Introduction to the EU AI Act

The European Union (EU) has one law about artificial intelligence that applies to every member state at once: Regulation (EU) 2024/1689, called the AI Act. It has applied in steps since February 2025, and it is the main example of regulation for AI use at work. In this lesson we read the Act the way a knowledge worker needs to. We start from the tool on your desk and ask what the Act says about it: the definition of an AI system, the risk tiers, the provider and deployer roles, and the duties that reach your daily work. Where the Act sets a duty, the lesson links to the article, so that you can read the sentence that binds you instead of a paraphrase of it.

This lesson is specific to the EU on a site that is otherwise global. If your organization is outside the EU, the Act can still apply, because it covers providers that sell into the EU and any provider or deployer whose system’s output is used in the EU (Article 2(1)). It does not apply to a person using AI for a purely personal, non-professional activity (Article 2(10)), so it is a law about professional and public use.

The lesson is general information, and it is not legal advice. The Act was amended in July 2026, guidance from the Commission keeps coming, and a decision about your organization needs a lawyer who knows your member state. Every article number and date below comes from the consolidated text of the Regulation as of July 27, 2026 [1], and the page shows a review date for that reason.

One duty explains why this page exists. Article 4 asks providers and deployers to “take measures to support the development of AI literacy” of their staff and of anyone operating AI systems for them, fitted to their role and to the context of use (Article 4). The same article says the duty does not require any specific level of literacy from any individual. A course like this one is one way an organization meets it.

Is it an AI system at all?

The rules in this lesson apply to an AI system, and Article 3(1) defines one as a machine-based system, designed to operate with some autonomy, that “infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions” [1]. The word that matters is infers. The system works out how to produce its output from the input, rather than following a fixed procedure a person wrote down. The Act also has rules for general-purpose AI models, the engines behind chat assistants, in Chapter V. Those fall on the model’s provider, and this lesson doesn’t cover them.

Recital 12 of the original text explains what that excludes: systems “based on the rules defined solely by natural persons to automatically execute operations” [2]. A spreadsheet macro that applies fixed rules is not an AI system. A rules engine with two hundred hand-written if-then rules is not one either. A model trained on past data that predicts, ranks or generates is one, and so is a chat assistant. For the border cases, Article 96(1)(f) asks the Commission to publish guidelines on how the definition applies [1]. Start from the Commission’s AI Act pages when you need them [3].

Checkpoint · choice

Is this an AI system?

Your team lists the software it uses for a compliance inventory. Which of these is an AI system under Article 3(1)?

A thermostat that switches the heating on whenever the room drops below 19 degrees, on a schedule your team set.A tool that predicts the cost center of each incoming invoice, from the invoices that were coded by hand last year.A mail rule that moves every message from a named sender into a folder, set up by hand in the mail client.

CheckHintSkip

The risk tiers

The Act regulates uses of AI rather than the technology, and it sorts those uses into tiers. To place a use, walk down this list, and stop at the first tier that fits.

1. Prohibited. Article 5 lists practices no one may put on the market or use.
   • Manipulation or deception that causes “significant harm”.
   • Exploiting people’s vulnerabilities due to age, disability, or social or economic situation, causing “significant harm”.
   • Social scoring that leads to unfair treatment.
   • Predicting a crime from a person’s profile or personality traits alone.
   • Building face-recognition databases by scraping images from the internet or closed-circuit television (CCTV).
   • Inferring the emotions of people at work or in education, except for medical or safety reasons.
   • Sorting people by their biometric data into categories such as race, religion, or sexual orientation.
   • Added in July 2026: sexual deep fakes made without consent, and material showing sexual abuse of children.
2. High-risk. Article 6 gives a product route and a use-case route. The product route: the AI is a safety component of a product, or is itself a product, that EU product law in Annex I already requires a third party to certify. Toys, lifts, and medical devices are on that list. The use-case route: Annex III names areas where AI decisions touch people’s lives. Its headings are biometrics, critical infrastructure, education, employment and workers’ management, access to essential services such as credit and insurance, law enforcement, migration and border control, and justice and democratic processes. Ranking job applicants and scoring a consumer’s creditworthiness are both on the list. Article 6(3) lets a provider argue that an Annex III system isn’t high-risk when it only does a narrow procedural task or prepares a human assessment, but a system that profiles people is always high-risk.
3. Transparency duty. Article 50 puts a duty to inform on systems that talk to people, generate content, recognize emotions, or produce deep fakes, whatever their tier. The section below on everyday work is about this tier.
4. Everything else. Most AI in use falls here, and the Act doesn’t set a specific duty for it. The Commission gives spam filters and AI in video games as examples [3]. The AI literacy duty of Article 4 still applies to the organization, because it attaches to the provider and the deployer rather than to the tier.

