The AI Act gives a separate role and separate obligations to an organization that uses a ready-made AI system in its own operations, even if it did not develop the model. An accounting firm that puts a chatbot on its website, or a manufacturer that screens resumes with an AI-based tool, already falls within the scope of the regulation, just with different obligations.
What Regulation (EU) 2024/1689 covers
Its full title is Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence. It was published in the Official Journal of the European Union in July 2024 and entered into force on August 1, 2024. As a regulation, it applies directly in every member state and needs no national transposition law, unlike the NIS2 Directive. Member states are responsible for designating the supervisory authorities and for the detailed rules on penalties.
The regulation applies to AI systems. It defines an AI system as a machine-based system that operates with varying levels of autonomy, may adapt after deployment, and infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. The definition is deliberately broad. A rule-based program running on fixed, predefined logic typically falls outside it, while a classifier built with machine learning or an assistant based on a language model falls inside. The European Commission issued guidelines on interpreting the definition in February 2025.
The regulation covers deployers established in the EU, meaning organizations operating here that use AI, and providers that place AI systems on the market or put them into service in the EU, whether they are located inside or outside the EU. It also applies when the output of a system belonging to a provider or deployer outside the EU is used in the EU. It does not apply to systems used for military, defense or national security purposes, to systems developed solely for scientific research, or to people who use AI for personal, non-professional activities. Data protection rules continue to apply alongside it: if an AI system processes personal data, the GDPR applies to it just the same.
Risk categories: prohibited, high-risk, transparency, minimal
The regulation classifies the purpose of use. The same language model falls into the lightest group as the search engine of an internal knowledge base, but used to evaluate employee performance it can be high-risk.
Prohibited practices
Article 5 lists the uses that may not be applied in the EU at all. These include subliminal or deliberately manipulative techniques that distort people’s decisions and cause harm, exploiting vulnerabilities due to age, disability or social situation, and social scoring. Also prohibited are the untargeted scraping of facial images from the internet or camera footage to build facial recognition databases, emotion recognition in the workplace and in education (with an exception for medical and safety purposes), and biometric categorization that infers, for example, religion, political opinions or sexual orientation. Real-time remote biometric identification in public spaces for law enforcement is also prohibited, with narrow exceptions. The 2026 amendment added two items to the list: a system that creates realistic intimate or sexual images, video or audio of an identifiable person without their consent is prohibited, and so is a system that generates child sexual abuse material. These two prohibitions apply from December 2, 2026. Fines for violating the prohibited practices can reach 35 million euros or 7% of total worldwide annual turnover. For large companies the higher of the two applies, for SMEs and startups the lower.
High-risk systems
A system can be high-risk in two ways. The first is when it is a safety component of a product covered by EU product safety legislation, or is itself such a product, for example a machine, a medical device or an elevator (Annex I). The second is when it is used in one of the areas listed in Annex III. Of these, the ones that come up most often in a typical organization are:
- recruitment and workplace decisions: targeting job ads, screening applications, evaluating candidates, decisions on promotion or termination, task allocation, performance monitoring;
- critical infrastructure: systems acting as safety components in the management of critical digital infrastructure, road traffic, or the supply of water, gas, heating and electricity;
- access to essential services: creditworthiness assessment of natural persons, risk assessment and pricing for life and health insurance;
- education and vocational training: admission decisions, evaluating exam results, monitoring behavior during exams.
Being in a listed area does not settle the question on its own. Under Article 6(3), a system can fall outside the high-risk group if it does not pose a significant risk to health, safety or fundamental rights and one of four conditions applies: it performs a narrow procedural task; it improves the result of a previously completed human activity; it detects decision-making patterns or deviations from them but does not replace or influence the previous human assessment without proper human review; or it performs a preparatory task for an assessment. The provider must document this assessment before placing the system on the market, present it to the authority on request, and still register the system in the EU database. If the system profiles natural persons, it always remains high-risk.
Systems with transparency obligations
Article 50 covers systems where the main risk is that people do not know they are dealing with machine-generated content or with a machine. If a chatbot talks with people, they must be told that they are communicating with an AI system, unless this is obvious from the circumstances. The provider of a generative system must mark its output in a machine-readable way. The deployer must disclose when it publishes a deepfake, that is, an image, audio or video that imitates a real person or event in a way that appears authentic, and when it has used AI to generate text that informs the public on matters of public interest, unless a human has reviewed that text under editorial responsibility.
Minimal risk
Most business uses of AI fall here: spam filters, translation tools, code completion, summarizing internal documents. The regulation imposes no new obligations on these and recommends following voluntary codes of conduct.
Alongside the categories, a separate chapter covers general-purpose AI models, meaning foundation models that can be used for many kinds of tasks, such as large language models or image and audio generation models. The technical documentation, copyright and information obligations for these fall on the model provider. Using such a model through an API or as a ready-made service does not make you a model provider.
Provider and deployer: who has which role
The regulation names several actors (importer, distributor, authorized representative), but for an organization that uses AI, two roles are what really count.
The provider is whoever develops an AI system or has one developed and places it on the market or puts it into service under its own name or trademark, even free of charge. The deployer is whoever uses an AI system under its own authority in the course of its professional activity. For most of its AI tools, an organization is a deployer: it buys or subscribes to a service, and its employees work with it.
