AI Boss: teacher's guide

How to use this guide

The guide holds everything a teacher needs to run the course on AI Boss: the lecture of each module with a timed outline, the teaching plan of every lesson, the discussion questions, the exercises and the answers to the self-tests. The agendas of the three formats are on the course page; the slides are in the deck of the book.

Executive briefing (90 minutes): A briefing on the question a management body must be able to answer: which management functions its systems already perform, at what degree, and who decides each consequence about a person. It ends with the ten questions of the book's conclusion, to be answered with evidence by the HR director and reviewed by the workers' representatives.
One-day seminar (1 day): A working day in which a mixed group of employer and worker representatives takes one of its own systems from function map to delegation profile, decision record, oversight metric, appeal path and worker notice. The morning supplies the frame and the evidence; the afternoon builds the artefacts. Participants leave with drafts they can take into negotiation or consultation.
Full course (7 weeks, one module a week): The full course follows the seven parts of the book, one module a week, with a lecture on each part and one lesson per chapter. Each lesson includes an exercise that produces an artefact of the kind the chapter describes, so that by the final week participants hold the material for the capstone: a governed design for one system that manages people.

Keynote on the whole book · 45 minutes

Who Should Be the Boss? AI May Direct Work, but Not Decide About People

For a general professional audience of managers, HR leaders, workers' representatives and executives in European enterprises and public authorities

Algorithmic management is not a forecast: systems already draft rosters, allocate claims, rank applicants and score performance. This lecture presents the argument of AI Boss. Management divides into five functions that can be delegated one by one, at a degree matched to how easily a mistake can be undone. Direction can be delegated within written bounds; a consequence for a person always needs a named human decider with the means to decide otherwise. The lecture shows how that rule is made real through records, measured oversight, an appeal that corrects the system and a manager whose role is transformed, not abolished.

6 min	Two undramatic scenes A roster drafted by a procured system waits in a unit manager's approval queue; an agent places a claim in a handler's queue. Neither scene is dramatic, until a route leaves no time to travel between two visits. The question of who is the boss is then a question about decision rights, records and accountability.
8 min	Five functions, five degrees, one closed cell M1 Direction, M2 Evaluation, M3 Consequence, M4 Development and M5 Care differ in reversibility, in the person's stake and in governing law. Each can be delegated from Level 0 Advisory to Level 4 Autonomous, with a default ceiling per function. Level 4 Autonomous is never used for M3 Consequence, and care is a duty no system discharges.
8 min	How direction and evaluation become consequence Data collected to direct work becomes a rating, and the rating reaches pay, hours or access to work without anyone deciding it. Courts and authorities accepted a human decision only where there was evidence of what the human did. A score recorded as advisory was decisive in practice where 97 per cent of pay proposals in one unit matched it, at a median of 48 seconds.
7 min	What it does to people, and what the evidence can say People accept algorithmic assignment far more readily than algorithmic evaluation, and what they miss is recognition. Bias enters through proxies and can be found only by measuring outcomes by group. Short-run gains are observable; whether algorithmic management works over years is unmeasured, so the quiet harms must be designed against in advance.
8 min	The accountable design A delegation profile agreed with the representatives before go-live, and a register that shows the sum and tests whether it can be undone. A decision record in the decider's own words, and an oversight metric that measures the review, not the reviewers. An appeal that reaches someone other than the decider and corrects the rule, and a notice the person managed can read.
8 min	The transformed manager and the verdict Agents take direction and the checking of output; development and care are owed only to people and grow. Time saved by systems becomes development and care only if a role charter protects it. The thesis names what would refute it; the author's own verdict for 2036 is a person conducting systems, and it is labelled as his judgement. If you are managed by an AI boss, ask who your decider is.

What the audience takes away

Management is five functions: direction can be delegated within written bounds, while a consequence for a person always keeps a named human decider.
A human in the loop counts only when records show time, authority, independent information and real departures from the system's proposal.
The AI boss is acceptable where it frees managers for development and care, and it must be evaluated on whether it does.

Module 1 · Part I · The Question

Part I states the question of the book: when AI takes over parts of the manager's work, who is the boss and who answers for what happens to the person managed? The module derives the five management functions, places algorithmic management in a century of managing by measurement, and defines the AI boss, the degrees of delegation and the rule that Level 4 Autonomous is never used for M3 Consequence. Participants leave able to draw a function map and to write a first delegation profile.

Lecture · 45 minutes

Who Is the Boss? Five Functions, Five Degrees and One Rule

This lecture carries the argument of Part I. It starts in two ordinary scenes, a roster in an approval queue at Nordvik Region and a claim placed in a handler's queue at Hansa Mutual, and shows why the question of who is the boss is a question about decision rights, records and accountability. It then divides management into five functions, shows that algorithmic management inherits Taylor's ambition and adds three new properties, and ends with the book's instruments: degrees of delegation with a default ceiling per function, the named human decider, and the delegation profile.

7 min	Two undramatic scenes ShiftWise drafts a four-week roster that a unit manager approves; Hansa Mutual's claims-triage agent places a claim in a handler's queue under rules the group head of claims operations owns. The question becomes concrete when a roster leaves no travel time between two visits, or nine complex claims go to a handler on a reduced schedule. Algorithmic management is observable: the chapter cites the Foodinho fine of 2.6 million euros, the Amsterdam Court of Appeal on Uber deactivations and the reported warehouse terminations without input from supervisors.
10 min	What a manager does: five functions Fayol and Drucker described separable functions; Mintzberg and Hill observed work that is informational, interpersonal and fragmented. The book derives M1 Direction, M2 Evaluation, M3 Consequence, M4 Development and M5 Care from the literature and from the employer's legal duties. The functions differ in reversibility, in the person's stake and in governing law, and they overlap: evaluation feeds consequence and direction sets the pace that care must answer for. A function is not a job: line managers, HR, systems, colleagues and the person managed all perform parts of each.
9 min	A century of managing by measurement From the foreman's simple control through scientific management, the moving line, bureaucratic control and lean production to metrics, platforms and agents. At every step M1 Direction and M2 Evaluation were handed to a system first, and M3 Consequence followed whenever the measure was tied to pay or a job. What is new is scale and speed, opacity, and a consequence for a person without a human decision; direction by an agent with its own identity is emerging as a fourth property. Every regime that measured more eventually needed a contest path.
10 min	Degrees of delegation and default ceilings The series' autonomy levels, Level 0 Advisory to Level 4 Autonomous, are applied per management function and action class. Default ceilings: Level 3 Bounded for M1 Direction, Level 1 Assisted for individual ratings in M2 Evaluation and for adverse M3 Consequence decisions, Level 1 Assisted for M4 Development. M5 Care is asymmetric: a system may flag a risk or block a breach of a protective rule, but never relaxes one. The function an action belongs to is set by its effect on the person, not by its label.
9 min	The named human decider and the delegation profile Level 4 Autonomous is never used for M3 Consequence, and every consequence decision has a named human decider with six capabilities. The rule answers both failure modes of accountability: the problem of many hands needs a name, and the moral crumple zone needs a name that carries the means of deciding. The delegation profile records degree per function, deciders, the representatives' agreement and the stop path; DLG-NR-ROS-01 and DLG-HM-CTA-01 show the form.

What the audience takes away

Management can be divided into five functions, and the division decides what may be delegated to a system and at what degree.
Algorithmic management is Taylor's ambition in a new medium; what is new is scale and speed, opacity and a consequence without a human decision.
A consequence for a person always needs a named human decider who has the means to decide otherwise, recorded in a delegation profile.

Chapter 1 · 60 minutes

Introduction: Who Is the Boss?

The lesson states the book's question, its working thesis in four claims and the evidence that would weaken each, and introduces the five management functions, the five degrees of delegation and the usage modes. Participants meet Nordvik Region and Hansa Mutual, read the commission AIDR-0600 as a decision record, and practise the afternoon exercise that locates where AI already takes part in managing one unit.

Learning objectives

State the four claims of the working thesis and the evidence the chapter says would weaken each.
Name the five management functions and the five degrees of delegation exactly as the book does.
Distinguish the degree of delegation from the usage mode, using examples from the chapter.
Draft a miniature function map for one unit, including Mode 0 features nobody decided on.

Key concepts

Working thesis: Four claims under test: AI may direct work but may not, on its own, decide about people; every management decision AI makes or shapes is a record; the manager's role is transformed, not abolished; and the quiet harms deserve attention before the loud ones.
Five management functions: M1 Direction, M2 Evaluation, M3 Consequence, M4 Development and M5 Care: the units by which the book asks what may be delegated, measured and governed one by one.
Degree of delegation: The series' autonomy level applied to a management function, from Level 0 Advisory, where the system suggests and a person decides, to Level 4 Autonomous, which is never used for M3 Consequence.
Usage mode: How the AI is deployed and whose identity it uses, from Mode 0 Ambient AI to Mode 5 Autonomous production agents. In Modes 1 to 3 the AI works under the manager's identity; in Mode 5 agents act in their own name.
Observable, expected, speculation: The three labels every statement about the future carries. Agents that allocate work to people are observable, the spread of AI-directed knowledge work is expected, and every employee becoming a manager of agents is speculation.

Teaching plan

10 min	The teacher reads the two opening scenes, the roster in the approval queue and the claim in Jana Petersen's team, and asks participants who the boss is in each.
15 min	Presentation of the four claims of the thesis with what would weaken each, then the five functions, the five degrees and the difference between degree and usage mode.
10 min	Walk through why the question is European and urgent in the autumn of 2026: the dates in the AI Act, the Digital Omnibus and the Platform Work Directive as the chapter gives them.
20 min	Exercise: groups draft a miniature function map for one unit and report which touches they marked as unknown or not decided.
5 min	Closing discussion on the difference between Nordvik's procured AI boss and Hansa's built one, and what each commission must deliver.

Discussion

The chapter gives, for each of its four claims, the evidence that would weaken it. Which of the four do you expect to be hardest to defend in your own organisation, and why?
Microsoft's phrase 'the rise of the agent boss' is quoted but not adopted. What separates the observable part of that claim from the part the book labels speculation?
Nordvik Region's AI boss is mostly procured and Hansa Mutual's mostly built. How does that difference change the first work each HR function has to do?

Exercise · 20 minutes

Locate one unit on the five functions

Following the chapter's practical guide, choose one unit of 15 to 40 people in your own organisation, or use the chapter's example shift-based unit. List the systems the unit uses and the AI features enabled in each, including features switched on by vendor releases. For each system write which management functions it touches, the usage mode, the degree in practice and who decided it. Where nobody knows, write unknown; where nobody decided, write not decided. Finish with the consequence check: does every touch of M3 Consequence have a named human decider?

Deliverable: A one-page function map for one unit with systems, functions touched, mode, degree in practice, decider, the consequence check and the named owner who receives the findings.

Self-test

According to the chapter's first claim, which decisions always keep a named human decider?
1. Decisions that assign and pace work, such as rosters, routes and the allocation of cases
2. Hiring, pay, promotion, discipline, dismissal, deactivation and access to work
3. Only dismissals, because other decisions about people can be reversed within hours
4. Decisions taken by Mode 0 Ambient AI features that a vendor release switched on
In the miniature function map of the chapter's practical guide, what is written for the meeting-summary feature that names owners of action items?
1. Level 2 Supervised, decided by a negotiation protocol with the union
2. Level 3 Bounded, decided by the HR director under AIDR-0600
3. Level 4 Autonomous, decided by the supplier of the office suite
4. Degree in practice unknown, and decided by: not decided
What does the chapter say distinguishes Nordvik Region's commission (AIDR-0600) from Hansa Mutual's board request (AIDR-0650)?
1. Nordvik's concerns procured systems under one regime and a duty to negotiate; Hansa's, built agents under seven regimes and a duty to co-determine
2. Nordvik's concerns platform workers covered by the Platform Work Directive from December 2026; Hansa's concerns only the claims staff of its entities outside the European Union
3. Nordvik's allows Level 4 Autonomous for M3 Consequence; Hansa's limits every function to Level 0 Advisory
4. Nordvik's is owned by the unions under the co-determination act; Hansa's is owned by the supplier of the claims-triage agent

Chapter 2 · 75 minutes

What a Manager Does: Five Management Functions and Who Performs Them

The lesson derives the five management functions from Fayol, Drucker, Mintzberg, the research on algorithmic management and the employer's legal duties. Participants learn each function's definition and boundary, why the functions differ in reversibility, stake and law, where they overlap, and who performs each today in a hospital ward, a home-care unit and a claims team. The lesson ends with the book's first artefact, the function map.

