GUIDE EUROPE · AI REGULATION

The European AI Act explained for businesses: system classification, obligations and governance

The European Regulation on Artificial Intelligence (AI Act) introduces fines of up to 7% of global turnover for companies that adopt or develop AI systems without adequate traceability, risk assessment and human oversight measures. The fines are graduated by type of breach, and prohibited practices carry the highest ceilings.

Updated october 2026
8 min read
Altovar Technical Editorial Team

1. What the AI Act regulates and who it applies to

Regulation (EU) 2024/1689, known as the AI Act, governs the development, placing on the market and use of artificial intelligence (AI) systems in the European Union. It does not concern only those who build models or software. A company that uses an AI system in one of its own processes has obligations as a deployer, even if the system was developed by an external supplier and bought as a standard service.

Many businesses assume they are excluded because they develop nothing. That is not the case. Buying an AI function built into a management system, a staff selection tool or a customer service chatbot is a use. The question is not 'do we develop AI?', but 'do we use systems whose outputs can influence people or important decisions?'. If the answer is yes, the regulation must be analysed carefully.

2. The four risk categories

The regulation takes a risk-based approach. Classification depends on the concrete use, not on the product name: the same software can be minimal risk in one department and high risk in another, because its effect on people changes. The table summarises the four categories and their main obligations in brief form. Use it as a starting point, and check the details with your adviser.

Obligations for high-risk systems apply in successive stages, and the timetable should be verified with your adviser, because it depends on how the regulation is implemented. The classification must be made before the system is put into use, not afterwards. A hurried and undocumented classification is one of the most frequent problems to emerge during inspections. Do not wait for an obligation's deadline to start the inventory: it is work that is needed in any case.

CategoryGeneral examplesObligations in brief
Unacceptable riskProhibited practices, such as certain forms of social scoring or harmful manipulationBan: the system cannot be used
High riskSystems used to select staff, evaluate performance or access essential servicesRisk management, data quality, human oversight, logging and information
Limited riskChatbots and systems that generate contentInform people who interact with the AI and flag generated content, in the cases provided for
Minimal riskSpam filters and support functions with no effect on peopleNo specific obligations, but good practice

3. Obligations for businesses using AI

As a deployer you must use the system in line with the provider's instructions, assign oversight to competent people and monitor how it performs over time. For high-risk systems, this also means checks on input data, keeping operational logs for an appropriate period and informing workers when a system is used in the workplace. These measures must be planned before activation.

The regulation also requires a sufficient level of AI literacy for staff who use or oversee these systems. This is not a one-off course: it is the ability to understand what the tool does, what errors it can make and when not to trust its output. Short, targeted and documented training is a concrete first step, and one that is easy to demonstrate in the event of an inspection.

4. Governance: inventory, roles and the relationship with the GDPR

Create a register of the AI systems in use, including what is embedded in software and what employees use on their own initiative, so-called shadow AI. For each system record the supplier, purpose, data processed, people involved, estimated risk level and internal owner. The register is the document that links all other decisions, and it must be updated every time a new tool arrives.

The GDPR applies in parallel. If the system processes personal data, you may need to carry out a data protection impact assessment (DPIA), verify the legal basis for the processing and inform the data subjects. These are not separate tasks to be done in sequence: the AI systems register and the impact assessment should be aligned. Finally, define a simple internal policy, with the approved tools, the data that must not be entered and who authorises new tools.

5. Compliance checklist

Take this checklist to the next meeting with the IT manager, the HR manager and the legal adviser. Every unchecked item is a point to clarify before use, not after. The list does not replace a legal assessment, but it organises the work so that it can be verified. Every item must have an answer that can be demonstrated with a document.

Review the checklist every time a new tool arrives or its use changes. The most important control is often the one on new tools, which frequently arrive without passing through the IT department. Every new tool must go through the same check, even when it is free or is being trialled by a single department. The department introducing it must complete the tool's record before using it.

  • Complete inventory of AI systems, including those embedded in software.
  • Risk classification for each system, justified in writing.
  • List of use cases in which the system influences decisions about people.
  • Internal owner and human supervisor for each high-risk system.
  • Information to employees and customers, where required.
  • Contracts with suppliers stating roles, data processed and documentation available.
  • Training plan and a record of the sessions held.
  • Periodic review of the inventory, because new tools often arrive without a check.

6. Frequent mistakes and underestimated risks

The first mistake is to think the AI Act concerns only technology manufacturers. Anyone who uses a system to select candidates or evaluate staff has obligations of their own, even if the software comes from an external supplier. The second is to rely on the supplier's declaration of conformity without reading what it covers and what it leaves to the customer. Product compliance and compliance of use are different things.

The third is entering personal or confidential data into free tools without a prior assessment. The fourth is not documenting the decision not to classify a system as high risk: if an inspection arrives, the written reasoning is the proof that the thinking was done. Fines are graded according to the type of breach, and prohibited practices carry the highest maximums. Do not estimate your exposure roughly: ask for legal advice on what applies to your company.

7. How to start: an inventory of the systems in use

The first concrete step is to ask each department head which tools they use that contain AI, including those that seem simple: writing assistants, scoring systems, recognition features in security software, the website chatbot. Collect the answers in a table with supplier, purpose, data processed and people involved. This collection usually takes a few meetings, not a long project.

With the inventory, identify the cases that affect people or important decisions: selection, evaluation, access to services, employee monitoring. These are the first to classify with an adviser. The expected outcome is not a certificate, but a map and a reasoned decision for each case, which management can review whenever a use changes. No dedicated software is needed to start.

Frequently Asked Questions

Does using internal chatbots for employees fall under the AI Act?

Yes. It usually falls under the limited-risk systems, with the obligation to let people know they are interacting with an AI system. It becomes high risk if it is used to select staff, assign tasks based on assessments or evaluate performance. The type of use matters more than the tool itself.

Does a management system with built-in AI functions need to be assessed?

Yes. Ask the supplier which functions use AI, what data they work on and what documentation is provided. Then check how you use them: a function that suggests an accounting entry is different from one that ranks candidates. Classification is based on use, not on the name of the module.

Do I need to carry out a data protection impact assessment?

It depends on the data and the type of processing. Where the system processes personal data in a way that may pose a high risk to people, an impact assessment is generally required under the GDPR. Your privacy adviser will decide with you whether it is needed in your case, and the outcome must be documented so that it can be verified.

Can I leave compliance to the software supplier?

Only in part. The supplier has obligations for the product it places on the market, but you have obligations for how you use it: choosing the purpose, supervision, training and information. A clear contract should state who does what and which documents the supplier provides to support your assessment.

What do I do if an employee uses a free tool without authorisation?

It is a compliance and data protection risk. The answer is not simply a ban: first record the tool, assess what data ends up in it and offer an approved alternative. A clear, short policy with a channel for requesting approvals reduces the problem far more than a generic ban that nobody checks.

About this content

General informational guide. Regulations and technical requirements change over time: always check the official sources and assess your specific case with a qualified technical or legal advisor.

GET IN TOUCH

Talk to our team.

We'll show how Altovar integrates into your infrastructure. No commitment, no sales pitch. Just a technical conversation.

BOOK A DEMO →