# EU AI Act: What Companies Need to Know About the AI Regulation

Author: Marin Niehues, Niehues Consulting · Published: 2024-09-23 · Updated: 2026-11-16 · AI Law, Risk & Governance
Source: https://www.niehuesconsulting.com/blog/eu-ai-act/

The EU AI Act is the European regulation that sorts AI systems into risk classes and places specific obligations on companies according to their role and use case. Formally it is Regulation (EU) 2024/1689. It applies directly in all member states and also reaches companies outside the EU as soon as their AI systems are offered in the European market or their output is used there. This article provides an overview and does not constitute legal advice.

For the executive board this creates a management task: which AI applications are running in the company, which class does each fall into, and who is accountable for the consequences? The article sets out the structure of the regulation and shows how it translates into a controllable approach.

**A company that knows the risk class of every AI system in use, and has clarified its own role for each, meets the obligations of the regulation in a targeted way and with manageable effort. Classification is the first work step, operation with clear responsibilities the second.**

## The structure of the regulation: risk determines the obligations

The regulation follows a risk-based approach. The greater the potential harm to health and safety or to fundamental rights, the stricter the requirements. For decision-makers the principle has a practical advantage, because it orders the work. Before a company considers individual obligations, it needs to know which AI systems it uses and which tier each belongs to.

The top tier consists of practices the regulation classes as unacceptable and prohibits, such as certain forms of social scoring. Below them sit high-risk systems. These include AI applications in sensitive fields such as recruitment or education, in critical infrastructure, and as a safety component of regulated products. They are subject to requirements on risk management and data quality, on documentation and human oversight, and on robustness.

Certain systems are subject to transparency obligations. With chatbots, it must be clear that a person is interacting with AI. Synthetically generated content must be marked in a machine-readable way, and deepfakes must be disclosed. Most of today's applications, such as spam filters, fall into the minimal-risk tier, for which the regulation sets no specific obligations. A separate regime covers general-purpose AI models and is aimed primarily at their providers.

**The four tiers at a glance**
- Unacceptable risk: prohibited practices, such as certain forms of social scoring.
- High risk: sensitive fields of use such as recruitment or critical infrastructure, with extensive requirements for process and evidence.
- Transparency obligations: chatbots, synthetic content and deepfakes with disclosure rules.
- Minimal risk: applications without specific obligations under the regulation.

*Interactive tool: AI Act risk check, assigns an AI system to a risk class under the AI Act (prohibited, high risk, transparency obligations, minimal) based on a few questions and lists the key obligations. Overview, not legal advice. Available at https://www.niehuesconsulting.com/blog/eu-ai-act/*

## When does the EU AI Act apply?

The EU AI Act applies in stages: its obligations become applicable in a fixed order according to risk and addressee. The regulation was published in the Official Journal of the EU in 2024, and its individual parts apply step by step.

**Order of application**
- First: the bans on unacceptable practices and the duty to ensure AI literacy among staff.
- Next: the rules for general-purpose AI models, along with the provisions on supervision and penalties.
- Then: most of the remaining obligations, including the requirements for high-risk systems in the sensitive fields of use.
- Last: the obligations for AI as a safety component of products already covered by their own EU product legislation.

The binding dates are set out in the regulation itself. They can only be moved by an amending act of the European legislator, so planning must always rest on the version in force in the Official Journal. For the executive board this yields a clear rule: the inventory and AI literacy come first, because they apply first and every later obligation builds on them.

## Who is a provider, who is a deployer?

A provider develops an AI system, or has one developed, and places it on the market under its own name, while a deployer uses such a system under its own authority in the course of business. Most mid-sized companies and large groups are deployers, because they use systems from software vendors or model providers.

The role can change. A company that markets a high-risk system under its own name, substantially modifies it, or repurposes a system so that it becomes high-risk is treated as a provider under the regulation and takes on the provider's obligations. The company settles this question before introduction, so that it does not surface during live operation.

For deployers of high-risk systems the obligations lie mainly in organizing the use. The system is used as intended, qualified people exercise human oversight, operation is monitored and logged, and employees are informed about the use. In addition there is the AI literacy duty under Article 4, which applies to all deployers and providers regardless of risk class.

These obligations are organizational tasks at their core. Oversight and monitoring only work if someone takes them on with the necessary authority and the process is anchored in day-to-day operations. This is where the legal text turns into delivery.

## What are the penalties for breaching the AI Act?

Breaches of the AI Act can lead to fines of up to 35 million euros or 7 percent of worldwide annual turnover, whichever is higher. This ceiling applies to prohibited practices. The regulation grades penalties by the severity of the breach.

**Fines under the regulation**
EUR 35m or 7% | Ceiling for prohibited practices, the higher amount applies
EUR 15m or 3% | Ceiling for breaches of the other obligations, for instance as deployer of a high-risk system
EUR 7.5m or 1% | Ceiling for incorrect or incomplete information supplied to authorities

For small and medium-sized enterprises, including start-ups, the lower of the two amounts applies in each case. Beyond the fine, the operational consequences matter to the executive board: an authority can require a system to be withdrawn from the market or its use to be stopped. How fines are calculated in individual cases is a matter for specialist lawyers.

## From regulation to delivery: a four-step approach

Companies that implement the regulation in a controlled way proceed in a fixed order.