A tier is a property of a use, so the same model can be in more than one tier at once. A general-purpose chat assistant is a transparency-tier system when it answers customers, a high-risk system when it ranks candidates, and an everything-else system when it drafts your meeting notes.

Checkpoint · sort

Which tier?

Unplaced (drag an item to a bucket, or select it and then click the bucket)

A school's camera system that infers whether students are bored or frustrated during lessonsA city that scores residents on their online behavior and denies the low scorers a place in housing programsAn insurer's model that sets the price of a life insurance policy from the applicant's health dataA university's tool that flags students who may be cheating during an online examA city council's website assistant that answers residents' questions about waste collectionA warehouse system that predicts how many boxes each product sells next week

Prohibited

High-risk

Transparency duty

No specific duty in the Act

CheckHintSkip

The provider and the deployer

The Act puts duties on roles, and the two that matter for most organizations are defined in Article 3. A provider 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, whether for payment or free of charge” (Article 3(3)). A deployer uses an AI system “under its authority”, outside a personal non-professional activity (Article 3(4)) [1]. Ask: did we build this, or sell it as ours? Then we’re the provider. Do we use it in our work? Then we’re the deployer. A vendor that sells a chatbot is its provider, and the hospital that books appointments with it is a deployer.

One organization can hold both roles, and a deployer can turn into a provider without noticing. Article 25(1) names the moments. You put your own name or brand on a high-risk system someone else built. You make a substantial modification to a high-risk system. You change the intended purpose of a system so that it becomes high-risk. That last one is the everyday case. A team that buys a general-purpose chat tool for drafting and then wires it into hiring decisions has changed its purpose into an Annex III use, and the team’s organization now owes the provider duties for it.

Pitfall: Renaming a vendor's tool

A company licenses a CV screening service, which ranks job applicants and so is an Annex III use, gives it its own product name, and offers it to customers as its own. When a customer asks who certified the system, the company points at the vendor. Under Article 25(1)(a), a party that puts its own name or trademark on a high-risk system already on the market is treated as its provider, unless the contract allocates the duties otherwise, and the original vendor stops being one for that system. Before you brand a tool as yours, check which role the branding gives you.

Checkpoint · match

Which role?

Match each organization to its role under the Act.

A startup tunes an open model into a contract-review assistant and sells it to law firms under its own nameChoose…ProviderDeployerBoth

A clinic uses a vendor's triage model exactly as delivered to sort incoming patient messages by urgencyChoose…ProviderDeployerBoth

A city council buys a general chat assistant for drafting letters and later uses it to rank applications for social housingChoose…ProviderDeployerBoth

CheckHintSkip

What each role owes

For a high-risk system the duties are substantial, and this lesson only gives their headings. Each one links to its article. Engineers who build such a system read the articles themselves.

A provider of a high-risk system must run a risk management system (Article 9), govern its training data (Article 10), write technical documentation (Article 11), make the system keep logs (Article 12), give deployers instructions for use (Article 13), design for human oversight (Article 14), reach an appropriate level of accuracy, robustness, and cybersecurity (Article 15), pass a conformity assessment (Article 43), register the system in the EU database (Article 49), monitor it after release (Article 72), and report serious incidents to the authorities (Article 73). Article 16 collects the list.

A deployer of a high-risk system has a shorter list, in Article 26. Use the system as the instructions for use say. Assign human oversight to people who have “the necessary competence, training and authority”. Keep the logs the system generates for at least six months. Monitor the system, suspend its use when it presents a risk, and tell the provider and the authority about serious incidents. Tell workers and their representatives before a high-risk system is used on them at work, and tell any person that a decision about them involved the system. A deployer that is a public body, or that provides a public service, or that scores credit or prices life and health insurance, also carries out a fundamental rights impact assessment before first use (Article 27).

