Try each exercise first, then open its model answer.
The EU AI Act for builders · Workbook
Four exercises, one for each thing this module teaches. Big Widget Company and everything it builds here are made up. Do not use real names, customers or documents from your own work. This is not legal advice: where you are not sure, write "not checked" and say who you would ask.
Answer as of 18 November 2026, using only the dates the lesson gives and the sources it names.
Exercise 1 (300.7-a)
Every duty, every date. Big Widget plans four builds. All four are used by people in the EU, and Big Widget builds and uses each one itself.
- Build A. A camera tool that guesses how warehouse workers feel during their shift, so managers can "spot low morale". It is not for medical or safety reasons.
- Build B. A tool that makes product pictures for dealers to post on their websites.
- Build C. A tool that scores and ranks applicants for internal job openings. It does not chat with anyone.
- Build D. An AI tool that suggests the order in which to pick warehouse shelves. It does not talk to anyone or make content, and its use is not on any list in the Act.
For each build:
- Write the four facts: intended use, where its users are, the as-of date, and every duty that may apply (more than one can).
- For each duty, give the date it applies from, and say which text that date comes from.
- Say whether the duty is already in application on 18 November 2026.
Then answer: a colleague says "Big Widget is not an EU company, so none of this applies." What do you tell them? And name one thing this module has not checked that you would look up before relying on your list.
Model answer
| Build | Duty that may apply | Applies from | Where the date comes from | In application on 18 Nov 2026? |
|---|---|---|---|---|
| A, B, C and D | AI literacy: take measures to support the AI literacy of the staff who run and use each tool (Article 4) | 2 February 2025 | Commission overview | Yes |
| A | Prohibited practice: inferring the emotions of people in the workplace, except for medical or safety reasons (Article 5(1)(f)) | 2 February 2025 | Commission overview | Yes: the practice is prohibited, and this build does not fit the medical or safety exception |
| B | Transparency: generated pictures must be marked in a machine-readable format as AI-made (Article 50(2)) | 2 August 2026 | Commission FAQ; the amended Article 113 names no separate date for Article 50 | Yes |
| C | High-risk: an employment use under Annex III (Article 6(2)) | 2 December 2027 | Regulation (EU) 2026/1744, which replaced Article 113 point (c) | No, not yet |
| D | Nothing more from its risk level: the minimal-risk level adds no rules of its own; AI literacy (first row) still applies | Not applicable | Not applicable | Not applicable |
A good answer also says:
- Every build has at least one duty. AI literacy is set for providers and deployers of AI systems in general, not for one risk level. The duty is to take measures; the Act does not require a guaranteed level of AI literacy for any one person. "Minimal risk" means no extra rules from that level, not no duty, so Build D is not "nothing to do".
- Build A is the one to stop. It is not "high-risk with paperwork"; it is on the prohibited list.
- Build C and the Article 6(3) exit. A tool that scores applicants is likely to profile them, and an Annex III system that profiles people is always high-risk. Do not count on the exit without a legal owner's written answer.
- Build C's date. The 2024 text dated most of the Act from 2 August 2026. That is no longer the date for hiring tools: the 2026 amendment moved it to 2 December 2027. Name the capacities and plan the oversight now anyway.
- The colleague. The Act covers providers placing AI on the EU market whether or not they are established in the EU, and providers and deployers outside the EU where the output is used in the EU. The Commission says it applies to actors "inside and outside the EU". What matters is where the users are and where the output is used.
- Not checked. For example: the transition rules for systems already on the market (Article 111), or whether the Omnibus changed anything else about Article 50. Look it up on the amending regulation and the Commission's current page, or ask the legal or compliance owner.
Exercise 2 (300.7-b)
Who holds which role? For each row, decide whether Big Widget is the provider, the deployer, both, or neither, and name one duty that follows from each role it holds.
- The warehouse team buys a vendor's high-risk screening tool for seasonal hiring and uses it exactly as supplied.