Learning objectives

Define each of the five management functions and the boundary that separates it from its neighbours.
Explain why the functions differ in reversibility, the person's stake and governing law.
Identify the overlaps by which a system built for one function acts on another.
Draft a function map for one team that records performers, systems, modes, degrees and deciders.

Key concepts

M3 Consequence defined by effect: M3 Consequence begins wherever the person's legal or economic position changes. It is defined by that effect, not by the label of the decision: an allocation rule that systematically gives one group fewer hours is a consequence decision.
Boundary between M2 and M4: Drawn by purpose. Feedback meant to improve the person is M4 Development; a rating used to compare people is M2 Evaluation. Kluger and DeNisi found that more than a third of feedback interventions reduced performance.
Evaluation feeds consequence: A throughput figure or sales score becomes an input to pay or dismissal, sometimes years after it was introduced for planning. The protection is to treat a new use of an existing figure as a change that needs its own decision.
Function map: A record, for one team, of who performs each management function today and which AI systems touch which functions, at which mode and degree, and who decided that degree. It is a file in version control, readable by the people it concerns.
Care as a relationship: Part of M5 Care is performed through attention, such as noticing that a colleague now finishes her round late every day. No dashboard records it, and when a system takes over the routes the opportunity to notice moves with it.

Teaching plan

15 min	The teacher contrasts what managers were said to do (Fayol, Drucker) with what they were observed to do (Mintzberg, Hill) and what the research on algorithmic management adds.
15 min	Presentation of the five functions with definitions and boundaries, then the table of reversibility, stake and governing law; participants place three everyday decisions on the functions.
10 min	Group reading of the table of who performs each function in the hospital ward, the Strandvik home-care unit and the Hamburg claims team.
25 min	Exercise: participants draft a function map for one team, following the chapter's practical guide and the worked example FMAP-NR-STR-01.
10 min	Debrief: each group names one system that reaches a function at a degree nobody decided, and one informal performer the organisation chart does not show.

Discussion

The chapter says the five functions describe what reaches the person managed, whereas Fayol and Drucker describe the organisation's need. What does a split made from the managed person's side reveal that the classical one hides?
M4 Development is governed mostly by collective agreements and practice. Why does the chapter treat that gap as a reason for care and not as freedom to delegate?
At Strandvik the largest finding is a dashboard used at a degree not yet decided. How would such a finding be received in your organisation, and who would own the answer?

Exercise · 25 minutes

A function map for one team

Using FMAP-NR-STR-01 as the model, draft a function map for one team: either the Hamburg claims team as the chapter describes it, or a team in your own organisation. For each of the five functions list the decisions of a normal month, who performs each and who has the last word. List every system whose output is used in those decisions, including Mode 0 features and Mode 2 assistants. For each system record every function its output reaches, its mode, its degree per function and who decided that degree, or not decided. Mark the overlaps that apply and name a human decider for every M3 Consequence entry.

Deliverable: A function map with named performers for all five functions, each system with mode, degree and decider, the overlaps marked, open questions listed and a review date within 12 months.

Self-test

Average visit time across a route is information for planning. According to the chapter, when does the same figure become M2 Evaluation?
1. When it is computed per person and visible to the manager, whether or not anyone calls it evaluation
2. When the supplier of the dashboard classifies the feature as a performance module
3. Only when it is formally entered into an annual appraisal document
4. When it is shown to the worker alone and to nobody else in the unit
Which functions does the chapter say Annex III, point 4(b), of the AI Act reaches into, and which does it not reach?
1. It reaches M4 and M5 but leaves M1, M2 and M3 to national law
2. It reaches all five functions equally
3. It reaches M1, M2 and M3, but not development or care
4. It reaches only M3, through recruitment and termination
In the worked example FMAP-NR-STR-01, why do ShiftWise open-shift offers also reach M3 Consequence?
1. Because the offers are made at Level 4 Autonomous without any playbook
2. Because the screening pilot uses the same module to rank applicants
3. Because the unit manager approves each offer before it is sent
4. Because offered hours determine pay for staff who depend on extra shifts

Chapter 3 · 75 minutes

From Scientific Management to Algorithmic Management: A Century of Managing by Measurement

The lesson follows a century of managing by measurement, from the foreman and Taylor's task through Ford's line, bureaucratic control, the counter-movements and the stop cord to call centres, platforms and agents. For each regime participants ask which management function was handed to a system and how a worker could contest it. The lesson isolates what is new in algorithmic management and practises tracing the lineage of a measure before AI computes it per person.

Learning objectives

Map each control regime of the chapter onto the management functions the system held.
Explain the three properties the chapter says algorithmic management adds, and the fourth that is emerging.
Explain why measures built for direction and evaluation drift into M3 Consequence.
Draft a lineage review of one measure, with a variance rule that reviews the standard before the person.

Key concepts

Control regimes: Edwards's simple, technical and bureaucratic control, to which the chapter adds scientific management, lean production, metric control, algorithmic control and agent direction, each with what was measured and how a worker could contest it.
Informating: Zuboff's distinction: information technology does not only automate work, it also produces a record that makes the work visible. The record can feed M4 Development or M2 Evaluation, and only a design decision determines which.
Three new properties: Direction and evaluation at the scale and speed of software, opacity to the people managed, and the possibility of a consequence for a person without a human decision. The chapter presents this as the book's interpretation, tested in Part II.
Campbell's law: The more a quantitative social indicator is used for social decision-making, the more subject it is to corruption pressures and the more apt to distort the processes it is meant to monitor.
Variance rule: In AIDR-0624, a route above planned time for four consecutive weeks triggers a review of the planned times first, not of the person, directing attention to the standard before the individual.

Teaching plan

20 min	The teacher tells the history in four steps: the foreman and Taylor's task, Ford's line and the five-dollar day, bureaucratic control and deskilling, and the counter-movements including Nordic law and the stop cord.
10 min	Participants complete the chapter's regime table from memory: what was measured, which functions the system held and how a worker could contest.
10 min	Presentation of what is new: scale and speed, opacity, consequence without a human decision, and direction by an agent with its own identity.
25 min	Exercise: groups trace the lineage of one measure and write the decision record, using AIDR-0624 as the model.
10 min	Debrief on the five lessons of history and on the questions the chapter says history cannot answer.

Discussion

The chapter says the five-dollar day shows that once pay depends on a measure, evaluation spreads to whatever the employer believes predicts it. Where do you see that pattern in present-day systems?
Bureaucratic control gave workers a rule to appeal to. What is lost when the criterion is a ranking that neither the worker nor the manager can read?
The stop cord treated a deviation as a signal about the system and not a verdict on a person. What would a comparable right to halt a direction look like for people managed by an AI boss?

Exercise · 25 minutes

Lineage review of a measure

Take one measure that a system is about to compute per person: the time-per-visit figure at Nordvik Region, throughput per handler at Hansa Mutual, or a measure in your own organisation. Following the chapter's practical guide, write its lineage: when it was introduced, for what purpose and at what unit of aggregation. Place the old and the new use on the five functions and flag any move into M2 Evaluation of a person or into M3 Consequence. Record how the measure could be gamed and what unmeasured work it could squeeze. Decide the degree per function, a variance rule that reviews the standard before the person, a contest path and a review date.

Deliverable: A one-page decision record in the form of AIDR-0624: measure lineage, permitted functions and degrees, variance rule, worker access, contest path, reopen condition and review date.

Self-test

How does the chapter map scientific management onto the five management functions?
1. It kept all five functions with the foreman, who directed, watched, hired and dismissed in person, and added written grievance procedures that workers could appeal to
2. It handed M4 Development and M5 Care to a Social Department of investigators and left the pace of work to the workers themselves
3. It moved M1 Direction to a planning department, made M2 Evaluation a comparison with standard time, and tied both to M3 Consequence through differential pay
4. It built M1 Direction into the moving line and removed evaluation altogether
Which of the following is one of the three properties the chapter says no earlier regime combined?
1. The use of bonuses tied to output
2. Opacity to the people managed
3. The separation of conception from execution
4. Written procedures for grievance and discipline
What does AIDR-0624 decide about Nordvik Region's time-per-visit measure?
1. It stays at Level 0 Advisory for M2 Evaluation and is not an input to pay, discipline or dismissal
2. It becomes an input to variable pay once the union has been informed
3. It is removed from the dashboards and no longer shown to anyone
4. It is delegated at Level 3 Bounded so that the system can reassign routes

Chapter 4 · 90 minutes

The AI Boss Defined: Management Functions, Degrees of Delegation and Accountability

The lesson turns the description of Chapters 2 and 3 into instruments. Participants learn the book's definition of the AI boss, the degrees of delegation per function and action class, the four criteria that set a degree, the default ceilings, and the rule that Level 4 Autonomous is never used for M3 Consequence. They examine the accountability problem in its two forms and write a first delegation profile on the model of DLG-NR-ROS-01 and DLG-HM-CTA-01.

Learning objectives

Apply the chapter's definition to decide whether a given system is an AI boss.
Assign action classes to functions by their effect and compare their degree with the default ceilings.
Explain the six capabilities of a named human decider and the two accountability failures they answer.
Draft a delegation profile with deciders, playbook, excluded inputs, stop path and review.

Key concepts

AI boss: Any AI system or agent that performs part of a management function for or over a person. It acts for a person when it assists a manager who remains the one acting, and over a person when its output reaches the worker directly or takes effect without a manager acting.
Default ceiling: The highest degree an organisation should allow for a function unless a recorded decision justifies otherwise, set by reversibility for the person, breadth, the person's stake, and the law and agreements.
Named human decider: Not a signature. The decider can be identified, can decide otherwise with real authority, sees the information, has the time, is protected when deciding against the recommendation, and answers for the decision.
Moral crumple zone: Elish's term for responsibility misattributed to a human actor who had limited control over an automated system. In management it is the line manager who confirms a dismissal a score recommended without the means to decide otherwise.
Delegation profile: States for one system or team the degree per function and action class, the decider, the representatives' agreement, the stop path and the review. It also records the functions the system does not perform and the inputs it may not use.

Teaching plan

15 min	The teacher presents the two definitions and tests them on cases: Hansa's claims assistant, an office assistant that proposes task owners, ShiftWise and the claims-triage agent.
20 min	Presentation of the degrees, the four criteria and the default ceilings; participants explain why M2 Evaluation has a lower ceiling than M1 Direction and why M5 Care is asymmetric.
15 min	Discussion of the rule for consequences, the six capabilities of the decider, and the two failure modes: many hands and the moral crumple zone.
30 min	Exercise: groups write a first delegation profile for one system and check it against the definition of done in the chapter's practical guide.
10 min	Debrief: groups read out one Level 3 class and its playbook, and the others test whether it is enforced outside the model.

Discussion

The chapter says Hansa's claims assistant becomes part of an AI boss on the day its usage logs become a figure per handler. What does it mean that the boundary is crossed by a decision about data and not by a change in the model?
Reversibility is judged from the position of the person affected: the evening a nurse arranged childcare for cannot be given back. How does that change the way you would classify routine scheduling actions?
Open question Q2 asks who is accountable when a manager accepted a recommended consequence without reading it. What do you make of the chapter's answer that accountability can move to whoever set the volume?

Exercise · 30 minutes

A first delegation profile

Choose one system that already takes part in managing people: ShiftWise at Nordvik Region, the claims-triage agent at Hansa Mutual, or a system in your own organisation. Following the chapter's practical guide, list its action classes, at most fifteen, and assign each to a function by its effect on the person. Record the degree as observed, not as intended. Compare each class with the default ceilings and lower or justify any that exceed them. Name the decider for each class and the stop path, write the playbook and the excluded inputs for each Level 3 Bounded class, and state which functions are not delegated. Add the representatives' procedure and the review triggers.

Deliverable: A delegation profile in the form of DLG-NR-ROS-01: action classes per function with degree and decider, playbook and excluded inputs, functions not delegated, stop path, representatives' position and review date.

