Try each exercise first, then open its model answer.
Designing a governance regime · Workbook
As of 18 November 2026. Big Widget Company and its hiring assistant are made up. Every fact in the model answers comes from this module's gears and the sources they cite (see sources.yml). Where the module has not checked something, the model answers say "not checked". Nothing here is legal advice.
Exercise 1 (500.8-a): Find the gaps, then fill them
Big Widget Company's AI committee has sent you this draft plan for its hiring assistant:
"Big Widget has adopted the NIST AI Risk Management Framework. The business decides when to start using the hiring assistant. Applicants below the assistant's confidence threshold are rejected automatically. The assistant keeps its own logs. Complaints go to hr-info@ like any other query. Stopping the assistant needs sign-off from the AI committee."
- Using this course's four parts of a regime (decider, gate, evidence, remedy), say whether the draft has each one and, if not, how the regime breaks without it.
- Write corrected rows for two actions, "Reject an applicant" and "Stop the assistant", with all four parts filled in.
- In two or three sentences, say why adopting the NIST framework does not by itself answer question 1. Say which parts of your answer come from NIST and which are this course's own.
Model answer
- Decider: missing. "The business decides" names no role, so no one does. And because the assistant rejects applicants by itself, no person makes the rejection call at all. A named decider on paper would not fix that; the test is whether the person actually decides. Gate: broken. The only check before a rejection is the assistant's own confidence score, so the governed system is checking itself. Evidence: weak. Logs exist, but nothing links a rejection to the applicant, system version, rules in force, rank, reviewer and reason, so no one else can reconstruct the decision. Remedy: missing in practice. A general complaints inbox with no power to reverse anything is not a challenge route; harm can be recorded and never put right. The stop rule is also a problem: making a stop wait for a committee sign-off defeats the point of a stop.
Good rows:
Consequential action Decider Gate (checked before the action) Evidence kept Challenge and remedy Reject an applicant An HR reviewer named in the go-live record. The assistant has no path to send a rejection A person reviewed the application itself, not only the rank, and recorded a reason Decision record linking applicant reference, where the application is kept, model version, rules-pack version, rank, reviewer, reason and time Rejection notice says how to challenge, how to ask for an explanation, and by when; a reviewer who did not make the decision re-examines the record and the rules in force, and can reverse Stop the assistant HR governance lead; any named overseer may pause None, by design: stopping is always allowed, and the regime says so in writing Stop record and reason Restart only through the "start ranking" gate again The empty gate on the stop row is not a gap. It is a written decision that a stop never waits.
- A framework document decides nothing; people do. NIST's framework is voluntary guidance, not law. Some of its outcomes overlap with the four parts: it asks that roles and responsibilities are documented and clear, for feedback from people outside the team, and for post-deployment plans that include mechanisms for appeal and override. But the four-part test itself is this course's design, not NIST's, and Big Widget still has to fill it in for each consequential action. NIST says organisations may borrow "as many – or as few – suggestions as apply" from its Playbook. That following them is not, by itself, evidence that a legal duty is met is this course's point, not the Playbook's.
Strong answers also note that the remedy depends on the evidence: a challenge reviewer cannot re-examine a rejection that left no record.
Exercise 2 (500.8-b): Test procurement's label, then date and scope the duties
Scenario date 18 November 2026. Big Widget's EU subsidiary uses a purchased hiring assistant, under its own authority, to rank EU job applicants for HR review. Procurement writes: "We are the deployer. The duties start on 2 December 2027. Nothing else to do." The packet does not say who developed or commissioned the system; who placed it on the market or put it into service, under whose name, or when; whether Big Widget has put its own name on it; or whether Big Widget's configuration changed its intended purpose or substantially modified it.
- State your role finding in one sentence, citing the AI Act articles you rely on.
- List the missing role facts as open questions.
- Walk through the classification question in order, saying which steps this module could not check.
- Explain why "the duties start on 2 December 2027" is not the end of the analysis, and add the missing transition questions.
- Fill in this table for five rows: the role clauses, the transition rule, Article 26(6) log keeping, Article 86 explanations, and ISO/IEC 42001.
Clause Applies from (scenario reading) Gate Evidence Owner Open uncertainty
- A colleague says the deployer duties "applied from 2 August 2026". Another says "the 2026 amendment pushed the whole Act back". Correct both.
Model answer
- Deployer, supported on the stated facts; additional provider capacity unresolved. Using the assistant under its own authority fits the Article 3(4) deployer definition. But Article 25(1) of the original Regulation (EU) 2024/1689 treats a deployer as the provider of a high-risk system already placed on the market or put into service if (a) it puts its name or trademark on it, (b) it substantially modifies it so it remains high-risk, or (c) it modifies the intended purpose of a system not classified as high-risk so it becomes high-risk. Buying the system does not settle that Big Widget is only a deployer.