- IT builds an applicant-ranking tool on a general-purpose model bought from a vendor, calls it "Big Widget TalentRank", and uses it on Big Widget's own applicants.
- Procurement takes another company's high-risk tool and sells it to dealers under the Big Widget name.
- For row 2: does the model vendor have any duties? If so, name one, and say what Big Widget should ask the vendor for.
Then explain, in two sentences, why "a company is either a provider or a deployer, never both" is wrong.
Model answer
| Row | Role | One duty that follows |
|---|---|---|
| 1 | Deployer: it uses the system under its authority (Article 3(4)) | Assign oversight to people "who have the necessary competence, training and authority" (Article 26(2)). Keeping the logs for at least six months (Article 26(6)) is also acceptable. |
| 2 | Provider and deployer. Provider: it developed the system and put it into service for its own use under its own name (Articles 3(3) and 3(11)). Deployer: it uses it under its authority (Article 3(4)) | As provider: design it so people can effectively oversee it (Article 14). As deployer: keep its automatic logs for at least six months (Article 26(6)). |
| 3 | Provider, under Article 25(1)(a): it put its name on a high-risk system already on the market | The provider duties for that system now sit with Big Widget. The first company is no longer the provider of that system but must cooperate. |
| 4 | The vendor is a general-purpose AI model provider | It must keep technical documentation and make information available "to providers of AI systems" that build on the model (Article 53(1)). Big Widget should ask for that information, because its own provider duties depend on it. |
Why "never both" is wrong. The definitions describe what an actor does with a system, not a type of company: building a tool and putting it into service for your own use makes you a provider, and using it makes you a deployer. The Commission's example splits the roles across two firms (a CV-screening tool's developer and a bank that uses it), but nothing in the definitions stops one firm holding both.
A good answer also notes that a hiring tool's high-risk duties apply from 2 December 2027, and that the roles should be named now anyway, because they decide who must do what on that day. In every row Big Widget also owes the Article 4 AI literacy measures, as provider, deployer or both, and those already apply. Whether a relabelling, change or new use fits one of the Article 25 routes is for a legal owner to confirm; each route has its own conditions.
Exercise 3 (300.7-c)
Say it is AI. Big Widget's dealer website has two new features.
- Feature 1. A chat assistant called "Wendy", with a photo of a smiling woman. Its first message is: "Hi! I'm Wendy. What can I do for your widget order today?"
- Feature 2. The product-picture tool from Exercise 1, Build B, whose pictures dealers post on their own sites.
- For each feature: does an Article 50 duty apply, who holds it, and what does it ask for?
- Write the notice for Feature 1 in plain words, and say where and when it appears.
- Mark which part of your notice the Act asks for, and which part is this course's good practice.
- A colleague says: "Everyone knows website chats are bots." Reply in one or two sentences.
Model answer
- Feature 1: Article 50(1) applies. Big Widget built the assistant, so it is the provider and the duty is its own. People must be "informed that they are interacting with an AI system", unless that is obvious to someone "reasonably well-informed, observant and circumspect". A human name and a photo work against it being obvious, and nothing on the screen says it is AI, so the notice is missing. Feature 2: Article 50(2) applies to Big Widget as provider of a tool that generates images: the outputs must be "marked in a machine-readable format and detectable as artificially generated or manipulated". A good answer adds that specific exceptions exist, so check the article for the case.
A notice like this, shown as text before or with Wendy's first message:
You're chatting with Big Widget's AI assistant, not a person. It can get things wrong. To reach a person, type "agent".
It must be clear and distinguishable, appear at the latest at the first interaction, and meet accessibility requirements (Article 50(5)). Showing it as text, not only as an image, lets screen readers announce it.
- The first sentence is what the Act asks for. "It can get things wrong" and the way to reach a person are this course's good practice, not a quotation of the law. A good answer may also suggest dropping the human photo or the human name.
- The Act does not assume people know: it requires the notice unless it is obvious to a reasonably well-informed person, and a chat with a human name and photo is built to look like a person.