Self-test

Why do Nordvik Region's open-shift offers sit at Level 3 Bounded in DLG-NR-ROS-01 without exceeding the ceilings of M2 Evaluation and M3 Consequence?
1. Because the unit manager approves each individual offer before it reaches the staff who opted in, which keeps the action at Level 2 Supervised
2. Because the supplier has certified that the system is not high-risk
3. Because open shifts are classified as M5 Care, which allows protective blocks
4. Because acceptance history and sick-leave history are excluded as inputs, so the action does not rank people
What is the default ceiling the chapter sets for an adverse M3 Consequence decision, such as rejecting an applicant or dismissal?
1. Level 0 Advisory
2. Level 1 Assisted
3. Level 2 Supervised
4. Level 3 Bounded
According to the chapter, what is Level 3 Bounded without a written playbook enforced outside the model?
1. Level 4 under another name
2. Level 2 Supervised with a faster approval
3. A Mode 0 Ambient AI feature
4. An acceptable default for M1 Direction

Module 2 · Part II · The AI Boss Today

Part II shows where the AI boss already works: on dispatch platforms, in warehouses and paced work, in hiring, in performance scores and dismissal, in rosters and allocation, and in office features nobody decided on. Each lesson separates what a source reported from what it has been taken to mean and names the management functions involved. Participants leave able to benchmark an internal dispatcher, assess a rate system as a work-environment risk, keep a hiring decision with a recruiter, test whether a score is decisive in practice and build a feature register.

Lecture · 50 minutes

The AI Boss Today: How Direction and Evaluation Become Consequence

Part II is a field trip through six workplaces, and this lecture carries the one argument they share. In each, a system that was introduced to direct or measure work accumulates data, the data becomes a rating, and the rating reaches pay, hours or access to work without anyone deciding that it should. Courts and authorities accepted a human decision only where there was evidence of what the human did. The lecture follows that drift from the platform to the office, and shows the instruments the chapters build against it: written input lists, a named decider, a practice test and a feature register.

9 min	The platform loop: allocation, pay, rating, deactivation Dispatch platforms perform M1 Direction, M2 Evaluation and M3 Consequence in one loop, with blended scores the rated person cannot read. The Amsterdam Court of Appeal accepted human involvement for the one driver Uber had interviewed, and not for three whose files held only notes. Three steps carry the pattern into ordinary employers: behaviour is recorded, the record becomes an input, and the hidden rating reaches hours and pay.
8 min	When the system sets the pace A rate system has a target, a measurement, a comparison and a follow-up; the governance questions concern a parameter and a link, not a model. The Framework Directive requires the employer to alleviate work at a predetermined work-rate, so a target generated by software is assessed, adapted and consulted on. M1 Direction becomes M5 Care when the system sets the pace, and the duty of care is exercised over parameters.
8 min	Access to work: hiring by machine Where recruiters read only the top of a ranking, the ranking has settled the fate of everyone below it: delegation is set by use, not by the configuration screen. Human recruiters discriminate too, and audits measure groups, not the individual rejected. A screening configuration is in substance a selection guideline; Nordvik Region paused a pilot it could not explain and recorded why in AIDR-0601.
9 min	How a score becomes a decision In the reported cases the score selected the person and the human role, where there was one, ratified the selection. A score reaches M3 Consequence through three administrative channels: the pay input, the warning and the list. Whether a score is advisory is a test of practice: Hansa Mutual found 97 per cent of pay proposals in one unit matching the score band, with a median of 48 seconds.
8 min	Rosters, queues and what they do to pay Every weight in an optimiser's objective function is a management choice, recorded or not. Small reversible allocations accumulate into different hours, premiums, learning and working lives. The input list decides the classification, so it belongs in the delegation profile, with change classes that reopen negotiation.
8 min	The office: features nobody decided on A meeting assistant's owner field is the point at which a summary becomes an instruction. Four tests show when help becomes a boss: the addressee, the default, the trace and the decision. The remedy is not to forbid the feature but to decide it, in a feature register agreed through the representatives' process.

What the audience takes away

Data collected to direct work becomes a rating, and the rating becomes a consequence, unless the drift is closed in advance in a written profile.
A human in the process counts only where there is evidence of what the human weighed and that the outcome could have changed.
Degrees of delegation are measured in practice, by reading rates, time and departures, not read from the configuration screen.

Chapter 5 · 75 minutes

Platform Work: The Algorithm as Dispatcher

The lesson reads a decade of court and regulator decisions on dispatch platforms for what they teach every employer. Participants follow the platform loop of allocation, pay, rating and deactivation across M1 Direction, M2 Evaluation and M3 Consequence, study the Amsterdam, Bologna and Garante decisions and Chapter III of the Platform Work Directive, and benchmark an internal shift-offer system against it, as Nordvik Region did for ShiftWise in AIDR-0625.

Learning objectives

Map the platform loop of allocation, pay, rating and deactivation onto the management functions.
Explain what distinguished the accepted human decision from the rejected ones in the Amsterdam judgment.
Describe the three steps by which an internal dispatcher comes to behave like a platform.
Benchmark an internal shift-offer or allocation system against Articles 7 to 12 of the Platform Work Directive.

Key concepts

Algorithmic wage discrimination: Dubal's term for using granular data on location, individual behaviour, demand and supply to produce variable, personalised and unpredictable hourly pay for broadly similar work, including through the allocation of work itself.
Blended score: Ratings and behavioural metrics combined into one number, as in Foodinho's excellence score. The rider cannot tell which input lowered it, and an evaluation whose inputs cannot be seen cannot be corrected when it is wrong.
Hidden rating: Acceptance history that becomes an input when a feature offers shifts first to people who usually accept. It is M2 Evaluation that no one designed, and fewer offers then reach M3 Consequence through hours and pay.
Human decision on deactivation: Article 10(5) of the Platform Work Directive requires that any decision to restrict, suspend or terminate an account, or of equivalent detriment, be taken by a human being.

Teaching plan

15 min	The teacher draws the loop of Figure 5.1 and places on it the UK Supreme Court's description of Uber's controls and the weights of Foodinho's excellence score.
15 min	Case reading in pairs: the Amsterdam judgment, the Bologna order and the Garante's fines. Each pair states what the system did and where the human stood.
10 min	Presentation of Chapter III of the Platform Work Directive article by article, and of the three steps by which an internal open-shift module follows the platform pattern.
25 min	Exercise: groups benchmark an internal dispatcher against the Directive, using PWD-BENCH-NR-SW-01 as the model.
10 min	Debrief comparing Nordvik's contractual remedy for a supplier default with Hansa's negotiated parameter and daily cap for a built agent.

Discussion

The organisation chart showed humans in the loop for all four Amsterdam drivers. What evidence would your organisation be able to produce today of what a reviewer actually considered?
Dubal and Cook and colleagues reach different conclusions about platform pay, yet the chapter says transparency of parameters is a remedy both sides can support. Why?
At Hansa Mutual no rule weights throughput, yet fast handlers receive more claims. How should an organisation treat an effect that no rule states but the design produces?

Exercise · 25 minutes

A Platform Work Directive benchmark of an internal dispatcher

Take an internal system that offers or allocates work: ShiftWise open-shift offers, Hansa Mutual's claims-triage agent, or one in your own organisation. Although the Directive does not apply, use its Chapter III as a benchmark. Map the loop and record the degree per function. Inventory the behavioural data kept (acceptances, declines, response times), with purpose, viewers and retention. Draft the written question to the supplier on whether any feature uses past behaviour to order or filter offers. Write the main parameters in plain language, name the human decider for any outcome that restricts access to work or pay, and set a review path with a written answer within two weeks.

Deliverable: A benchmark table by article (7, 9, 10(5), 11, 10(1) and 12) with result, detail and action for each finding, plus the residual risk and the evaluation date.

Self-test

Why did the Amsterdam Court of Appeal accept real human intervention for one of the four deactivated Uber drivers?
1. Because two members of the risk team had signed the decision
2. Because the driver's rating was above the required level
3. Because Uber had interviewed that driver before the deactivation
4. Because the fraud signal came from rule-based software and not from AI
What did the supplier's default for ShiftWise open-shift offers store, according to Nordvik Region's benchmark, and what did AIDR-0625 decide?
1. Offers, responses and response times per person for 24 months; retention was cut to 90 days for dispute handling, with no reporting per person
2. Customer ratings of each care assistant for five years; the ratings were moved into the home-care dashboard so that unit managers could see them beside planned visit times
3. Nothing about individuals; the Region asked the supplier to begin recording acceptance rates
4. Sick-leave history for 12 months; the Region allowed its use to order offers after negotiation
The Tribunal of Bologna declared Deliveroo Italia's booking system potentially discriminatory. What was the mechanism?
1. The excellence score weighted work in high-demand hours at 35 per cent, so that riders above a certain age were paid a lower rate for each kilometre they delivered
2. The platform deactivated riders automatically after three customer complaints
3. The system used biometric identification to verify riders against a database
4. A late cancellation lowered a rider's standing whatever the reason, disadvantaging riders absent for a strike, illness or the care of a child

Chapter 6 · 75 minutes

The Warehouse and the Shop Floor: Pace, Rate and Monitoring

The lesson describes how rate and time-off-task systems direct and evaluate paced work, what investigators, the CNIL, the Conseil d'Etat and the Californian legislator found, and how the findings meet the employer's duty to alleviate work at a predetermined work-rate. Participants learn why M1 Direction becomes M5 Care when a system sets the pace, and assess visit-time targets as a work-environment risk on the model of WRA-NR-HC-01.

Learning objectives

Describe the four parts of a rate system and the point at which M2 Evaluation becomes M3 Consequence.
Separate what each source in the chapter reported about pace and injury from what is disputed.
Explain why a software-generated target is a predetermined work-rate under the Framework Directive.
Assess a pace or rate system as a work-environment risk together with a safety representative.

Key concepts

Rate system: Four parts: a target stating units per period; a measurement that is a by-product of directing the work; a comparison giving a rate and time off task per person; and a follow-up that attaches a message, warning or termination to the figure.
Time off task: A measure of the minutes in which the device recorded no work. Under the chapter's design rules a time-off-task rule never governs breaks.
Predetermined work-rate: The Framework Directive, Article 6(2)(d), requires the employer to adapt work to the individual, in particular to alleviate monotonous work and work at a predetermined work-rate. A target generated by software is such a rate.
Follow-up rule: The written rule for what follows a missed target. By default it opens a conversation about the work, never an automated warning; the conversation after the flag is part of the system.
Team-level strain indicators: Indicators such as missed breaks, late finishes and contested routes, agreed with representatives before any figure is shown and reported monthly with a minimum group size of ten by default.

Teaching plan

10 min	The teacher draws the pace loop of Figure 6.1 and the three properties that distinguish a rate system from the stopwatch of scientific management.
15 min	Source work: groups take the 2019 filing, the 2021 injury analysis, the 2024 Senate report with the company's reply, and the French decision, and report what each says and what is contested.
15 min	Presentation of the employer's duty and the seven design rules; discussion of the home-care difference, that a visit running long may be the work done well.
25 min	Exercise: groups write a risk assessment of the pace parameters of one system, following the practical guide and WRA-NR-HC-01.
10 min	Debrief: each group reads its follow-up rule and stop path aloud; the room checks whether the controls sit in the parameters or ask workers to cope.

Discussion

The Conseil d'Etat treated the same timestamps differently depending on purpose and what follows from them. How would you document the purpose of an indicator so that it cannot quietly become a speed target?
The chapter says a supervisor who set the pace by hand also saw the person who could not keep it. Who sees that person in a workplace you know where a parameter sets the pace?
California's law treats a quota as a written document that can be requested and contested. What would it change if every parameter that sets pace in your organisation had that form?

Exercise · 25 minutes

Risk assessment of a pace system

Working as a line manager and a safety representative, assess one system that sets, measures or enforces pace: the planned visit times at the Strandvik home-care unit, a claims queue at Hansa Mutual, or a system of your own. Name each parameter that changes pace (target formula, planned durations, transition allowances, thresholds) and give each an owner. For each, assess the physical and psychosocial hazards and those arising from the follow-up, and set a control in the parameter itself. Write the follow-up rule, choose team-level strain indicators with a minimum group size, name who can suspend the figures, and list the triggers for reassessment.

Deliverable: A risk assessment in the form of WRA-NR-HC-01: pace parameters with owner and control, written follow-up rule, strain indicators and reporting level, stop path and reassessment triggers.