Open questions for the role memo:
- Who developed the system, and did Big Widget have it developed?
- Who placed it on the market or put it into service, under whose name or trademark, and when?
- Did Big Widget put its own name or trademark on it afterwards? If Big Widget's name appears, who put it there, and what does the contract say about who carries the provider's obligations? A logo on a screen is a fact to explain, not an answer.
- Did Big Widget's configuration substantially modify it, or change its intended purpose? Was it classified as high-risk before that change, and is it after?
Under Article 3(3), a provider both develops the system (or has it developed) and places it on the market or puts it into service under its own name. Both halves need facts. Whether the 2026 amendment changed any part of Article 25 is not checked; counsel checks a consolidated text.
Classification, in order:
- Is the intended purpose within Annex III point 4(a): recruitment or selection, in particular to analyse and filter job applications and to evaluate candidates? Use the provider's stated purpose and how Big Widget actually uses the tool.
- Does it profile natural persons? If yes, it is always high-risk and the Article 6(3) derogation is closed. The Act's definition of profiling was not read for this module: not checked.
- If not, does it pose a significant risk of harm to health, safety or fundamental rights, including by materially influencing the outcome of decision making? A ranking that sets which applications HR reads first, or at all, may shape the outcome. That is a fact to examine, not to assume.
- The rest of Article 6(3), Article 6(4), and any change the amendment made to Article 6 were not read: not checked. Counsel owns this step.
Record the classification as a documented question with the facts relied on, not as a conclusion.
A start date says when a duty could first apply, not that it reaches this tool. Article 111(2), as amended by Regulation (EU) 2026/1744, applies the Regulation to operators of high-risk systems placed on the market or put into service before the Chapter III date "only if, as from that date, those systems are subject to significant changes in their designs". Its separate 2 August 2030 deadline is for providers and deployers of high-risk systems "intended to be used by public authorities". That turns on the system's intended use, not on who the operator is. Nothing in the stated facts says the assistant is intended for use by public authorities; the transition memo records that as a fact to confirm. Two more questions for the memo: 5. When was the first unit of this assistant's type and model placed on the market or put into service in the EU: before or after the Chapter III date? Recital 39 of 2026/1744 says that for this grace period "the decisive factor is the date on which the first unit of that type and model of high-risk AI system was placed on the market or put into service on the Union market for the first time". So Big Widget's own go-live date is not the answer. The memo records the provider's evidence for that first date and whether Big Widget's unit is the same type and model. The same recital extends the grace period where a unit "has been lawfully placed on the market or put into service" before the date. That lawful unit need not be the first one. So the memo keeps two separate facts: the first unit's date, and the provider's evidence that at least one unit of the same type and model was lawfully placed on the market or put into service before the Chapter III date. The dates alone do not show the second. If any of these is unclear, or lawfulness is disputed, counsel rules, including on how the recital applies. 6. On or after that date, and by the day of the check, has any change been made that counts as a significant change in its design? Who decides, and on what evidence? A change that is only planned does not count yet; it is a reason to run the check again on the day it is made.
Legal counsel rules on this in writing before the "start ranking" gate opens. Until then, the gate's answer is "wait for a person".
A good table:
Clause Applies from (scenario reading) Gate Evidence Owner Open uncertainty Art. 3(3), 3(4) definitions; Art. 25(1)(a)-(c) role shift (original 2024 text) Art. 3 is in Chapter I: 2 Feb 2025 under original Art. 113. Art. 25 is in Chapter III Section 3: 2 Dec 2027 for Annex III systems under Art. 113 as amended by Regulation (EU) 2026/1744 Art. 1(40)(b) Before configuration and before ranking use Role memo: stated facts; the four missing role facts; contract, market and branding evidence HR governance lead, with legal counsel Provider capacity unresolved until development, placing-on-market, branding and modification facts are known Art. 111(2) as amended: systems already placed on the market or put into service Turns on the Chapter III date and on design changes on or after it Before ranking use, and again at every design change Transition memo: first EU market or service date for this type and model, and, as a separate fact, the provider's evidence that at least one unit was lawfully placed on the market or put into service before the Chapter III date, design changes made on or after the Chapter III date Legal counsel, with the HR governance lead Whether Chapter III duties reach this system at all; interaction with Art. 86 not checked Art. 26(6): keep logs automatically generated by the system, to the extent under Big Widget's control, for a period appropriate to the intended purpose of at least six months, unless applicable Union or national law (in particular data-protection law) provides otherwise 2 Dec 2027 at the earliest Before ranking use Log retention setting, owner, and the data-protection review of it HR governance lead with IT and the privacy owner Covers only automatic logs under Big Widget's control; depends on the logging capability the provider supplies Art. 86(1)-(3): explanation of the AI system's role and the main elements of the decision, for an affected person subject to a deployer's decision based on an Annex III high-risk system's output with legal or similarly significant effects they consider adverse to their health, safety or fundamental rights; limited by paras. 2 and 3 Chapter IX: the general date in original Art. 113 is 2 Aug 2026 Before rejection Decision record able to produce the explanation HR governance lead Not every rejection is in scope. Whether the right reaches this tool before the Chapter III duties apply, its interaction with Art. 111(2), and any amendment to Art. 86 are not checked; counsel rules ISO/IEC 42001:2023 Not a legal date None claimed None claimed Not assigned Only the ISO catalogue page was read; no clause is checked and no certification is claimed Keep Big Widget's own promise to explain every rejection on request in a separate row marked "company policy, not law". It is wider than the Article 86 right.