Exercise 4 (300.7-d)
Oversight with power, and records. Big Widget TalentRank (Exercise 2, row 2) is a high-risk hiring tool, and Big Widget is both its provider and its deployer. The current design:
TalentRank scores each applicant and emails rejections automatically every night. A weekly report of all scores goes to the HR inbox, which anyone in HR can read. There is no way to pause the nightly run except by asking IT to delete the job. Logs are kept for 30 days.
- Score the design on the three properties from Article 14: can a named person understand each output, override it before it takes effect, and stop the tool so it halts in a safe state?
- List the changes needed so it meets each property.
- What does the Act say about logs, for the provider and for the deployer? Does 30 days meet it?
- Name two other deployer duties for this tool that involve telling people something.
- Map your fixed design to the operate gate: which gate element lines up with each of oversight people, automation bias, the stop and the logs? Say where the link is not a full match.
- A colleague says: "Once TalentRank runs behind the operate gate, we comply with the Act." Name three provider duties the gate does not cover, and say who checks what the Act requires for this tool.
- When do these high-risk duties apply, and why build the oversight now?
Model answer
- Understand: no. No named overseer sees each score with its reasons; a report anyone could read is "someone could look". Override: no. Rejections are sent before any person sees them, so nothing can be reversed before it takes effect. Stop: no. Deleting a job through IT is not a stop button or similar procedure that brings the tool to a halt in a safe state.
- Name a trained recruiter with the competence, training and authority to act (Article 26(2)). Show them each recommendation with its reasons before anything is sent. Let them override or reverse it. Give them a pause that stops all sends. Train them to watch for over-relying on the score, which the Act calls automation bias (Article 14(4)(b)).
- As provider, Big Widget must build the tool so it records events (logs) automatically over its lifetime (Article 12(1)). As deployer, it must keep those logs for at least six months, unless other Union or national law says otherwise (Article 26(6)). Thirty days does not meet that.
- As an employer, inform workers' representatives and the affected workers before using the tool at work (Article 26(7)); and inform the applicants that they are subject to the use of a high-risk AI system (Article 26(11)). Using the tool according to its instructions for use and monitoring its operation are also acceptable deployer duties, though they are not about telling people.
5.
| Operate gate (course) | The Act |
|---|---|
| A named holder, with an approval that expires | Oversight assigned to people with competence, training and authority (Article 26(2)) |
| A rubber stamp is not review | Awareness of automation bias (Article 14(4)(b)) |
| A stop that fails closed | A stop that brings the system to a halt in a safe state (Article 14(4)(e)) |
| A receipt for every gate decision, holding no private content | Automatic logs (Article 12(1)) that record the events Article 12(2) names, kept by the deployer (Article 26(6)) |
The last row is a link, not a match. The receipt holds the decision (what was asked, what the gate answered, who approved), never applicants' details or scores. Article 12(2) asks for logs that give traceability: events relevant to spotting risk, to post-market monitoring and to the deployer's monitoring. Those logs are a separate, controlled work record, and whether they give that traceability is a question to check, not something the receipt proves. 6. The colleague is wrong. Running the gate is not the same as complying with the Act. As provider of a high-risk system, Big Widget also needs, among other things: a risk management system (Article 9(1)); a quality management system (Article 16(c)); the required documentation (Article 16(d)); the conformity assessment procedure before the tool is put into service, an EU declaration of conformity and the CE marking (Article 16(f), (g), (h)); and corrective actions when something is wrong (Article 16(j)). Any three is a good answer. The gate is stricter than the law in places (its approvals expire, and the course uses it for every system) and silent in others. A use-specific check of what the Act requires is done with the legal or compliance owner, not by the gate, this course or an AI assistant. 7. For a hiring tool under Annex III, the high-risk duties apply from 2 December 2027, a date set by Regulation (EU) 2026/1744. Build the oversight now: the capacities and the named holder decide who must do what on that day, and the gate is the course's habit for any system. Building it early does not by itself meet the Act; the legal owner's check of the other provider and deployer duties, and of the transition rules this module has not checked (Article 111), still has to happen.