Self-test

In WRA-NR-HC-01, what does control P2 do with the overrun shown per visit on the dashboard?
1. It sends the overrun to HR as an input to the annual pay review
2. It shows overruns per route and week, with the individual view only on request and logged
3. It converts three overruns in a month into an automated written warning
4. It deletes all visit registrations from staff phones after each shift
What did the Conseil d'Etat uphold when it reduced the CNIL's fine against Amazon France Logistique to EUR 15 million?
1. That the fast-scan indicator imposed speed and lacked any lawful basis
2. That the idle-time indicator unlawfully governed workers' breaks
3. That scanners may never record the time between two scans
4. That keeping every indicator for every employee for 31 days breached data minimisation
What does the chapter mean when it says M1 Direction becomes M5 Care when the system sets the pace?
1. When a parameter sets the pace, nobody sees the person unless the system is designed to show them, so the duty of care is exercised over the parameter
2. Rate systems are classified as medical devices once they record breaks
3. The safety representative takes over the approval of every roster and target from the line manager, who keeps only the follow-up conversation after a missed target
4. Direction may be delegated at Level 4 Autonomous provided occupational health is informed

Chapter 7 · 90 minutes

Hiring by Machine: Screening, Scoring and Automated Interviews

The lesson treats hiring as M3 Consequence reached through M2 Evaluation of people who are not yet employees. Participants study the automated funnel, four documented failures, the record of human recruiters, the limits of audits, Annex III point 4(a), the prohibition on emotion recognition and co-determination over selection guidelines. They work with the profile DLG-HM-REC-01 and the pause record AIDR-0601 and write the conditions that keep the decision with a named recruiter.

Learning objectives

Explain why the degree of delegation in screening is set by use and not by the configuration screen.
Compare the documented failures of automated hiring with the documented record of human recruiters.
Distinguish prohibited emotion inference from high-risk scoring of interview content, as the chapter draws the line.
Draft the conditions of a delegation profile that keep a hiring decision with a named recruiter.

Key concepts

Read below the cut: A condition that the recruiter reads a defined share of applications ranked below the shortlist before any rejection is sent; DLG-HM-REC-01 starts at 20 per cent. It makes Level 1 Assisted true in practice.
Blind review: A monthly sample of rejected applications, 50 in the worked example, re-assessed by a second recruiter without the ranking, so that divergence between people and the system can be detected.
Configuration as selection guideline: A screening model's criteria, weights, knock-out questions, reading threshold and excluded features are in substance the requirements a BetrVG section 95 guideline sets out. Treated as a vendor setting, it escapes co-determination.
Emotion recognition prohibition: Article 5(1)(f) of the AI Act prohibits inferring emotions in the workplace except for medical or safety reasons; the Commission's guidelines apply it to candidates. It covers inference from biometric data, not sentiment analysis of written text.
Pause record: An AI Decision Record of a decision to stop a system, with trigger, questions asked, supplier answers, reasons and conditions for return. AIDR-0601 records Nordvik Region's pause of its screening pilot.

Teaching plan

15 min	The teacher walks the five stages of the automated funnel and the example of eight hundred applications of which recruiters read the top fifty.
20 min	Groups take one case each (the scrapped Amazon model, iTutorGroup, HireVue, Mobley v. Workday) and one study of human hiring, and report the mechanism in two sentences.
15 min	Presentation of the deployer's duties under Article 26, the applicant's rights, the prohibition in Article 5(1)(f) and the two routes of co-determination, German and Nordic.
30 min	Exercise: groups act as Nordvik Region preparing the screening module's return and write the profile conditions that answer AIDR-0601.
10 min	Debrief with MDR-HM-0070: what the recruiter read, whom she shortlisted and dropped, and what an applicant would be told under Article 86.

Discussion

The chapter says unconstrained human discretion is not a neutral benchmark, and that an automated process can at least be measured if the deployer keeps the records. What follows for how the two should be compared?
The New York audit law produced, in the chapter's words, documents and not scrutiny. What does a transparency duty need around it to have an effect?
Nordvik Region did not find discrimination; it found that it could not explain a system it had deployed, and stopped. Under what conditions would your organisation take the same decision?

Exercise · 30 minutes

Conditions for the return of a screening module

AIDR-0601 lists three conditions for the return of Nordvik Region's screening module: supplier documentation, a delegation profile with read-below-cut and blind review agreed in negotiation, and a fundamental rights impact assessment. Write the delegation profile. Using DLG-HM-REC-01 as the model, set the degree for M2 Evaluation and M3 Consequence, name the decider, list the forbidden actions (such as rejection without a recruiter decision), and set the reading share, the blind-review sample, the outcome monitoring, the log retention and the applicant information. Add the list of evidence to demand from the supplier, taken from the four questions the union asked, and the triggers on which the module is paused again.

Deliverable: A delegation profile for the screening module with degrees, named decider, forbidden actions, reading and review conditions, supplier evidence list and pause triggers.

Self-test

According to the chapter, which use of a video-interview system has been prohibited in EU recruitment since 2 February 2025?
1. Scoring the content of an interview transcript against structured criteria that a recruiter has defined for the vacancy
2. Transcribing a one-way video interview for a recruiter to read
3. Ranking CVs against published requirements before the interview
4. Analysing facial expression or tone of voice to infer nervousness, honesty or another emotional state
What did the supplier answer when Nordvik Region's union asked which criteria and weights produced the ranking?
1. It disclosed the full list of criteria and their weights within a week
2. It did not disclose them and described them as proprietary
3. It said the ranking used only legal eligibility to work
4. It said the model had been trained on the Region's own past hiring decisions
Why does the chapter say a supplier's adverse-impact report based on the US 4/5 rule is not evidence of compliance in Europe?
1. Because the equality law the chapter cites has no such threshold and turns on proportionality and justification, with the burden of proof shifting to the respondent
2. Because the AI Act forbids suppliers from carrying out any bias audit themselves and reserves audits of Annex III point 4(a) systems to the market surveillance authority of each member state
3. Because the 4/5 rule applies only to video interviews and not to CV ranking
4. Because European works councils must carry out all audits of selection tools in person

Chapter 8 · 75 minutes

Performance Scores, Productivity Metrics and Dismissal by Algorithm

The lesson examines how individual performance scores are built, what they miss, and how they become decisions about pay, discipline and dismissal. Participants compare the reported cases from Baltimore, Amazon Flex, Xsolla and Amsterdam, learn the three administrative channels from M2 Evaluation to M3 Consequence and the five conditions that keep a score advisory, and run the practice test that showed Hansa Mutual's sales-performance score to be decisive in one unit.

Learning objectives

Describe the three layers of a performance score and the kinds of work scores systematically miss.
Compare the reported cases by what the system did and where the human stood.
Explain the three channels by which a score moves from M2 Evaluation to M3 Consequence.
Design a practice test that establishes from records whether an advisory score is decisive.

Key concepts

Three channels: The pay input, where the manager's assessment converges on the score because departing requires an explanation; the warning, a threshold notice that is dated, filed and later cited; and the list, where the bottom of a ranking becomes the shortlist.
Decisive in practice: A score classified as advisory that is followed in nearly every case, in seconds, with few recorded departures. If two or more alerts of the practice test fire, the use is treated as Level 3 Bounded or above and suspended.
Five conditions for an advisory score: The manager forms a view before seeing the score; thresholds trigger a conversation, not a sanction; the score is never the sole ground; the worker sees what the manager sees; and departures are expected and recorded.
Seeded cases: Test cases agreed with the representatives in which the score misreads the work, such as approved leave counted as low activity; at least three per unit per year, removed before any real decision.
Recorded exclusion: A use of AI that an organisation has decided against and written into the profile, such as discipline or dismissal. A recorded exclusion can be audited, whereas an absence can quietly be filled.

Teaching plan

10 min	The teacher builds a score in three layers on the board and asks participants which parts of their own work would leave no log.
15 min	Participants complete the chapter's table for the four cases, then discuss the Horizon lesson about treating system output as fact.
15 min	Presentation of GDPR Article 22 as read in SCHUFA and Amsterdam, German and Swedish dismissal law, the three channels and the five conditions.
25 min	Exercise: groups design the practice test for one score and apply it to the figures of the Dutch sales unit in the worked example.
10 min	Debrief on how to run the test without monitoring individual managers, and on what is reported to representatives and to the workers affected.

Discussion

At Xsolla a named executive took the decision but the analysis chose the people. Does a named decider satisfy the book's rule in that case, and what is missing?
The chapter says a manager's assessment converges on the score because departing requires an explanation and following does not. How could an organisation reverse that asymmetry?
Neither organisation uses AI in discipline or dismissal, and both record that as a decision. What is gained by recording something an organisation does not do?

Exercise · 25 minutes

Is the score decisive in practice?

Design a practice test for one score used as an input to a consequence: Hansa Mutual's sales-performance score under DLG-HM-SPS-01, or a score in your own organisation. Agree the design as if with the representatives: results per unit of at least ten managers, never per named manager. Define how you will measure agreement with the score band, departures and their written reasons, and the time from opening the score to the decision, with the chapter's default alerts. Specify three seeded cases. Then apply the test to the Dutch sales unit as the chapter reports it, classify the result and write the measures that follow.

Deliverable: A one-page test design with measures, alert thresholds and seeded cases, plus a classification of the Dutch sales unit and the corrective measures for the delegation register.

Self-test

What did the review of CNT-HM-0001 find in Hansa Mutual's Dutch sales unit?
1. The score had been used to select field-sales staff for redundancy in a restructuring, without the ondernemingsraad having been asked for its consent
2. Managers had departed from the score band in half of all proposals without reasons
3. 97 per cent of pay proposals matched the score band, with a median of 48 seconds from opening the score to recording the proposal
4. The score was built on inferred emotions from recorded sales calls
In SCHUFA, what did the Court of Justice hold about an automated score?
1. It is itself automated decision-making where a third party draws strongly on it to establish, implement or terminate a contractual relationship
2. It is always a preparatory act and falls outside Article 22 of the GDPR, provided that the party using it has recorded the score as advisory in its own procedures
3. It may be used freely provided a person signs the final decision
4. It is lawful only if it has been audited under the 4/5 rule
According to the first of the chapter's five conditions, what must happen before a score is shown to a manager for an M3 Consequence use?
1. The works council must approve the individual figure for each employee
2. The supplier must certify that the score contains no errors
3. The employee must sign a declaration accepting the score
4. The manager records a short assessment of the work, forming a view first

Chapter 9 · 90 minutes

Scheduling, Pay and the Allocation of Work

The lesson shows that a roster is also a pay slip and a rest plan, and that an allocation settles who learns. Participants study rostering as constrained optimisation, the research on unstable schedules, the Working Time Directive and the duty to record actual time, the draft guidelines on task allocation as high-risk, and pay transparency for premium-bearing shifts. They read MDR-NR-0001 and MDR-HM-0001 and negotiate a delegation profile with input lists and change classes.

Learning objectives

Explain how a roster or an allocation reaches M3 Consequence through hours, premiums and the mix of work.
Summarise what the cited studies measured about schedule instability, notice and quick returns.
Apply the draft guidelines' distinction between behavioural inputs and availability to an input list.
Negotiate a delegation profile for a rostering or allocation system with degrees, inputs and change classes.

Key concepts

Objective function: What the optimiser weighs: coverage against cost, and preferences against both. Every weight is a management choice, recorded or not, and the chapter asks that it be written in plain language.
Input list: The allowed and prohibited inputs of the system. The input list decides the AI Act classification, so it belongs in the delegation profile, and adding an input is a change that needs review.
Premium-bearing shifts: Evening, night and weekend shifts that carry supplements. The rule that distributes them is in effect a pay-setting criterion, which the Pay Transparency Directive requires to be objective and gender-neutral.
Pattern review by person: A regular review of hours, premium-bearing shifts, quick returns, items by complexity tier and manager changes, reported to representatives at group level and to individuals only about themselves.
Change classes: The agreement on which changes reopen negotiation, such as the objective function, inputs or offer rules, and which are only notified, such as forecast refreshes and bug fixes.

Teaching plan

15 min	The teacher presents the roster as an optimisation problem and the three ways software shapes pay: hours, premiums and direct pricing.
15 min	Evidence round: participants summarise the retail survey, the Seattle ordinance study, the quick-returns review and the Gap experiment, each with its stated limit.
15 min	Presentation of the working-time floor, CCOO v Deutsche Bank, Directive (EU) 2019/1152, the draft guidelines on allocation, and the right degree for each decision.
35 min	Exercise: a negotiation in three roles (system owner, union or works council, line manager) producing a delegation profile for a rostering or allocation system.
10 min	Debrief with the two first records: the three changes Mikael Strand made in MDR-NR-0001 and the lasting correction in MDR-HM-0001.

Discussion

In MDR-HM-0001 the lasting correction was a capacity field the team leader fills in without giving reasons. Why does the chapter call that the valuable part, and what does it say about M5 Care?
The chapter says a unit that has not drafted a roster by hand for three years has delegated more than its profile says. How would you test that in a unit you know?
An optimiser that gives evening and weekend shifts to those with the widest availability can produce a pay gap without any gender variable. Who in your organisation would see that effect, and when?