- 2 August 2026 is the general date in the original Article 113. Regulation (EU) 2026/1744, adopted on 8 July 2026, amended Article 113: its Article 1(40)(b) makes Chapter III Sections 1 to 3, except Article 6(5), apply from 2 December 2027 for Article 6(2)/Annex III systems and 2 August 2028 for Article 6(1)/Annex I systems. So on 18 November 2026 the deployer duties in Article 26 are not yet applicable, and even then the transition rule may keep them from reaching this tool. But this is not a blanket deferral: the amendment moves specific provisions, and Article 3, for example, has applied since 2 February 2025. A future date means "not yet", not "never". The European Commission's AI Act page is an overview, not the operative legal text, so it is not the source for a date. Only Article 1(40), Article 4, the amended Article 111(2) and recital 39 were opened; other changes the amendment makes are not checked.
Exercise 3 (500.8-c): From a report's recommendation or finding to a design requirement
- Name the two public-sector failures this module uses, the body that reported on each, and the date of each report.
- For each of these Big Widget design features, name the recommendation or finding it answers and its source, say which of the two it is, and say whether the feature fully meets that need or only partly: (a) a rejection notice that says how to request human review and by when, with a reviewer who did not make the decision and can reverse it; (b) an applicant notice at go-live saying, in plain words, that an AI system ranks applications; (c) a versioned change log of the tool's settings; (d) a decision record linking each outcome to its input, system version, rules-pack version and decider.
- For (c), say what Big Widget should add so an independent expert can actually scrutinise the rules.
- Explain in a short paragraph why an audit after deployment is not enough, and state two limits on what these cases can prove.
Model answer
- Robodebt (Australia): the Royal Commission into the Robodebt Scheme presented its report on 7 July 2023, and it was tabled that day; an updated edition followed on 11 July 2023 after corrections. Ongekend onrecht (the Netherlands): the Dutch parliamentary inquiry committee into the childcare benefit reported on 17 December 2020. It found that fundamental principles of the rule of law were violated in administering the childcare benefit.
(a) Recommended: a clear review path for people affected by automated decisions (Robodebt Recommendation 17.1). It meets that need if it exists before the first decision and the reviewer really can reverse. (b) Recommended: plain-language information about automated decisions (Recommendation 17.1). Partly: it says the tool is used, not how it decides, so the regime adds an explanation on request. (c) Recommended: access to business rules and algorithms so that independent experts can scrutinise them (Recommendation 17.1). Partly: a log of setting changes is useful but may not show the rules the tool actually applies. (d) Found: late, incomplete and incorrect information to Parliament and deficient information management within ministries (Ongekend onrecht, p. 8). The record also gives something to audit from the first decision to the body that Recommendation 17.2 says the Commonwealth should consider, one that monitors and audits automated decisions.
The Robodebt rows are recommendations: what should exist from now on. On their own they are not findings that each thing was missing in the Robodebt scheme. This module read only the report page, so the report's chapters and whether the recommendations were carried out are not checked.
- A rules pack: the configuration, thresholds, ranking criteria and provider documentation that actually drive a ranking, kept as a numbered version. A named reviewer from outside the HR team can inspect it under written access terms, and each decision record names the rules-pack version and where the application is kept, so the input can be recovered within the retention period. If the provider will not let anyone outside see how the ranking works, record that as a gap with an owner. Do not call a change log "access to the rules".
- Missing records and missing remedy powers limit what a later audit can reliably rebuild or put right. Other records or testimony may fill some gaps, and a remedy created later may help some people, but a regime cannot promise either to the people it has already decided about. So the challenge route, the notice, the rules pack and the records have to exist before the first decision. This is why Raji and colleagues place audit documents at each stage of development rather than at the end, and why NIST lists appeal and override among the mechanisms a post-deployment plan should include. Limits: neither case involved a generative AI system, and neither source says that any one design choice would have changed what happened. The Dutch report also leaves nationality-based risk profiles outside its remit and refers to the data-protection authority's separate investigation, so that investigation's findings are not attributed to the report.