The EU AI Act
- Lesson 5
- 35 min
- Module 2: Laws and standards in plain words
Where we are
In the last lesson, we drew the map: a law binds, while a standard or a framework is chosen. The document assistant's row still has an empty cell, its category under the EU AI Act. In this lesson, we read that law in plain words: its categories, its two main roles and its dates as changed in 2026, starting with the rules for selling food.
Lesson 5 of 19 · AI Governance: Policies, Risk and AccountabilityThe EU AI Act
1. What governance is for
2. Laws and standards in plain words
3. Inventory, risk tiers and impact assessment
4. Policies people can follow
5. Controls, documentation and evidence
What you will be able to do
Larkfield sells to customers in the European Union, so one law matters at once. You learn what the EU AI Act asks, of whom and from when, and you place one Larkfield system in it.
Before you startWhat to know and to have ready
What you need to know
- The six kinds of rule and the four questions, from A map of the landscape.
- If you took it: the short version in Governance in practice, in the course Leading AI Products and Teams.
What you need
- The EU AI Act timeline and your rules per system table from the last lesson.
This course is not legal advice. The sources are the official texts: the EU AI Act, Regulation (EU) 2024/1689 (official text, checked 22 October 2026) and the Digital Omnibus on AI, Regulation (EU) 2026/1744 (official text, checked 22 October 2026), which changed some of its dates.
The rules for selling food
Section titled “The rules for selling food”Think of the rules for selling food in most countries. A few things may not be sold as food at all. Baby food has heavy rules: tests, records and inspections before it reaches a shelf. Packaged food needs a label, so the buyer knows what is inside. A plain apple needs nothing new. The rules follow the harm that the food can do, not how clever the recipe is.
Two more ideas come with them. The maker and the shop have different duties: the maker tests the recipe, and the shop stores and sells the food correctly. And a maker abroad that sells into the country follows that country’s rules. New rules often start later than the day they are written, so that makers have time to change.
The EU AI Act works in a similar way.
The problem: two opposite mistakes and an old date
Section titled “The problem: two opposite mistakes and an old date”Farida asks the team what the EU AI Act means for Larkfield. Larkfield is a Canadian company, and it sells to customers in Canada and the European Union (constructed for this course). She gets three answers:
- Tomás: “We are a Canadian company. A European law does not apply to us.”
- Omar: “Our triage service is AI, so it is high-risk. We must stop every launch until we have a lawyer.”
- Grace: “I read an article. The high-risk rules start on 2 August 2026.”
All three are wrong, in different ways. Tomás thinks that the law follows the company’s address. Omar thinks that every AI system is high-risk. Grace has the date from the original text, before a 2026 law changed it. A team that follows any of them either ignores the law or wastes months on rules that may not apply.
The idea: rules in proportion to the harm
Section titled “The idea: rules in proportion to the harm”The EU AI Act is the European Union’s law on artificial intelligence. It uses a risk-based approach: the duties grow with the harm that an AI use can do. The law itself does not list “four levels”. This course groups its rules into four, as a teaching picture, plus a separate set for the large models behind many tools.
Text description of the diagram
The flow asks three questions about an AI use, in order. First: is it a banned practice? If yes, it may not be used in the EU. If no, the second question: is it high-risk? That is, is it part of a product’s safety under the Annex I laws, or in an Annex III area? If yes, heavy duties apply, for example technical documentation, quality management and human oversight. If no, the third question: does it talk to people, or make generative content or deep fakes? If yes, transparency duties apply: tell people, and mark generated output. If no, it is minimal risk, with no new duties from the Act. A separate box is for general-purpose AI models, the large models behind many tools. Their own chapter of rules is for the companies that provide them.
Each level, in plain words:
- Banned practices. Examples are harmful manipulation, exploiting people’s vulnerabilities, social scoring, and emotion recognition at work and in schools, with exceptions (Article 5). The Omnibus added two: AI that makes realistic intimate images of an identifiable person without consent, and AI that makes child sexual abuse material.
- High-risk systems. An annex is a list attached at the end of the law. A system is high-risk in two ways (Article 6). It is a safety part of a product under the EU laws listed in Annex I. Or it is used in one of the eight Annex III areas. They are biometrics, critical infrastructure, education, employment, essential private and public services (for example credit scoring), law enforcement, migration and border control, and justice and democratic processes. An Annex III system that does only a narrow procedural task can fall outside, but one that profiles people is always high-risk (Article 6(3)).
- Transparency duties. People must be told when they talk to an AI system, unless it is obvious. Generative AI output must be marked in a machine-readable way, and deep fakes must be disclosed (Article 50).
- Minimal risk. Everything else. The Act adds no new duties. Other laws, such as privacy law, still apply.