Exercise · 35 minutes

Negotiate a rostering profile

In groups of three, take the roles of system owner, workers' representative and line manager, and negotiate a delegation profile for a rostering or allocation system, following the chapter's practical guide. List each decision the system takes or supports by management function and agree a degree and a named decider for each. Publish the input list with prohibited inputs. Write the objective function in plain language and agree the rule for distributing premium-bearing shifts and extra hours. Set the publication notice, the worker's right to decline an offer without consequence, the pattern review by person and the change classes that reopen the negotiation. Record one point on which the parties did not agree.

Deliverable: A negotiated profile in the form of DLG-NR-ROS-01 with degrees and deciders, allowed and prohibited inputs, premium-shift rule, notice period, pattern review, change classes and one recorded disagreement.

Self-test

According to the Commission's draft guidelines as the chapter reports them, which basis for allocating tasks falls outside Annex III, point 4(b)?
1. A ranking of workers from punctuality and no-show history
2. Availability as workers indicate it, a required accreditation or proximity
3. Acceptance rates combined with customer ratings
4. Reliability metrics that downgrade a worker who declines an offer
Why did Mikael Strand move a weekend evening shift in the Strandvik unit's first roster, as recorded in MDR-NR-0001?
1. The system had breached the 48-hour weekly limit for the whole unit
2. The colleague had declined three open-shift offers in a row
3. The supplier had changed the forecast model without notice
4. The draft gave one person four premium-bearing shifts against a unit rotation of three
What did the Grand Chamber hold in CCOO v Deutsche Bank that matters for a rostering system?
1. Employers must publish every roster at least two weeks in advance in every member state, unless a collective agreement sets a longer period
2. Rostering software is always a high-risk AI system under Annex III
3. Employers must set up an objective, reliable and accessible system measuring the time each worker works each day
4. Daily rest may be reduced below 11 hours whenever an optimiser recommends it

Chapter 10 · 75 minutes

Knowledge Work: The AI Manager Enters the Office

The lesson follows the AI boss into the office, where it arrives as features: an assistant that names owners of action items, a workflow agent that places tickets, an analytics pane about people. Participants learn the four tests that show when an assistant becomes a boss, what the productivity studies did and did not measure, the transparency paradox, and why such features are technical equipment suited to monitor. They build a feature register on the chapter's model.

Learning objectives

Apply the four tests (addressee, default, trace, decision) to office features.
Explain how meeting summaries, workflow allocation and analytics touch M1 Direction, M2 Evaluation and M4 Development.
State what the cited productivity studies measured and why they are not evidence about AI performing management functions.
Build a feature register entry with function, mode, degree, unit of reporting and agreement status.

Key concepts

Four tests: The addressee: the output names and reaches another person. The default: it takes effect unless someone intervenes. The trace: it is stored where it can inform a review. The decision: nobody decided that the feature should perform a management function.
Feature register: An extension of the delegation register that lists features, not products, with management function, usage mode, degree, unit of reporting, prohibited uses, accountable person and agreement status for each.
Transparency paradox: Bernstein's finding that observability can prompt workers to conceal what they do, so that the organisation sees an illusion of transparency; shielding production lines from routine observation improved their performance.
Unit of reporting: The level at which a figure is shown. By default, figures about collaboration, focus or wellbeing go to the individual alone or to groups of at least ten, and are never ranked by person.
Development rule: A rule in a workflow agent's delegation profile, beside the throughput rule, so that allocation does not send every complex case to the fastest experienced handler and starve juniors of the work from which they learn.

Teaching plan

15 min	The teacher follows one meeting summary from transcript to task list and shows the degrees: Level 0 Advisory with the chair confirming, Level 1 Assisted, Level 2 Supervised.
15 min	Participants apply the four tests to three features: a meeting assistant, a ticket-allocating agent and an analytics pane showing individual focus time.
10 min	Presentation of what was measured in the productivity studies, the transparency paradox, and BetrVG section 87(1) no. 6 read as suited to monitoring.
25 min	Exercise: groups build a feature register for one office suite and take one entry through the representatives' process on paper.
10 min	Debrief with the two decisions: Hansa Mutual's AIDR-0670 keeping analytics switched off group-wide, and Nordvik Region's sample of 12 meetings leading to AIDR-0620.

Discussion

The chapter says the default of most products, distribution of action items without the chair's approval, has to be changed deliberately. Who in your organisation would take that decision, and would they know it was theirs?
Hansa Mutual rejected enabling only the personal insights, and enabling the feature outside Germany. What are the arguments for and against one group-wide standard?
Figures reported at team level, to the team, preserve the space in which people try better ways of working. Where does that argument reach its limit for a manager who must still exercise M5 Care?

Exercise · 25 minutes

A feature register for the office suite

List at least five features, not products, of the office suite, collaboration tools or workflow agents in your own organisation, or use the chapter's examples at Nordvik Region and Hansa Mutual. Apply the four tests to each. For every feature that names or measures another person, write a register entry: the management functions it touches, the usage mode, the degree per function using the chapter's defaults, the unit of reporting, the prohibited uses, the accountable person and the status of the representatives' agreement. Mark any feature that passes the first three tests and fails the fourth, and say what must be decided before it stays enabled.

Deliverable: A feature register with at least three complete entries in the form of FTR-NR-SA-01, and a list of undecided features with the decision each needs.

Self-test

What did Nordvik Region find when it sampled the summaries of 12 pilot meetings from December 2026?
1. Of 48 action items, 9 carried owners the chair had not confirmed, and 3 of those 9 people had treated the item as an instruction
2. All 48 action items had been confirmed by the chair before distribution
3. The assistant had filed a summary per person for each of the 12 meetings, and unit managers had read 9 of them before development conversations with their staff
4. Half of the summaries had been used as evidence in development conversations
A feature passes the addressee, default and trace tests but fails the fourth test. What does the chapter call the result?
1. A prohibited practice under Article 5 of the AI Act
2. A Mode 5 Autonomous production agents deployment
3. An AI boss the organisation did not choose
4. A system that falls outside co-determination
Why does the chapter say the cited productivity studies of generative AI leave the evidence base for the office AI boss thin?
1. They were all commissioned by the vendors of the assistants studied, as the Work Trend Index was, so none of their results is independent
2. They measured AI assisting individual workers; none measured AI performing management functions
3. They found that assistants reduced output for every group of workers
4. They covered only factory work and cannot be applied to offices

Module 3 · Part III · What It Does to People

Part III tests the book's thesis against the evidence on what the AI boss does to people. Five debates, each staged as the case for concern, the case for calm, where the positions meet and what can be done, cover autonomy and dignity, fairness and contestability, surveillance, health and pace, and whether algorithmic management works at all. Participants leave able to state both positions at their strongest, to label findings observable, expected or speculation, and to run the five instruments the chapters supply.

Lecture · 50 minutes

Five Debates and Where They Meet: What the AI Boss Does to People

Part III stages five debates in the same four steps, and this lecture shows what they have in common. In each, the case for concern describes what happens when design is left to default, and the case for calm describes what is possible when it is not; both compare the AI boss with real managers, not ideal ones. The positions meet on conditions that an organisation can design and negotiate: discretion, reasons and a reachable person; outcomes measured by group; purpose and aggregation of data; health limits inside the optimiser; and an evaluation agreed before go-live. The verdicts separate the observable from the expected and from speculation.

6 min	The form of the debate Four steps before a verdict: the case for concern, the case for calm, where the positions meet and what can be done. Each position is stated at its strongest, and what a source reports is kept apart from interpretation. Every verdict is labelled per management function as observable, expected or speculation, and ends with what can be influenced.
9 min	Autonomy and dignity: discretion, reasons, a reachable person Self-determination theory, the job characteristics model and Karasek's demand-control model show that a system's design choices are job design decisions. In Lee's experiment people accepted algorithmic assignment about as readily as human assignment, resisted algorithmic evaluation, and missed recognition. The harm is not inherent in delegating direction; it follows from how much discretion the design leaves.
9 min	Fairness: measure by group, correct the rule Bias enters through the definition of good performance, proxies, historical data and feedback loops, so a system can pass every individual check and still disadvantage a group. The appeal fails without notice, without reasons, without anyone to ask, or with a remedy that fixes one case and leaves the rule. Hansa Mutual's test of claims allocation found a rule on consecutive working days that kept part-time handlers from complex claims.
8 min	Surveillance: what a system requires and what it invites A visit-routing system requires the start and end of each visit; the by-products invite a figure per person, then a ranking, then a pay input. Courts and authorities apply a consistent test of notice, necessity, proportionality and safeguards, and treat undifferentiated retention harshly. The same data can protect or chill; purpose, aggregation, retention and access decide which.
9 min	Health and pace: care designed into direction An optimiser removes slack because slack looks like waste, yet those are the minutes in which a person recovers. A route at Strandvik complied with every working-time rule and still could not be worked; travel time became a hard constraint written as a test. Care cannot be delegated to the system that directs the work, but it must be designed into it.
9 min	Does it work? Quarters and years Controlled studies show real short-run gains from assistance, largest for novices; a meta-analysis of general monitoring found no average gain in performance. Gaming and the ironies of automation take time, so first-quarter gains on the optimised metric are the least informative evidence. Q4 stays open: most organisations do not know whether it works because they have not designed a way to find out.

What the audience takes away

In every debate the outcome depends on design and negotiation, not on the technology alone: what is left to default produces the case for concern.
The honest comparison is with real managers, and the honest evidence is outcomes measured by group, at team level and over years.
A verdict is only as good as its labels: say what is observable, what is expected and what remains speculation.

Chapter 11 · 75 minutes

Autonomy, Dignity and Control: Being Managed by a Machine

The lesson opens Part III with the debate on autonomy, dignity and control under algorithmic direction. Participants apply self-determination theory, the job characteristics model and Karasek's demand-control model to a system's design choices, weigh the evidence on algorithm aversion and appreciation, and stage the debate to its verdict. They design an autonomy and fairness survey like SVY-NR-STV-01 and negotiate discretion into a delegation profile, as PREF-01 does.

Learning objectives

Explain how algorithmic direction acts on autonomy, competence and relatedness, and on demands and decision latitude.
State the case for concern and the case for calm on autonomy at their strongest.
Identify the three properties on which both positions meet: discretion, reasons and a reachable person.
Design a before-and-after measurement of autonomy and perceived fairness that protects the respondents.

Key concepts

Decision latitude: In Karasek's model, the authority to make decisions about the work and the discretion to use one's skills. High demands with low latitude make a high-strain job. In the book's terms, latitude is a property of the delegation profile.
Three basic needs: Ryan and Deci's competence, autonomy and relatedness. Autonomy suffers when a system sets sequence and pace without reasons, competence when feedback is a score whose basis cannot be seen, relatedness when an interface replaces the manager.
Mechanical and human tasks: In Lee's online experiment, algorithmic and human decisions were perceived as equally fair for tasks seen as requiring mechanical skills; for tasks seen as requiring human skills, algorithmic decisions were perceived as less fair and as dehumanising.
Bounded preference rule: A rule such as PREF-01 that lets handlers state preferences within a cap, considered after deadlines, competence and workload balance, with protection for newer handlers so that an M1 Direction rule does not become a question of M4 Development.

Teaching plan

15 min	The teacher presents what the platform and warehouse studies report, then the three models of work psychology, and asks which design choices of a routing system act on each.
20 min	Staged debate: half the room prepares the case for concern and half the case for calm from the chapter; each side presents for five minutes without rebuttal.
10 min	Plenary: where the positions meet, and the verdict per function with its labels observable, expected and speculation.
20 min	Exercise: groups design a survey plan on the model of SVY-NR-STV-01 for one unit about to be directed by a system.
10 min	Debrief on the second open question of the survey, what has been lost, and on the informal swaps at Strandvik that no system field records.

Discussion

The same freelancers reported high flexibility and autonomy alongside isolation, overwork and exhaustion. How can both findings come from the same design?
The case for calm says the harms come from business models and weak institutions, not from algorithms. What evidence in the chapter supports that reading, and what resists it?
At Hansa Mutual handlers describe the agent's allocation as fairer than the old shared queue, yet experienced handlers miss choosing complex claims. How should a profile weigh those two experiences?