1. Inventory: it records every AI system in use, including systems developed by the company itself, embedded functions in purchased software and applications that business units have introduced on their own. Use is regularly broader than central IT knows.
2. Classification: for each system, the intended purpose and the company's role are documented together with the risk class. Business units and IT work with legal expertise, because the intended purpose determines the classification and the business unit knows it best.
3. Responsibilities: who approves a new system? Who exercises human oversight? Who reports incidents, and to whom? These questions belong in a governance model with named roles.
4. Anchoring in operations: this includes approval processes for new applications, reporting with indicators, training for the roles concerned and a fixed rhythm in which the executive board reviews the AI portfolio.

Only with the fourth step does an inventory become a procedure that also works for new systems and changing personnel. The legal assessment of individual cases rests with qualified lawyers.

## Where implementation stalls in practice

Three patterns recur. First, the inventory remains incomplete because it is set up as a one-off survey. Second, responsibility sits with a single function, such as the legal department or data protection, while decisions on use are taken in the business units. Third, policies sit on the intranet because nobody translates them into the workflow.

From a delivery perspective the remedy is the same in all three cases: each requirement is translated into a role and a process and given an indicator. An inventory has an owner and an update rhythm. An approval has a process with deadlines. A training course has a target group and a measure of success.

**Three routes to implementation compared**
Criterion | Legal review alone | Individual measures in the business unit | With Niehues Consulting
Starting point | Legal text | Single application | Inventory and portfolio of all AI systems
Result | Legal opinion | Local rules | Roles and processes with indicators in regular operation
Decision paths | Open | Different in each unit | Uniformly defined and named
Anchoring in operations | Downstream | Selective | Introduction with enablement of the business units
Link to AI strategy | Weak | Weak | Sequence of applications and business case from one source

## Niehues Consulting as a partner for implementation

The requirements of the regulation can be met reliably when AI management and organization design work together. [Niehues Consulting](/) in Düsseldorf combines both from the delivery perspective. Founder Marin Niehues builds organizations that deliver and brings AI into operation, in projects ranging from VC-funded scale-ups to DAX 40 groups across 20+ countries on six continents.

In AI management, the inventory of applications is built with business analysis and requirements, followed by the business case, a reasoned sequence and portfolio steering. In organization design, roles and decision paths are defined so that approval and oversight work in daily business. The approach follows the Niehues Delivery Model with its three phases. Three of its review questions matter most for implementing the regulation: Who decides, who delivers, who uses? Which roles and decision paths does operation require? Who owns operation from day one? A free 30-minute initial consultation is the starting point and can be arranged via the [contact page](/#kontakt).

**What decision-makers should take away**
The regulation sorts AI systems into risk classes and assigns graduated obligations to providers and deployers, which apply in stages. For the executive board this results in a clear work plan: build an inventory and classify, then set responsibilities and anchor the rules in operations. A company that backs these four steps with named owners and indicators creates the basis for expanding its use of AI safely.

**Sources**
- Regulation (EU) 2024/1689 laying down harmonized rules on artificial intelligence (Artificial Intelligence Act), Official Journal of the European Union, 2024

## Frequently asked questions

### What is the EU AI Act?

The EU AI Act is Regulation (EU) 2024/1689 on artificial intelligence. It creates a uniform legal framework for the development, distribution and use of AI systems and divides them into classes according to risk. The higher the risk, the more extensive the requirements for documentation and oversight.

### Who does the EU AI Act apply to?

The regulation applies to providers and deployers of AI systems that are placed on the market or used in the EU. This includes companies based outside the EU where the output of their systems is used in the EU. For most mid-sized companies and large groups the deployer role is the relevant one.

### Which AI systems count as high risk?

High-risk systems are AI applications in fields of use that the regulation classes as particularly sensitive, such as recruitment and employee evaluation, credit assessment, education or critical infrastructure. AI used as a safety component of regulated products is also covered. What matters is the specific intended purpose of the system within the company.

### When does the EU AI Act come into force?

The EU AI Act applies step by step, starting with the prohibitions and the AI literacy duty. The rules for general-purpose AI models follow, then the high-risk obligations. The binding dates are set out in the version of the regulation in force in the Official Journal of the EU.

### Can the AI Act be delayed?

The application dates of the AI Act can only be moved by a formal amending act of the European Parliament and the Council. Planning should therefore rely solely on the version in force in the Official Journal. Inventory, classification and AI literacy pay off regardless of the date, because every later obligation builds on them.

### What are the fines under the EU AI Act?

Prohibited practices can draw fines of up to 35 million euros or 7 percent of worldwide annual turnover, and breaches of the other obligations up to 15 million euros or 3 percent. For incorrect information supplied to authorities the ceiling is 7.5 million euros or 1 percent. For small and medium-sized enterprises the lower amount applies.

### Does the AI Act apply to using ChatGPT at work?

Yes, a company that uses ChatGPT or a comparable assistant for work is a deployer of an AI system and must at least ensure the AI literacy of its staff. If it uses the assistant with customers, transparency obligations are added. If it is used for a high-risk purpose such as screening job applications, the high-risk obligations for deployers apply. The obligations for the underlying model lie with its provider.

### Which authority supervises the AI Act?

Each member state designates its own market surveillance authorities. In Germany the legislator sets out responsibility in a separate implementing act, and the Federal Network Agency is regarded as the obvious choice for the central authority. Existing sector supervisors remain responsible for individual areas such as financial services. At European level, the Commission's AI Office supervises providers of general-purpose AI models.