Read that deployer list again with the earlier safety lessons in mind. Human oversight by competent people is the human in the loop from Why agent safety is different. Log retention is logging and audit. Suspending use and informing the provider is escalation. The Act turns the habits of this course into duties, for the high-risk tier, and a written policy is how an organization shows it meets them.

Transparency duties in everyday work

Article 50 is the part of the Act that most people meet, because it applies whatever the tier and its duties are simple.

• A provider of a system that interacts with people must design it so that people know they are dealing with an AI system, unless that is obvious to a reasonably observant person (Article 50(1)). A customer-service chatbot says that it is one.
• A provider of a system that generates text, images, audio, or video must mark the output “in a machine-readable format and detectable as artificially generated or manipulated” (Article 50(2)). Content credentials are one technical answer to this duty.
• A deployer of an emotion-recognition or biometric-categorization system must tell the people exposed to it (Article 50(3)).
• A deployer who publishes a deep fake, which Article 3(60) defines as generated or manipulated content that resembles real people, places or events and “would falsely appear to a person to be authentic”, must disclose that the content is generated or manipulated (Article 50(4)). For a work that is clearly artistic or satirical, a disclosure that doesn’t spoil the work is enough.
• A deployer who publishes AI-generated text “with the purpose of informing the public on matters of public interest” must disclose that, unless a person reviewed the text and holds editorial responsibility for it (Article 50(4)).

The information goes to the people concerned “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure” (Article 50(5)) [1]. Most everyday use at work falls under none of these. Drafting a memo with an assistant is neither a deep fake nor a publication on a matter of public interest, so the disclosure habit from What may go into an AI tool stays a matter of what your audience expects rather than of law. The moment the output is a chatbot for the public, a realistic image of a real person, or a public article, Article 50 applies.

Checkpoint · match

Which transparency duty applies?

For each situation, pick the Article 50 duty that applies.

Your company puts a chatbot on its public website to answer customer questionsChoose…People must be told they are dealing with an AI systemThe output must carry a machine-readable mark that it was generatedThe content must be disclosed as a deep fakeThe text must be disclosed as AI-generatedNo Article 50 duty

A vendor sells an image generator to design agenciesChoose…People must be told they are dealing with an AI systemThe output must carry a machine-readable mark that it was generatedThe content must be disclosed as a deep fakeThe text must be disclosed as AI-generatedNo Article 50 duty

Marketing publishes a realistic AI-made video of a real actor for an advertisementChoose…People must be told they are dealing with an AI systemThe output must carry a machine-readable mark that it was generatedThe content must be disclosed as a deep fakeThe text must be disclosed as AI-generatedNo Article 50 duty

A news site publishes an AI-drafted article about a local election after an editor reviewed it and takes responsibility for itChoose…People must be told they are dealing with an AI systemThe output must carry a machine-readable mark that it was generatedThe content must be disclosed as a deep fakeThe text must be disclosed as AI-generatedNo Article 50 duty

You use an AI assistant to draft a memo for your own teamChoose…People must be told they are dealing with an AI systemThe output must carry a machine-readable mark that it was generatedThe content must be disclosed as a deep fakeThe text must be disclosed as AI-generatedNo Article 50 duty

CheckHintSkip

The timeline

The Act applies in steps, set out in Article 113 and changed by the July 2026 amendment, Regulation (EU) 2026/1744, which moved the high-risk dates back [1]. Today is after the general application date, so most of the table is already in force.

From	What applies
August 1, 2024	The Act entered into force, twenty days after publication on July 12, 2024 [3]
February 2, 2025	Definitions, the AI literacy duty, and the prohibited practices (Chapters I and II)
August 2, 2025	Rules for general-purpose AI models, the governance bodies, and the chapter on penalties
August 2, 2026	The Act as a whole, including the transparency duties of Article 50
December 2, 2026	The prohibitions added in 2026 on sexual deep fakes and child abuse material (Article 5(1)(ba) and (bb)). Generators already on the market before August 2, 2026 must meet Article 50(2) by this date too (Article 111(4))
December 2, 2027	The high-risk duties for Annex III systems
August 2, 2028	The high-risk duties for AI in Annex I products

Article 111 gives transition periods to systems that were already in use, and the Commission keeps a page with the current state of application [3]. Dates have moved once already, which is why this page has a review date.