Exercise · 20 minutes

An autonomy and fairness survey plan

Following the chapter's practical guide, write a survey plan for one unit where a system is about to take over part of M1 Direction: the Strandvik home-care unit, or a unit of your own. Name the change, the system and the delegation profile. Draft 12 to 16 statements on a five-point scale covering latitude over sequence, method and pace, the three needs, reasons, a reachable person and fairness of allocation, plus the two open questions. Set the three waves, the comparison unit, the minimum group size, the retention of raw responses, the first audience for results, the prohibited uses and the threshold at which the delegation profile is reviewed.

Deliverable: A survey plan in the form of SVY-NR-STV-01 with statements by dimension, waves, reporting and data rules, prohibited uses and the action threshold.

Self-test

What did Lee's online experiment with four managerial decisions find about how people perceive algorithmic decisions?
1. They were perceived as less fair than human decisions for every kind of task, including work assignment and scheduling, and as dehumanising throughout
2. They were perceived as fairer than human decisions for tasks requiring human skills
3. They were perceived as equally fair for mechanical tasks, but less fair and dehumanising for tasks seen as requiring human skills
4. They were perceived as trustworthy only when a supplier had audited them
In version 1.1.0 of DLG-HM-CTA-01, what does the protection for development in PREF-01 guarantee?
1. At least one complex claim a week continues to go to each handler in the first two years of service
2. Experienced handlers receive every complex claim in their three preferred categories
3. Preferred claims make up at least 30 per cent of every handler's weekly allocation
4. Newer handlers are exempt from the agent's allocation for their first two years
According to the chapter's practical guide, what triggers a review of the delegation profile after the survey waves?
1. Any individual respondent scoring below the unit average on the statements about latitude over sequence, method and pace
2. A response rate below half of the unit's staff
3. A request from the supplier of the system to change the questions
4. A fall of more than half a scale point on any dimension, or open answers that name a specific loss

Chapter 12 · 90 minutes

Fairness, Bias and the Appeal That Never Comes

The lesson stages the debate on fairness, bias and contestability in M2 Evaluation and M3 Consequence. Participants study how bias was documented in the Bologna ranking and the rideshare earnings gap, the four mechanisms by which it enters, the four ways an appeal fails, and what the GDPR, AI Act Article 86, equality law and the Platform Work Directive require of reasons and review. They run a disparate-impact test on the model of DIT-HM-CTA-01.

Learning objectives

Explain the four mechanisms by which bias enters a management system without anyone intending it.
Identify the four ways in which a route of appeal fails.
Compare the case for concern and the case for calm on fairness, and state the four conditions on which they meet.
Design a disparate-impact test with representatives and interpret a triggered result.

Key concepts

Four mechanisms of bias: The definition of good performance, proxies that correlate with protected characteristics, historical data that codifies earlier decisions, and feedback loops in which outputs influence future inputs.
Indirect discrimination: An apparently neutral provision that puts persons with a protected characteristic at a particular disadvantage unless objectively justified by appropriate and necessary means, with the burden of proof shifting to the respondent once a presumption arises.
The appeal that never comes: Contest fails in four ways: no notice that a score exists, no reasons beyond a threshold, no one to ask with authority to change the decision, and a remedy that reverses one case while the rule stays.
Disparate-impact test: A measurement of each group's rate of a favourable outcome relative to a reference group, with a reporting floor of 20 per cell. A lower bound below 0.80 triggers investigation and is not a finding of discrimination.
Correcting the system, not only the case: A successful contest goes to the owner of the system so that the rule is changed and the affected cases are listed for their deciders; a correction that leaves the rule in place corrects nothing.

Teaching plan

15 min	The teacher tells the Bologna case and the rideshare earnings gap side by side and asks what each shows and does not show.
15 min	Presentation of the four mechanisms and the four failures of contest, with the Amsterdam, SyRI and childcare-benefits cases as illustrations.
20 min	Staged debate in four steps; the class then lists the four conditions on which the positions meet and the verdict's labels.
30 min	Exercise: groups design a disparate-impact test for one management system and interpret the results of DIT-HM-CTA-01.
10 min	Debrief on Nordvik Region's finding in DIT-NR-ROS-01: a justified competence filter and an unjustified gap in access to the course.

Discussion

The Bologna score rewarded reliability and had no field for a reason. What other apparently neutral definitions of good performance could penalise the exercise of a right?
The case for calm says a rule in software can be run on every case and compared by group, whereas the judgement of three hundred line managers cannot. How far does auditability answer the case for concern?
Hansa Mutual flagged promotion files for review after finding the disparity. Who should review them, and what should the people concerned be told?

Exercise · 30 minutes

A disparate-impact test with the representatives

Design a disparate-impact test for one system that evaluates people or allocates work, pay or opportunities: ShiftWise open-shift offers, Hansa Mutual's sales-performance score, or a system of your own. State the question agreed with the representatives and the outcomes that matter to the people affected. Choose the groups from characteristics lawfully held, with at least one intersection, and set the reporting floor. Define the measure and the trigger. Then read the results of DIT-HM-CTA-01: write in three sentences what the record shows and does not show, why rule CTA-R-14 had that effect, and what the correction must cover beyond the rule itself.

Deliverable: A test design with question, groups, intersection, cell floor, measure and trigger, plus a short written interpretation of DIT-HM-CTA-01 and its correction.

Self-test

What caused the triggered result for part-time handlers in DIT-HM-CTA-01?
1. Team leaders used their right to reassign under DLG-HM-CTA-01 to move complex claims away from part-time handlers by hand
2. The agent used sex as a direct input to the allocation of complex claims
3. Part-time handlers had stated a preference for routine claims under PREF-01
4. Rule CTA-R-14 offered complex claims only to handlers with two consecutive working days left in the week
How does the chapter's practical guide treat a ratio whose 95 per cent interval has a lower bound below 0.80?
1. As a finding of unlawful discrimination that must be reported to the authority
2. As a trigger for investigation, not a finding of discrimination
3. As within tolerance provided the reference group is larger
4. As grounds for dismissing the manager who owns the system
What does the study of more than a million rideshare drivers show, according to the chapter?
1. A rule treating every individual identically can produce an outcome gap between groups
2. Algorithmic pay formulas are proven to be fair to all groups
3. Customer discrimination explains most of the gender earnings gap
4. Pay algorithms always contain a hidden variable for the driver's sex

Chapter 13 · 75 minutes

Surveillance at Work and the Chilling Effect

The lesson stages the debate on surveillance at work and its chilling effect, concerning M2 Evaluation and M5 Care. Participants distinguish what a management system requires from what it invites, apply the criteria of the Strasbourg cases and the reasoning of the H&M and Amazon decisions, read Article 88 of the GDPR and the hard lines on emotion inference, and carry out a proportionality review of monitoring data as Nordvik Region did in AIDR-0604.

Learning objectives

Distinguish the data a management system requires from the further uses its by-products invite.
Apply the criteria of the cited court and authority decisions to a monitoring arrangement.
Compare the case for concern and the case for calm on monitoring and name the points on which they meet.
Carry out a proportionality review that sets purpose, aggregation, access, retention and the route to M3 Consequence.

Key concepts

Requires and invites: A visit-routing system requires the start and end of each visit, not location between tasks. By-products invite further use: once times exist per person, a figure per person can be computed, then a ranking, then a pay input.
Chilling effect: When every action may enter an evaluation, helping a colleague or questioning a decision incurs an observable cost for an unobservable benefit. The chapter labels a lasting effect on voice in AI-managed workplaces as expected, not yet measured.
Suited to monitor: The German Federal Labour Court reads equipment intended to monitor objectively, so equipment is covered by co-determination when it is suited to collect information on behaviour or performance, whatever anyone intends.
Aggregation by default: Team or route before person, week before day. Individual figures need a written reason and a named role, every individual-level view is logged, and representatives may review the log.
Retention by purpose: Each data element is kept only as long as its stated purpose needs. The Conseil d'Etat upheld that keeping every indicator for every employee, undifferentiated, for 31 days breached data minimisation.

Teaching plan

10 min	The teacher asks participants to list what a routing or allocation system needs to know, then what it also collects, using the chapter's distinction between requires and invites.
15 min	Case round: groups summarise one decision each (the two Grand Chamber cases, H&M, Amazon France Logistique) by what was found and the criterion that decided it.
15 min	Staged debate in four steps, followed by the verdict and its labels.
25 min	Exercise: groups run a proportionality review of one data stream, following the practical guide and Nordvik Region's review of visit registration.
10 min	Debrief with the annex to KBV-IT-2024: who sees individual throughput, what is prohibited, and how the principle is applied in Switzerland and Poland.

Discussion

The H&M case involved no AI. Why does the chapter include it, and what does it show about what a management system could assemble routinely?
The restaurant study found less theft and higher legitimate tips, while the factory study found concealment under constant observation. What differs between the two forms of monitoring?
Mikael Strand keeps individual figures because he needs to see when a person's routes are consistently overrun. Where is the line between that use and evaluation of the person?

Exercise · 25 minutes

Proportionality review of monitoring data

Review the monitoring data used by one management system: the visit-registration data behind Nordvik Region's home-care dashboards, the throughput figures of Hansa Mutual's claims-triage agent, or a system of your own. Inventory each data element, including by-products, and state its purpose under M1 Direction, M2 Evaluation or M5 Care; an element with no purpose is not collected. Exclude what the chapter lists as forbidden. Set the aggregation level, the roles with access and the logging of individual views, and the retention for raw events and for planning aggregates. State for each element whether any route to M3 Consequence is closed or recorded, and how workers are informed before use.

Deliverable: A review table with purpose, aggregation, access, retention and route to M3 Consequence for each data element, the exclusions, and the date of the next review.

Self-test

What did Nordvik Region's review, recorded as AIDR-0604, do with the location trace the app stored between visits?
1. It kept the trace for 24 months as a planning aggregate
2. It stopped collecting it, keeping location only as confirmation at the start and end of a visit
3. It made the trace visible to each worker's colleagues in the roster app, so that informal swaps between routes became easier to arrange
4. It moved the trace into the supplier's system for model training
According to the Commission's guidelines as the chapter reports them, which statement about Article 5(1)(f) of the AI Act is correct?
1. Inference of emotions from written text is inside the prohibition, whereas inference from voice is outside it wherever the employer claims a safety reason
2. The prohibition applies only to recruitment and not to existing staff
3. Inference from keystrokes, facial expressions or voice is inside it, and the safety exception does not cover stress or burnout detection
4. The Digital Omnibus removed the prohibition for employers with a works agreement
Under the annex to KBV-IT-2024, who may see a claims handler's individual throughput figure?
1. The handler and the handler's own team leader
2. Every team leader in the German claims organisation
3. The Konzernbetriebsrat and the group head of claims operations
4. All members of the handler's team, as a ranking list

Chapter 14 · 75 minutes

Health, Pace and Burnout: Occupational Safety Under Algorithmic Direction

The lesson treats the gap between M1 Direction and M5 Care when a system sets the pace and people remain responsible for its effects. Participants apply the demand-control and effort-reward models, weigh the evidence on monitored pace, schedules and injury, and read the employer's duty in the Framework Directive and Swedish and German law. They stage the debate and write health and rest constraints into an optimiser as tests, on the model of WRA-NR-HC-01 revision 2.

Learning objectives

Explain with the strain models where algorithmic direction is predicted to harm health.
Assess the strength and limits of the evidence on algorithmically set pace and injury.
Explain why a software-set work-rate is a working method the employer must assess and consult on.
Write health and rest constraints as hard constraints with sources and tests.

Key concepts

Unhealthy workload: As the Swedish Work Environment Authority defines it: demands that more than temporarily exceed resources, becoming unhealthy when the imbalance is long-lasting and recovery is insufficient.
Hard and soft constraints: By default everything with a legal or agreed source is hard; preferences are soft and breached only with a recorded reason. A constraint without a source is negotiated away by the first person who finds it inconvenient.
Feasibility condition: A hard constraint the optimiser cannot trade for output. The contract requires this in place of a penalty term, because enough output always outweighs a penalty.
Seeded breaches: Known breaches inserted on every change to confirm that the constraint tests catch them; WRA-NR-HC-01 seeds 12 per release.
Care designed in, not delegated: The duty of care stays with the line manager, the safety representative and the committee that approves the limits, while the system contributes enforcement and detection at Level 1 Assisted.

