AI Act Glossary: Do You Speak the Regulator's Language?

AI Act Glossary

When drafting IT contracts or auditing corporate processes, I often find that the biggest communication noise arises not between humans and machines, but between business and lawyers. With the arrival of the European AI Act regulation, which formally applies since August 2024 but its full enforceability is launching gradually until 2027, orienting yourself in the new terminology is a must. If you want to correctly classify the systems you use in your company, you must first speak the same language as the European regulator.

Article 3 of the regulation itself defines a total of 68 terms that form the legislative basis of the entire text. It doesn't make sense to go through them all, but without understanding the key ones, you won't get far in implementing internal rules. Let's focus on the most important terms and their real impact on your practice.

AI System

The first building block is the **AI System** itself. Forget definitions from science fiction movies; for the regulator, it is a machine-based system designed to operate with varying levels of autonomy and which, after deployment, can exhibit adaptiveness, and which infers from its inputs how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. In practice, this means that an AI system is not only an advanced neural network like GPT-4, but easily a simple algorithm in your e-shop that recommends related items based on previous purchases.

Today, we come into contact with artificial intelligence systems on a daily basis without thinking about it: whether it's unlocking a phone using Face ID, spam filters in your email inbox, or recommendation algorithms on social networks.

Key Roles in the AI Ecosystem

Provider

The provider is the one who develops an AI system and places it on the EU market under their brand. They bear the greatest share of responsibility and obligations. If your IT department programs its own internal AI tool for customer behavior analysis and you start offering it commercially, you become a provider with all obligations – from risk assessment to technical documentation.

Deployer

The deployer is a person who actively uses AI systems for their professional activities. In my office, clients often ask: *Does this apply to our employees playing with ChatGPT?* No, private and non-commercial use is exempt from the regulation. However, if you integrate that same ChatGPT into your corporate CRM to automatically reply to clients, you become a deployer. You are responsible for ensuring your people use the tool in accordance with the rules and that they are sufficiently trained.

This term replaced the previous unfortunate designation "user", which created the false impression that regulation falls on every average citizen writing poems with AI for fun.

Distributor

A distributor is an entity in the supply chain that supplies an AI system to the market without being its creator or importer. They do not participate in the development process in any way, nor do they change the properties of the AI system; their role is purely to distribute the product. Obligations are also associated with this role – the distributor must, for example, physically verify that the product has the necessary CE conformity marking and documentation.

Importer

An importer places an AI system on the European Union market from a provider located outside the EU. If you buy an enterprise license for an AI system from a US startup in Silicon Valley that does not have representation in the EU, the importer must guarantee that the software meets EU standards before anyone runs it in the EU for the first time.

Operator

Operator is an umbrella term that groups all the roles listed above (providers, deployers, importers, and distributors) in the text of the regulation. It is used when an obligation applies to the entire lifecycle of the software.

Risk

The entire AI Act is built on the logic of risk profiles. It is not about *what* the technology can do, but *how* it can cause harm. The regulation defines four basic risk categories. Fortunately, most systems fall into **minimal risk** (like spam filters) or **limited risk** (e.g., chatbots on websites, where you must only inform the customer they are not talking to a human). A fundamental turning point occurs with **high risk**. This covers anything that affects safety and human lives – including candidate screening systems in HR. If a system evaluates someone incorrectly, it can ruin their career, which is why the rules are strictest here. The final group is **unacceptable risk** (such as social scoring or predictive policing), which EU law bans outright.

Lifecycle and Market

It is interesting how the Act defines time milestones. **Placing on the market** is the moment when a provider makes an AI system available in the EU for the first time. Any subsequent distribution is referred to as **making available on the market**. However, a crucial term is **putting into service**. This is the moment a deployer starts using the AI system for their professional activities. This concept prevents companies from avoiding regulation by developing a high-risk system internally 'for their own use' – rules will apply to them from day one of deployment.

Other Important Terms

Intended Purpose

Refers to the specific use of the AI system for which its provider developed it. It is based on what the manufacturer states in official documentation. Precise definition of the purpose is the basis for assessing risk and safety.

Substantial Modification

It is a change to the system after it has been put into service which the original provider did not foresee and which changes the intended purpose of the software. The AI Act establishes a strict rule here: whoever carries out such a substantial modification takes on full responsibility and all obligations of the original provider.

Post-Market Monitoring

Developers are obliged to continuously collect and evaluate experiences from the real operation of their systems. If they encounter new risks in practice, they must react immediately – for example, by issuing a security update or withdrawing the tool from the market.

General Purpose AI (GPAI)

Classic AI models usually focus on one activity. General purpose AI models are universal large systems trained on huge amounts of data (such as OpenAI's GPT series) that can perform a wide range of tasks from writing code to translation. Due to their massive market influence, the regulator pays special attention to them and imposes specific obligations on their creators regarding documentation of training data.

Deep Fake

Synthetic content (image, video, audio recording) that mimics real persons or events and could appear authentic to a person. The AI Act does not ban deep fakes, but imposes a strict transparency obligation – the content must be clearly marked as generated to prevent fraud and disinformation.

AI Literacy

For all of this to work in a company, the legislation imposes an AI literacy obligation. This means it is not enough to just buy licenses – as a manager, you must ensure that your employees who work with AI are sufficiently trained, understand how the system works, what its limits are, and how to safely handle data.


Speaking the regulator's language is not a luxury; it is necessary prevention. If you are not sure which role your company falls into or what obligations apply to you when deploying external AI tools, I recommend starting not by writing guidelines, but by conducting a thorough review of your existing processes and non-disclosure agreements (NDAs).

← Back to Blog Overview