What the deployer has to do
For a minimal-risk system there is little to do beyond AI literacy, and for a system with transparency obligations, the labeling described above. Article 26 gives the deployer of a high-risk system several tasks:
- operate the system according to the provider’s instructions for use;
- assign human oversight to a person who has the necessary competence, training and authority;
- where the input data is under the deployer’s control, it must be relevant and sufficiently representative for the intended purpose;
- monitor how the system operates, and report risks or serious incidents to the provider and the authority;
- keep the logs the system generates automatically, where they are held by the deployer, for at least six months;
- before putting it into use in the workplace, inform employees and their representatives, and tell affected persons that decisions about them are made with the help of a high-risk system.
Public bodies and private entities providing public services, as well as deployers that assess creditworthiness or life and health insurance risk, must also carry out a fundamental rights impact assessment before putting the system into use. This does not apply to high-risk systems used in critical infrastructure. A person about whom a decision with significant effects was made based on the output of such a system can ask for an explanation of the role the AI system played in that decision.
When a deployer becomes a provider
The role is decided by what the organization actually does. Under Article 25, a deployer becomes a provider if it puts its own name or trademark on a high-risk system, modifies it substantially, or starts using a system that was not originally high-risk for a purpose that makes it high-risk. If, for example, the HR department sets up a general chat assistant to rank candidates, the organization can easily end up with provider obligations: a risk management system, technical documentation, logging, designing human oversight, accuracy and cybersecurity requirements, conformity assessment, and registration in the EU database. That is a much heavier burden than the deployer’s, so it is worth making this decision deliberately.
AI literacy: preparing your staff
Under Article 4, providers and deployers must take measures to support the AI literacy of their staff and of other persons operating AI systems on their behalf. The original text required them to ensure a sufficient level to their best extent. Since the 2026 amendment, the article states that providers and deployers do not have to guarantee a specific level of literacy for any individual, and the Commission publishes practical examples. The obligation applies regardless of risk category, so it also holds at a company where employees only use a general writing assistant. For deployers of high-risk systems, Article 26 additionally requires that the person performing human oversight receives adequate training.
The regulation defines AI literacy as the skills, knowledge and understanding that let someone use AI systems on an informed basis and be aware of their opportunities, their risks and the harm they can cause. The regulation does not prescribe an exam, a certificate or a number of training hours. The expected level depends on the employee’s technical knowledge, experience and the context of use, and also on who is affected by the system’s output.
In practice, this means different content for different roles:
- for every user: what the tool can and cannot do, why a confident but wrong answer sounds believable, what data must not go into it, and where the line runs between their own decision and a machine suggestion;
- for those who make decisions about people based on AI output: the sources of bias, what human oversight actually involves, and documenting the decision;
- for IT and procurement: recognizing the roles (provider or deployer), evaluating provider documentation, and managing logging and access.
The European Commission’s AI Office has published a Q&A on interpreting Article 4 and collects organizations’ training practices. It is worth recording the training content, the participants and the dates in writing, because in a regulatory or internal investigation this is what shows what the organization did.
Timeline: what applies when
The provisions of the regulation become applicable in stages. The first three dates have already passed; the two deadlines for high-risk systems are still ahead:
- February 2, 2025: the general provisions, including AI literacy, and the prohibited practices.
- August 2, 2025: the rules for providers of general-purpose AI models, the governance system and the penalties.
- August 2, 2026: most of the regulation, including the transparency obligations. Generative systems placed on the market before this date have until December 2, 2026 to mark generated content in a machine-readable way.
- December 2, 2027: the requirements for high-risk systems under Annex III.
- August 2, 2028: the rules for high-risk systems linked to the product legislation listed in Annex I.
The last two dates were pushed back by Regulation (EU) 2026/1744, the so-called digital omnibus on AI. The Commission proposed it in November 2025, the Parliament and the Council adopted it in June 2026, and it entered into force on July 27, 2026. The original deadlines were August 2, 2026 for Annex III systems and August 2, 2027 for Annex I systems, which is why older materials still give these dates. For planning, work from the consolidated text available on EUR-Lex.
Internal AI inventory and usage policy
None of the regulation’s obligations can be met unless the organization knows which AI systems it uses. In most places this does not match the list of officially procured tools. Chat assistants that employees use with their own accounts, browser extensions, and AI features that reached existing software through updates are all part of the picture. The article on shadow AI covers in more detail how to uncover this hidden use.
What goes into the inventory
A spreadsheet is enough if it records the following for each tool:
- the name of the system, its provider, and the version or subscription type in use (business or personal account);
- who uses it, for what task, and what data goes into it (personal data, trade secrets, customer data);
- the organization’s role: deployer or provider, and if the way the tool is used could change that, the reasoning;
- the risk classification under the regulation, with a short justification of the decision;
- the responsible person, where the contract and the data processing terms are kept, and the review date.
The justification for the classification matters at least as much as the category itself. If a tool is not high-risk because a person always reviews its output and it only serves as a suggestion, write that down, and day-to-day practice has to match it.
What the usage policy should settle
The inventory shows what needs rules. An AI usage policy typically settles these questions:
- which tools are allowed, which only with a business account, and which are prohibited;
- which data categories must not go into an AI tool, and which may go in after anonymization;
- when human review of AI output is mandatory, and who is responsible for it;
- how AI-generated content published externally must be labeled;
- whether using AI for decisions about people (hiring, evaluation) is prohibited or allowed only with separate approval;
- how an employee can request a new tool, and who decides on it.
The policy works when employees know it and their training builds on it. Review the inventory and the policy once a year, and also whenever a new AI tool or a significant new feature is introduced.
Related service: Shadow AI assessment with an AI usage policy →