- General-purpose AI models have their own chapter, with duties for the companies that provide them. Models with “systemic risk” have more duties (Article 55). Larkfield uses such models through a provider; it does not provide one.
Two more points matter for almost every organisation in scope. AI literacy first asked providers and deployers to “ensure” that their staff had enough AI literacy, to their best extent (Article 4, as first adopted). The Omnibus rewrote it as a duty to “support the development of” AI literacy (new Article 4, from Article 1(5) of Regulation (EU) 2026/1744). It does not require a guaranteed level for any person. And fines are large. They go up to EUR 35 million or 7% of worldwide annual turnover for banned practices, whichever is higher. For smaller companies, the lower of the two amounts applies (Article 99).
Two roles: provider and deployer
Section titled “Two roles: provider and deployer”The food maker and the shop have different duties. The Act has two main roles in the same way (Article 3(3) and (4)).
- A provider develops an AI system, or has it developed, and offers it or puts it into service under its own name.
- A deployer uses an AI system under its authority, for work. Personal, non-professional use does not count.
One company can hold both roles for one system: it builds the system and uses it. For a high-risk system, a deployer must use it as its instructions say. It must also give human oversight to people with the competence, training and authority to do it (Article 26). Some deployers must also assess the effects on people’s rights first. They are mainly public bodies, and those that score credit or price life and health insurance. Not every deployer must do this (Article 27).
Text description of the diagram
The diagram has two steps. Step one asks whether a company outside the EU is in scope. If it places an AI system on the EU market, or if the system’s output is used in the EU, the Act can apply. If neither, the Act is unlikely to apply, but ask a lawyer to confirm. Step two asks which role the company has, system by system. A provider develops the system, or has it developed, and offers it or puts it into service under its own name. A deployer uses it under its authority, for work. A company that builds a system and uses it is both, for example Larkfield with its triage service.
Outside the EU. The Act can apply to a company outside the EU. It does when the company places an AI system on the EU market, or when the system’s output is used in the EU (Article 2(1)). So Tomás’s answer is wrong: the question is where the output is used, not where the company lives, as for the food maker abroad.
The dates, as changed in 2026
Section titled “The dates, as changed in 2026”The Act entered into force on 1 August 2024 (Article 113). Its rules then start in steps. In 2026, the Digital Omnibus on AI changed some of the dates. It is a separate EU law, Regulation (EU) 2026/1744, in force since 27 July 2026.
Text description of the diagram
The timeline has eight dates in two groups. These already applied on the date the texts were checked. First, 1 August 2024: the AI Act enters into force. Then 2 February 2025: the definitions, AI literacy and the banned practices. Then 2 August 2025: the rules for general-purpose AI models, the governance bodies and the penalties. Then 27 July 2026: the Digital Omnibus on AI enters into force and moves some dates. Then 2 August 2026, the general date: most of the Act applies, including the transparency duties. These apply later. On 2 December 2026, two new banned practices start, and generative systems already on the market before 2 August 2026 must mark their outputs. On 2 December 2027, the rules for high-risk systems in the Annex III areas start; the original date was 2 August 2026. On 2 August 2028, the rules for high-risk systems in Annex I products start; the original date was 2 August 2027.
The two moved dates are in the Omnibus (its Article 1(40)): 2 December 2027 for Annex III systems and 2 August 2028 for Annex I products. The original Act had 2 August 2026 and 2 August 2027. The new bans and the marking date for older generative systems are 2 December 2026. So Grace’s article was right when it was written, and wrong now. Always read the date of a summary, and check it against the official text.
A tiny example: the document assistant
Section titled “A tiny example: the document assistant”Farida fills the empty cell of the document assistant’s row. She follows the flow, one question at a time, and writes each answer:
| Step | Question | Answer for the document assistant |
|---|---|---|
| 1 | Is its output used in the EU? | Not directly: staff in Canada read the answers. The partner shop’s staff: location unknown. |
| 2 | A banned practice? | No: it answers staff questions from documents. |
| 3 | An Annex I product, or an Annex III area? | No: it does not decide on jobs, credit, education or access to services, and it profiles nobody. |
| 4 | Does it talk to people? | Staff ask it questions, but they open it as an AI assistant, so it is obvious. |
| 5 | Category (most likely) | Minimal risk |
| 6 | Role | Larkfield built it (provider) and uses it (deployer): both |
| 7 | Not sure | Does serving the partner shop’s staff mean that Larkfield places it on the market for another company? |
Step 4 is the one that a learner most often skips. The duty to tell people applies unless it is obvious. Here the course reads it as obvious, but this is a judgement, so it goes to the lawyer too.