Penalties

Article 99 sets the ceilings, and each member state writes the rules for applying them. A prohibited practice can cost up to 35 million euros or 7% of worldwide annual turnover, whichever is higher. Breaking the provider duties, the deployer duties of Article 26 or the transparency duties of Article 50 can cost up to 15 million euros or 3%. Giving authorities wrong or incomplete information can cost up to 7.5 million euros or 1%. For small and medium-sized enterprises the lower of the amount and the percentage applies [1]. The order tells you what the legislator considers worst: the banned practices are at the top, and the everyday transparency duties are in the same band as the high-risk duties.

Who enforces it

Each member state designates at least one notifying authority and at least one market surveillance authority for the Act (Article 70). Those national authorities are who a deployer reports a serious incident to, and who can fine an organization in that state. At the EU level the Commission builds expertise through its AI Office (Article 64), and it has the exclusive power to enforce the rules for general-purpose AI models, a task it entrusts to the AI Office (Article 88(1)). The European Artificial Intelligence Board, with one representative per member state, coordinates the national authorities (Article 65) [1]. The Commission describes the AI Office’s powers over general-purpose models: requesting documentation, evaluating models, requiring corrective action and issuing fines [4]. For your own country, look up which market surveillance authority your member state designated. Article 70(2) asks each member state to publish how its authorities can be contacted and to name one point of contact [1].

Exercise

Pick one AI tool used at your work. In two lines, decide whether your organization is its provider or its deployer, and name the one Article 50 transparency duty that applies to your use of it, or write that none applies. Ten minutes at most. Those lines are the first entry of an inventory your organization needs for the AI literacy and deployer duties, and writing one shows you how quickly the roles become clear. A good result gives the reason with the role (“deployer: we bought the tool unchanged”) and the paragraph with the duty (“Article 50(1): a chatbot for customers”). Which of the Act’s duties touch your own daily work, and which only your organization’s?

Stretch: Check whether the tool's use appears in Annex III. If it does, write down which of the deployer duties of Article 26 your organization already meets and which it does not.

Recap

1. The Act applies to AI systems, which Article 3(1) defines by their ability to infer outputs from input. A tool that only executes rules a person wrote is outside it [1].
2. Uses fall into tiers: prohibited practices (Article 5), high-risk systems (Article 6 with Annex I and Annex III), systems with a transparency duty (Article 50), and everything else, for which the Act doesn’t set a specific duty [1].
3. Providers build or brand a system and deployers use one. Changing a system’s purpose into a high-risk use, or putting your name on it, makes you its provider (Article 25) [1].
4. A deployer of a high-risk system follows the instructions for use and assigns competent human oversight. It also keeps the logs for at least six months and reports incidents, and it tells the people affected (Article 26) [1].
5. Article 50 reaches everyday work through public-facing chatbots, generated media, deep fakes and public-interest text. Ordinary drafting has no Article 50 duty [1].
6. The AI literacy duty of Article 4 has applied since February 2, 2025. High-risk duties start on December 2, 2027 and August 2, 2028, and dates have moved once, so check the review date on this page [1].

You can now

• Sets policy, logging and escalation for agents in an organization
• Discloses AI use where the audience expects it

Mark lesson finishedI know this, skip it

References

1. European Parliament and Council of the European Union. Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text of 27 July 2026. EUR-Lex. Reference. AI Act
2. European Parliament and Council of the European Union. Regulation (EU) 2024/1689 (Artificial Intelligence Act), as published in the Official Journal on 12 July 2024, with recitals. EUR-Lex. Reference. AI Act OJ
3. European Commission. AI Act, the regulatory framework for artificial intelligence. European Commission, Shaping Europe's digital future. Reference. EC AI Act
4. European Commission. European AI Office. European Commission, Shaping Europe's digital future. Reference. EC AI Office