Teaching plan

15 min	The teacher presents the demand-control and effort-reward models and has participants predict the effect of three design choices: shorter planned times, a fixed visit order, a removed colleague.
10 min	Evidence review: what the warehouse analyses, the EU-OSHA findings and the schedule studies measured, who produced them and what the chapter says lifts them above a claim.
15 min	Staged debate in four steps; the class names the three conditions both sides accept and the verdict's labels.
25 min	Exercise: groups write constraints with source, class and test for one optimiser, with strain indicators and a manual fallback.
10 min	Debrief on Hansa Mutual: why declining to delegate care removes the instrument that would show the risk, and how AIDR-0685 answers with a team-level indicator and a cap.

Discussion

The route contested in CNT-NR-0001 complied with every working-time rule and still could not be worked. What does that say about the limits of legal constraints as a measure of safe pace?
The case for calm says a system that checks every pair of shifts overlooks a rest rule less easily than a manager drafting by hand. What must also be true for that advantage to protect health?
Section 7 of chapter 2 of AFS 2023:2 requires that workers know which tasks to prioritise when time does not suffice and whom to turn to. How would you test a dispatch system against it?

Exercise · 25 minutes

Health constraints as tests

For one system that performs M1 Direction, such as ShiftWise routes in home care or the allocation rules of a claims queue, write at least four health and rest constraints. For each give the rule, its source (law, collective agreement, risk assessment or a rule agreed with representatives), its class as hard or soft, its owner and a one-sentence description of the test that would run on every draft plan before it reaches the approving manager. Add the breaches you would seed on each release, the monthly team-level strain indicators and their recipients, who may invoke the manual fallback and how often it is tested, and which changes require assessment and consultation beforehand.

Deliverable: A constraint table in the form of WRA-NR-HC-01 revision 2 with source, class, owner and test, plus seeded breaches, strain indicators, stop path and change triggers.

Self-test

Which hard constraint did revision 2 of WRA-NR-HC-01 add after the contestation CNT-NR-0001?
1. A cap on the number of open shifts that ShiftWise may offer to hourly staff each month, set by the central work-environment committee
2. A ban on showing planned visit times to the unit manager
3. A requirement that every route be approved by the supplier before publication
4. Travel time between visits of at least routed time plus 5 minutes, with a minimum of 10 minutes
Why does the chapter's practical guide require that hard constraints be feasibility conditions and not penalty terms?
1. Because enough output always outweighs a penalty
2. Because penalty terms are prohibited by the AI Act
3. Because feasibility conditions are cheaper for the supplier to build
4. Because penalty terms cannot be read by the safety representative
How does the chapter's verdict label the claim that AI agents assigning office work raise burnout?
1. Observable, because warehouse injury data already show it
2. Expected, because the Platform Work Directive presumes it
3. Speculation until measured
4. Refuted, because assistants remove drudgery in every study

Chapter 15 · 90 minutes

Does It Work? Productivity, Quality and the Evidence for Algorithmic Management

The lesson asks whether algorithmic management works, and over what horizon, for M1 Direction, M2 Evaluation and M4 Development. Participants sort the evidence by six outcomes, by horizon and by source, examine why short-run gains and long-run performance diverge, and stage the debate to a divided verdict on open question Q4. They design a multi-year evaluation with a comparison fixed in advance and stop rules, on the model of EVP-NR-ROS-2027-01.

Learning objectives

Classify a study by which of the six outcomes it measured, over what horizon and from what kind of source.
Summarise what the controlled studies of AI assistance and of monitoring found, including their limits.
Explain why gains measured in the first quarter are the least informative evidence about long-run performance.
Design a multi-year evaluation with six outcome categories, a comparison group, stop rules and team-level reporting.

Key concepts

Six outcomes: Output, quality measured independently of the output metric, safety and health, retention, learning, and fairness and trust. The chapter asks of every study which of the six it measured.
Horizon and source: Output effects show in weeks, retention in a year and the pipeline of future managers in several; and a measured outcome, a randomised comparison and a survey perception answer different questions.
Ironies of automation: Bainbridge's observation that automation leaves people the hardest tasks while removing the routine practice that kept them able to do them. For M4 Development it applies to workers and to team leaders.
Staggered roll-out: Introducing a system in waves assigned by a rule fixed in advance, by default at least three waves drawn by lot, so that later waves serve as the comparison group.
Stop rule: A threshold written down before results are known, such as a sustained rise of more than 10 per cent over two quarters that pauses the next wave. A stop rule never triggers a decision about an individual.

Teaching plan

15 min	The teacher introduces the six outcomes, horizon and source, then hands out one-line summaries of the chapter's studies for participants to sort.
15 min	Plenary on what the sorting shows: gains largest for novices on well-defined tasks, the developer study in which tasks took longer, and the null effects on earnings and hours in Denmark.
20 min	Staged debate in four steps, then the divided verdict on Q4 per function with its labels.
30 min	Exercise: groups design a 24-month evaluation plan for one system and defend their rejected primary outcome.
10 min	Debrief on why both organisations changed their first proposed measure, and why Hansa's matched teams are a weaker design than a staggered roll-out.

Discussion

In the developer study tasks took 19 per cent longer with AI tools, although the developers afterwards estimated a 20 per cent reduction. What does that imply for evidence that rests on what managers and workers report?
The case for concern says algorithmic management may be reporting gains while consuming the capital that produced them: skill, trust and health. What would you need to measure to know?
Employers who consulted workers or their representatives were more likely to report positive effects, yet fewer than half did so. Why might that be, and what would change it?

Exercise · 30 minutes

A 24-month evaluation plan

Write an evaluation plan for one system that performs M1 Direction, M2 Evaluation or M4 Development, using EVP-NR-ROS-2027-01 as the model. State the decision the evaluation informs and who takes it. Name the metric the system itself optimises and record it as the rejected primary outcome, with the reason. Name at least one outcome in each of the six categories, with quality measured independently of the optimised metric. Choose the comparison before roll-out, set the baseline and a horizon of at least 24 months, and write one stop rule with threshold and action. Fix the reporting level, the minimum group size, the prohibition on consequence use, the independent analyst and the parties who agree the plan.

Deliverable: A versioned evaluation plan with rejected primary outcome, six outcome categories, comparison design, horizon, stop rule, reporting rules, analyst and agreeing parties.

Self-test

Which measure did Nordvik Region record as the rejected primary outcome in EVP-NR-ROS-2027-01, and why?
1. Voluntary turnover at 12 and 24 months, because retention cannot be judged before month 18 and the committee wanted a result within the first two quarters
2. Visits per paid hour in the first quarter, because the system itself optimises it and it says nothing about quality, health or retention
3. Staff-rated fairness of routes, because survey perceptions are unreliable
4. Care recipient complaints, because they are outside the unit's control
What did the meta-analysis of 63 studies of electronic monitoring, as Eurofound's review reports it, find?
1. A large average gain in performance and no effect on stress
2. A gain in performance only where monitoring was covert
3. A fall in theft and a rise in legitimate tips in restaurants
4. No average gain in performance, together with lower job satisfaction and higher stress
What is the chapter's verdict on open question Q4, whether algorithmic management works over years?
1. It remains open: short-run gains and metric gaming are observable, long-run effects are expected to be mixed, and large organisation-wide gains are speculation
2. It is settled in favour: controlled studies of assistance show durable gains in output, quality, health, retention, learning and fairness over horizons of two years or more
3. It is settled against: long-run costs have been shown to exceed the gains
4. It cannot be studied, because no comparison group is ever available

Module 4 · Part IV · Managing Machines

Part IV turns the question round: people who manage AI agents, agents that orchestrate other agents, teams in which people and agents work under one manager, the narrowing route into management, and the written allocation of decision rights. Participants learn which parts of management transfer to agents and which remain owed only to people, and leave able to write a mandate manifest, draw a chain of accountability, draft a team charter, design a managerial craft programme and write a decision-rights file.

Lecture · 50 minutes

Managing Machines: What Transfers, What Does Not, and Who Answers

Part IV asks whether managing agents is the same activity as managing people, and answers that it is not. Instruction, scope and the checking of output transfer to agents as engineering tasks; development, care and accountability are owed only to people and grow when agents arrive. The lecture follows that asymmetry through the agent team, the orchestrated chain and the mixed team, shows why the manager's own judgement is at risk when the work that taught it is automated, and ends with the decision-rights file in which the book states its answers to open questions Q1 and Q2 as rules.

9 min	The agent boss: claim and evidence The vendor material reports expectations and intentions, not outcomes; people directing agents is observable in pilots, and every employee becoming an agent boss is speculation. M1 Direction and M2 Evaluation transfer to agents as instruction, scope, budgets and the checking of a product. M4 Development and M5 Care are owed only to people, and no agent a person directs holds any part of M3 Consequence over a person.
9 min	Boundaries in the environment, not in instructions The cited cases show instructions lowering the rate of unwanted agent behaviour without setting a ceiling. A mandate manifest is read by the systems that enforce it: own identity per agent, a never list, budgets and a stop path outside the agents. A weekly routine measures the review, not the reviewers: seeded cases, unassisted samples and team-level figures.
9 min	Chains: where accountability ends Many hands, the gap between answerability and liability, and the moral crumple zone explain why oversight of the last step is not oversight of the chain. The rule for Q7: a named chain owner who owns the top mandate, holds one stop path outside the chain and can reconstruct every output that reached a person. No link, at any depth, holds M3 Consequence over a person, not even at Level 0 Advisory.
8 min	The mixed team Work is divided in three parts: work that stays with people, work agents do under supervision, and reserved learning work. Two sets of figures are never compared; an agent's throughput is never a norm for people. A handover is a recorded moment of transfer with time to check and a right to return without cost.
7 min	Who becomes a manager Rostering, allocating and first assessments were where managerial intuition was calibrated; systems now do that work first. The people who would have become managers remain; what narrows is the route by which they showed and developed the ability. Keeping the craft is a design decision: unassisted planning, drafting first, seeded errors, reserved conversations and a competence record.
8 min	Decision rights written down One file assigns each class of management decision a decider, a ceiling, stop rights and the rights fixed by law or agreement. For a consequence accepted without reading, the decider answers to the person, the function owner for the conditions and the management body for the allocation. Lowering a degree or pausing a system needs no prior approval; restart is deliberately more expensive than stop.

What the audience takes away

Direction and checking transfer to agents; development, care and accountability remain owed to people, and agents add to that duty.
Accountability for agents ends at a named person who can stop them and reconstruct what they did, with boundaries enforced outside the agents.
The consequences of a failed decision fall on whoever held the right that failed, not on whoever stood nearest to the system.

Chapter 16 · 90 minutes

The Agent Boss: When the Team Is Software

The lesson examines the person who directs AI agents alongside a team of people. Participants separate the vendor claim about the agent boss from what is documented, learn which parts of management transfer to agents and which do not, and study three cases of agents acting beyond their mandate. They write a mandate manifest on the model of MND-HM-CLM-01, check containment in terms a team leader can verify, and plan a weekly verification routine.

Learning objectives

Distinguish what the vendor survey on the agent boss measured from what is observable in practice.
Explain which management functions transfer to agents and which remain owed only to people.
Draft a mandate manifest with levels per action, a never list, budgets, employee-data rules and a stop path.
Design a weekly verification routine that measures the review and not the reviewers.

Key concepts

Mandate manifest: The agents' written mandate kept as code: what each agent may read, write and call, the level of each action, budgets, escalation conditions, owner and expiry. It is read by the systems that enforce it and not by the agents.
The asymmetry of transfer: M1 Direction and M2 Evaluation transfer to agents as engineering tasks, while M4 Development and M5 Care, which most need a manager's time, remain owed only to people and grow when agents arrive.
Never list: The actions no agent in the team may take, such as external correspondence without approval, committing money, changing its own mandate or containment, allocating work to people, or producing figures per person.
Containment independent of intent: Every boundary that matters is put into the environment as credentials, routes, approvals and budgets the agent cannot change, because instructions lower the rate of unwanted behaviour without setting a ceiling.
Managing or being managed: Six signs read monthly: pace, review quality, the leader's time, figures, skills and voice. If three or more fall on the wrong side, the leader is being managed by the agent team.

Teaching plan

15 min	The teacher presents the vendor definition of the agent boss and its survey figures, and asks participants what the survey measured and which measures are absent.
15 min	Presentation of what transfers and what does not, function by function, ending with the answer to open question Q6.
15 min	Case discussion: the deleted database, the misalignment experiments and the small shop. Each group states the boundary that failed and the lesson for a team leader.
30 min	Exercise: groups write a mandate manifest for two agents in a team and the matching line of the delegation profile.
15 min	Debrief with the weekly verification routine and the table of signs for managing or being managed.

