselves as a framework:

  1. Inventory: record all AI systems, including functions in purchased standard software and applications that business units have procured themselves.
  2. Purpose: for each system, record what it is actually used for and which decisions it influences.
  3. Mapping: check the purpose against Annex I and Annex III, involving legal advice in cases of doubt.
  4. Decision: record the result with its rationale and date as well as the responsible person, and reassess whenever the purpose of use changes.

The first step has the greatest leverage. In many organizations nobody has a complete picture of which AI functions are actually in use. Without an inventory, every classification remains incomplete. The article on shadow AI describes how to detect uncontrolled use and bring it under governance. The article on AI risk management shows how classification feeds into an overarching risk register.

Interactive tool

AI Act risk check

Answer a few questions about your AI system. The check assigns it to a risk class under the AI Act and lists the key obligations. It provides an overview and does not constitute legal advice.

This tool needs JavaScript to make a selection.

What obligations do providers and deployers have?

Providers carry the comprehensive catalog of obligations from risk management to conformity assessment, while deployers are responsible for use in line with the instructions under human oversight. The provider develops a system or has it developed and places it on the market under its own name. The deployer uses a system under its own authority. Most mid-sized companies are deployers.

For providers of high-risk systems, the program includes among other things:

  • a risk management system across the entire lifecycle of the system
  • requirements for data and data governance for training, validation and testing
  • technical documentation and automatic logging
  • transparency and instructions for use for deployers
  • arrangements for human oversight
  • appropriate accuracy, robustness and cybersecurity
  • a quality management system with conformity assessment and registration

Deployers use the system in accordance with the instructions for use. They assign human oversight to people with the necessary competence. Within their own sphere of control they check the input data and monitor operation. They retain the automatically generated logs. Before use in the workplace they inform the affected employees and their representatives. People about whom a high-risk system prepares or takes decisions are informed of its use.

An additional obligation applies to certain deployers: the fundamental rights impact assessment under Article 27. It applies to bodies governed by public law, to private entities providing public services, and to all deployers of systems for creditworthiness assessment or for risk assessment and pricing in life and health insurance. Before first use they describe the deployment, the groups of people affected, the risks to their fundamental rights and the measures for oversight and complaints. The legal assessment of each individual case belongs in the hands of specialist lawyers.

What penalties apply for breaches?

Breaches of the obligations for high-risk AI can be punished with fines of up to 15 million euros or up to 3 percent of worldwide annual turnover, whichever is higher. This range applies to providers and deployers alike. For small and medium-sized enterprises, including start-ups, the lower of the two amounts applies. Supplying incorrect or incomplete information to authorities costs up to 7.5 million euros or 1 percent of turnover. The highest range of 35 million euros or 7 percent is reserved for prohibited practices.

The obligations for high-risk systems apply after a transition period that is shorter for Annex III systems than for safety components under Annex I. The prohibitions and the AI literacy obligation for staff apply earlier. Companies that start with the inventory and classification now use the transition period for the build-up and avoid implementation under time pressure.

Implementation as a delivery task

The typical breaking points lie in operations. The inventory is missing, a business unit procures AI functions without a checkpoint, the designated oversight person has never been trained, and documentation is produced only after the fact. Each of these gaps can be closed with clear responsibility and a fixed place in the process.

A three-stage approach has proven effective. First, classification is embedded as a checkpoint in procurement and approval, so that no system goes into operation without an assessment. Next, responsibilities for oversight and logs as well as for reassessment are assigned. Finally, the people who carry out these tasks are enabled. The article on AI governance shows how to set up the bodies and policies for this.

Niehues Consulting supports companies in exactly this transfer into operations, following the Niehues Delivery Model with the phases Understand, Build, Introduce. Within AI management, the inventory and prioritization of the AI portfolio emerge together with metrics, guided by the test question of who decides, who delivers, who uses. In introduction and operationalisation, Niehues Consulting builds processes and responsibilities and clarifies who owns operations from day one. Business units and oversight staff are enabled until the routines have taken hold in day-to-day business. Within organization design, roles and decision paths are set so that they work reliably.

Frequently asked questions

What is high-risk AI under the AI Act?

High-risk AI comprises systems that either serve as a safety component of a regulated product or are used in one of the areas listed in Annex III, for example recruitment. An extended catalog of obligations applies to them. Classification depends on the purpose of use of each individual system.

What is Annex III of the AI Act?

Annex III is the list of eight areas in which AI systems count as high-risk. They cover biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration plus justice and democratic processes. The Commission can amend the list.

Is AI-supported applicant management a high-risk system?

Yes, Annex III expressly names the filtering and evaluation of job applications. An exemption under Article 6(3) is only available for narrowly defined ancillary functions and must be documented. If the system profiles applicants, the exemption no longer applies.

Is ChatGPT a high-risk AI system?

No, a general-purpose language assistant such as ChatGPT is based on a general-purpose AI model, for which separate provider obligations apply. Such a system only becomes high-risk through its purpose of use. If a company uses it to assess job applications, for instance, the company itself takes on the role of provider of a high-risk system.

What obligations do companies have that only deploy high-risk AI?

Deployers use the system in accordance with the instructions for use and assign human oversight to competent people. They monitor operation, retain the logs and inform employees and their representatives in advance of workplace use. Certain deployers also carry out a fundamental rights impact assessment.

What is a fundamental rights impact assessment and who needs one?

A fundamental rights impact assessment evaluates, before the first use of a high-risk system, which groups of people are affected and what risks to their fundamental rights arise. It is mandatory for public bodies and private providers of public services, as well as for deployers of systems for credit assessment and pricing in life and health insurance. The result is notified to the market surveillance authority.

When do the high-risk AI obligations apply?

The obligations for high-risk AI apply after a transition period that is shorter for Annex III systems than for safety components under Annex I. The prohibitions and the AI literacy obligation apply earlier. Companies should take the binding timetable from the version of the regulation currently in force.

What are the fines for breaching high-risk obligations?

Breaches of the high-risk obligations can lead to fines of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises, the lower amount applies. Supervisory orders up to a ban on use come on top.

Who in the company should be responsible for classification?

Classification is the responsibility of a named person in the line organization, for example the head of the AI portfolio, working closely with legal and IT. Classification belongs in the procurement and approval process. That way, no system goes into operation without an assessment.

Let’s talk about your initiative.

In 30 minutes we clarify where your organization stands and which step will pay off most now.

Request a consultation

30 minutes · free · no obligation