Larkfield: two more systems, two other answers
Section titled “Larkfield: two more systems, two other answers”The same steps give other answers for other systems. The reply suggester in the helpdesk tool comes from TicketHarbor (made up for this course). TicketHarbor is its provider; Larkfield, which uses it, is its deployer. The triage service reads EU customers’ tickets, so its output is used in the EU. Larkfield built it and uses it, so Larkfield is both provider and deployer. Its category follows the same flow. You place it, and the other six systems, in the module practice at the end of this module.
The surprise comes then. As the course reads the Act, none of Larkfield’s systems is high-risk. The Act still matters through its roles, its transparency duties and its dates. But its biggest effect on Larkfield is the habit that it builds: know each system, its role and its reason.
Discuss: “Then we can wait until 2027”
Section titled “Discuss: “Then we can wait until 2027””Omar reads the timeline and smiles.
- Omar: “The high-risk rules start in December 2027, and we are not high-risk anyway. We can wait.”
- Sam: “Does nothing apply today?”
- Farida: “The bans, AI literacy and the transparency duties already apply. And if we add AI to hiring, that is an Annex III area.”
Before you read on, answer these:
- Which parts of the Act already applied on the date the texts were checked?
- Which Larkfield change would bring a system into the high-risk rules?
- Is “not high-risk” a reason to do nothing?
The course’s answer. On the date checked, the banned practices, AI literacy, the rules for general-purpose AI models, the penalties and the transparency duties already applied. A use of AI in hiring, for example screening applicants, would most likely be in the employment area of Annex III. So Larkfield would need to check it before the December 2027 date, not after. “Not high-risk” is not “nothing to do”. Larkfield still needs to know which systems talk to EU customers, and, where the Act applies, it has a literacy duty. The work that high-risk rules ask for, such as an inventory and documentation, takes months. Most of it is good governance anyway.
What would change the answer: a new use in an Annex III area, or a system that talks to customers alone. A change in the official text would too. Then the lawyer looks again, with the new facts.
When to use this, and what it costs
Section titled “When to use this, and what it costs”Use the flow for every system that has a link to the EU: its output, its users or its market. Do it again when a system changes what it does or whom it serves. Record the category as “most likely”, with the reason, and keep a “not sure” line. The lawyer gives the final answer.
The cost is small for minimal-risk systems: a row in a table. It is large for high-risk systems: documentation, oversight, records and checks, often for months. That is why the category question comes first. Do not pay for high-risk work on a system that is not high-risk, and do not miss the one that is.
Common mistake
A plan built on an old summary
Symptom: the team’s plan says “high-risk rules from 2 August 2026”. Nobody can say where the date comes from.
Cause: the team used a summary written before the Digital Omnibus on AI changed the dates. Secondary summaries do not update themselves.
Fix: take every date from the official text, and write when you checked it. Keep one timeline for the whole company, owned by one person, as the EU AI Act timeline in the course files. Check it again before each decision that depends on a date.
Knowledge check
Section titled “Knowledge check”Check your understanding
Tomás says: 'We are a Canadian company, so the EU AI Act does not apply to us.' What is the right question to ask instead, and why?
Show the answer
Ask, system by system: does Larkfield place the system on the EU market, or is the system’s output used in the EU? The Act can apply to a company outside the EU in those cases. So the answer depends on where each system’s output is used, not on where the company is. A lawyer confirms the answer for each system.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
Larkfield uses TicketHarbor's reply suggester (made up for this course). Larkfield also built the triage service and uses it. Name Larkfield's role for each one, and the reason.
Show the answer
Reply suggester: Larkfield is the deployer, because it uses the system under its authority for work. TicketHarbor is the provider, because it develops the system and offers it under its own name. Triage service: Larkfield is both provider and deployer, because it developed the system, put it into service under its own name, and uses it.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
Predict, then check against the timeline: a company plans to use AI to screen job applicants in the EU. Which category is it most likely in, and from which date do those rules apply, as amended?
Show the answer
Most likely high-risk: employment is one of the Annex III areas. As amended by the Digital Omnibus on AI, the high-risk rules for Annex III systems apply from 2 December 2027. The original date was 2 August 2026. Because the work takes months, the company should start before that date, and ask a lawyer about its own case.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Recap
- The EU AI Act sets duties in proportion to the harm. The course’s picture: banned, high-risk (Annex I products or the eight Annex III areas), transparency duties, minimal risk, and a separate chapter for general-purpose AI models.
- A provider develops a system and puts it out under its own name; a deployer uses it at work. One company can be both. The Act can apply outside the EU when the output is used in the EU.
- The Digital Omnibus on AI moved the high-risk dates: 2 December 2027 for Annex III, 2 August 2028 for Annex I. Take dates from the official text, never from an old summary.
- As the course reads it, none of Larkfield’s systems is high-risk. That is a reason to record the reason, not to do nothing.
You can now place a system in the Act and name its role. But the Act says what must be true, not how to run governance every day. Next, you see the framework and the standards that organisations use for that.