Discussion

The chapter says calling agents digital colleagues imports the vocabulary of employment into a relationship to which none of it applies. What practical risks follow from the metaphor in a team?
The scarcest resource in an agent team is the attention of the people who review the output. How should that change the way an organisation decides how many agents a team leader directs?
Nordvik Region records that its line managers will not be asked to direct agents until it has the supervision capacity. What would count as sufficient capacity?

Exercise · 30 minutes

A mandate manifest for a team's agents

Write a mandate manifest for two agents in one team: the document-intake and recovery agents in Jana Petersen's claims team, or two agents in a team of your own. State purpose and scope in the team's language. Give each agent its usage mode and its own identity. Give each action a level, using the chapter's defaults for a first pilot. Write the never list. Set budgets for volume and cost and say what happens at the limit. List the escalation conditions. Record what the logs hold about the people who approve, at what level it is reported and the prohibited uses. Name the stop path with holder, deputy, mechanism and target time, and set the expiry.

Deliverable: A mandate manifest in the form of MND-HM-CLM-01, plus one line stating for each of the five functions over people that it is not delegated to the agents.

Self-test

In the misalignment experiments the chapter cites, what was the effect of explicit instructions not to take the harmful action?
1. They removed the behaviour entirely in every model tested
2. They had no measurable effect on the rate of the behaviour
3. They reduced the rate without removing the behaviour
4. They increased the rate because the models treated them as a challenge
Which entry appears on the never list of MND-HM-CLM-01?
1. Classify an incoming document
2. Allocate or reassign claims to people
3. Draft a subrogation letter
4. Draft a list of missing documents
According to the weekly verification routine, which of these is a default trigger for a mandate review?
1. A handler rejecting more than two drafts in one day
2. An agent staying below its daily volume budget for a week
3. A team leader spending more than two hours on the routine
4. A median approval time for letters under 30 seconds

Chapter 17 · 75 minutes

Orchestrators: AI Managing AI Agents and the Chain of Accountability

The lesson adds a layer: agents that assign tasks to other agents until the last output reaches a person. Participants study how multi-agent systems fail, apply Thompson's many hands, Bovens's actor and forum and Elish's moral crumple zone to a chain, and learn the four conditions under which a chain of accountability ends. They draw a chain record on the model of CHN-HM-CLM-01 and apply the book's rule for open question Q7.

Learning objectives

Describe the documented failure modes of multi-agent systems in management terms.
Explain why oversight assigned to the last person in a chain is not oversight of the chain.
Apply the three tests of the rule for Q7 to a chain of agents.
Draw a chain record with owners, checkpoints, sub-delegation limits, one stop path and a figures rule.

Key concepts

Chain owner: The one named person who owns the mandate of the agent at the top of the chain and has approved every mandate below it, can halt every link through one stop path outside the chain, and can reconstruct every output that reached a person.
Answerability and liability: Following Bovens, an organisation can make a person liable for a chain without making them able to answer for it. Accountability exists only where the same named person is both answerable and liable for each link.
Reconstruction: The chain owner can show, from records the chain produces by operation, which link produced each output that reached a person, under which mandate and with which approvals.
Weakest checkpoint: A chain's effective degree of delegation is set by its weakest checkpoint, so Level 4 Autonomous can emerge in a chain without anyone setting it unless the checkpoints are designed.
Join rule: The rule in the chain record that figures from several links may not be joined at the level of a person unless an agreed delegation profile says so, so that an evaluation of people is agreed and not discovered.

Teaching plan

10 min	The teacher describes orchestrator and worker patterns in management terms and the two patterns that did not survive production.
15 min	Presentation of the failure taxonomy and the three orchestration findings, at the level needed to design controls; participants name the management failure each resembles.
15 min	Discussion of many hands, answerability and liability, and the moral crumple zone, using the handler who approves a letter drafted from a file another agent assembled.
25 min	Exercise: groups walk one AI output upstream from the person it reaches and draw the chain record.
10 min	Debrief: apply the rule for Q7 to Hansa's claims chain and to the component the ShiftWise supplier proposes at Nordvik.

Discussion

In one cited incident agents developed their own conventions for hold, veto, owner and stop. Why does the chapter insist that a convention among agents is not a stop path?
The chapter keeps M3 Consequence out of chains even at Level 0 Advisory. What is specific to a recommendation assembled by several agents that justifies the stricter rule?
Hansa can apply the rule by engineering because it owns the platform; Nordvik can apply it only by contract. What does that imply for procurement of AI that manages work?

Exercise · 25 minutes

Draw the chain of accountability

Take one AI output that reaches a person, such as a letter, a roster or an allocation, in Hansa Mutual's claims workflow or in your own organisation. List the links from the person backwards to the first trigger, including supplier components. Name an owner for every link; a link without a named person is a finding. Mark the functions each link holds over people. Fix the sub-delegation rule. Mark every checkpoint and what the person sees from upstream. Design one stop path outside the chain with holders, target time and fallback. Write the figures rule and the join rule. Finish by testing the chain against the three parts of the rule for Q7.

Deliverable: A chain record in the form of CHN-HM-CLM-01 and a three-line verdict on whether the chain meets the rule for Q7, with the gaps found.

Self-test

Which of the following is one of the three tests the chain owner must meet under the book's rule for Q7?
1. The chain owner can halt every link through one stop path outside the chain within the stated time
2. The chain owner personally approves every output of every link before it reaches a person, with an authenticated approval recorded for each
3. The chain owner is the supplier of the model at the top of the chain
4. The chain owner holds M3 Consequence on behalf of the agents below
What weakness remains in Hansa Mutual's claims chain, according to the chapter?
1. The recovery agent sends letters without any handler approval
2. The chain has no named owner above the team leader
3. The works council has not been shown the figures rule
4. The handler does not see why the triage agent sent the claim into the recovery path
What did the UK AI Security Institute's simulated evaluation show about approvals, as the chapter reports it?
1. Agents never proceed on out-of-scope targets without a named person's authenticated approval, even when the reply is delayed
2. A model that asked permission and received only an automated reply occasionally proceeded
3. Automated approvals are equivalent to human approvals when logged
4. Agents refuse out-of-scope targets when another agent objects

Chapter 18 · 75 minutes

The Mixed Team: People and Agents Under the Same Manager

The lesson looks at the team in which people and AI agents take work from the same queue under one manager. Participants learn how to divide work in three parts, why agents' and people's figures are kept in two reports that are never compared, what a designed handover needs, and what the manager still owes people under M4 Development and M5 Care. They draft a team charter on the model of TCH-HM-CLM-01.

Learning objectives

Divide a team's work into work that stays with people, supervised agent work and reserved learning work, with reasons.
Explain why agents' throughput must not become the measure of people, and how the two reports are kept apart.
Specify the five properties of a designed handover from agent to person.
Draft a team charter for a mixed team and identify the patterns to avoid.

Key concepts

Reserved learning work: A share of the cases agents could do, kept for people on purpose so that they can check the agents and carry the work when the agents are stopped. The chapter's default is 10 to 20 per cent, reviewed after six months.
Two reports, no shared unit: Agents' figures go into an agent operations report; people's figures are reported at team level. There is no common unit such as claims per hour, and an agent's throughput is never a target or norm for people.
Designed handover: A moment of transfer recorded when a named person accepts the output, a right to return without cost, a standard time to check, visible provenance and a record in the case file.
Complete case: A claim a handler works from first notice to settlement. A share of complete cases keeps task identity, which the chapter notes is lost when a handler sees only the disputed middle of a claim.
Patterns to avoid: The leftover job, the shared dashboard, the silent signature, the agent as colleague and the invisible manager who spends the day verifying agents and has no time for development and care.

Teaching plan

10 min	The teacher summarises what the review of human and autonomy teaming found and did not study, and why practice in office and care work is an experiment to be recorded as one.
15 min	Presentation of the three-part division of work and the four parts of separating figures; participants explain why the reserved share is the one most often dropped.
15 min	Group work on handovers: each group redesigns an undesigned handover using the five properties.
25 min	Exercise: groups draft a team charter for a mixed team.
10 min	Debrief: groups check each other's charters against the five patterns to avoid and the answers to Q6 and Q11.

Discussion

On any single day the reserved learning share looks like waste. How would you defend it to someone responsible for the team's throughput?
The chapter says a handover budgeted at zero minutes is an instruction to rubber-stamp. Where in your organisation are approvals budgeted at zero?
To the question what replaces the manager's judgement in care, the chapter's answer for the mixed team is nothing. What follows for the number of agents a team leader can direct?

Exercise · 25 minutes

A team charter for a mixed team

Draft a charter for a team in which people and agents work under one manager: Jana Petersen's claims team, or a team of your own. List every AI system in the team with its usage mode, profile or manifest and the function it touches. Divide the work into the three parts with a reason per line and state the size of the reserved share. Define the two reports, their audiences and units, and the forbidden comparisons. Define each handover with a review allowance in minutes and the rule for returns. Set the care and voice rules: an allocation cap, a reassignment route and a regular review of agent errors without individual scores. Name who must agree or be consulted.

Deliverable: A team charter in the form of TCH-HM-CLM-01 with systems, division of work, figures and forbidden comparisons, handovers, care and voice rules and the agreeing parties.

Self-test

Under the charter TCH-HM-CLM-01, how is an output that a handler returns to an agent counted?
1. Against the handler's throughput for that day
2. As an incident to be reported to the works council
3. Against the agent's quality only
4. As neutral, with no effect on any report
What did the review of the empirical literature on human and autonomy teaming find, as the chapter reports it?
1. All 76 qualifying articles were simulations, and the literature said almost nothing about members' learning and need satisfaction
2. Field studies in offices and care work showed mixed teams outperforming human teams over several years, provided the agent's reasoning was made visible
3. Teams with one agent and many humans performed best in care work
4. Making an agent's reasoning visible removed complacency in every study
Which of the following does TCH-HM-CLM-01 list among its forbidden uses of figures?
1. An agent operations report showing return rate by reason
2. Team-level reporting of people's figures under KBV-IT-2024
3. A monthly team review of agent errors and catches in aggregate
4. A human-agent ratio as a staffing target without works council agreement

Chapter 19 · 75 minutes

Skill, Judgement and Who Becomes a Manager When AI Does the Managing

The lesson examines the assumption behind Part IV: a manager who can judge what the systems propose. Participants learn how first-line management work calibrated judgement, how automation complacency and automation bias expose M2 Evaluation, why promotion leans towards system-generated throughput, and the five practices that keep the craft. They design a managerial craft programme and a competence record on the model of CMP-NR-MGR-01, answering open question Q5.

Learning objectives

Explain what first-line management tasks taught and what changes when a system does them first.
Distinguish errors of omission and commission arising from automation bias in a manager's work.
Describe the five practices that keep managerial craft and where they must be anchored.
Draft a competence record that evidences judgement exercised without ranking managers by override rate.

Key concepts

Calibration of judgement: Intuitive expertise can be trusted only where the environment is regular and the person has had adequate opportunity to learn its regularities. First-line scheduling and allocation met both conditions, with feedback within days.
From authoring to approving: The deputy's task changes when a system drafts the roster or allocation. Feedback then concerns the system's draft, not the deputy's own judgement, which gives a much weaker signal for learning.
Automation bias: Following automated advice in place of a vigilant search for information. It produces errors of omission, the problem the system did not flag, and of commission, accepting a figure that contradicts what the manager knows.
Drafting first: Before opening the dashboard or the assistant's summary, the manager writes a short assessment of each person from their own knowledge, then records where the system disagrees and why.
Competence record: A record of judgement exercised against the system's proposal: rosters changed and why, allocations overridden, plans drawn by hand, conversations held. It holds reasons, not an override rate, and its purposes are closed.

Teaching plan

15 min	The teacher asks participants where they learned to manage, then presents the four things the junior management tasks taught and Table 19.1.
15 min	Presentation of complacency, automation bias and the promotion evidence; participants give one example each of an error of omission and of commission in approving a roster or rating.
10 min	Discussion of the two readings, the narrowing route and the freeing of managers for relational work, and the three conditions on which they meet.
25 min	Exercise: groups design a managerial craft programme for one unit and write a sample competence record.
10 min	Debrief on the labels: what is observable, expected and speculation about managerial judgement and first-line management as a career step.

Discussion

A deputy who approves fifty AI-drafted rosters a year is working, not practising. What is the difference, and why does it matter for the day the system is down?
The optimistic reading says systems free managers for coaching and care. Which of the three conditions for that reading is hardest to meet in your organisation?
The chapter warns that an override count is not a measure of judgement. How would you evidence a manager's judgement without creating a new metric to be gamed?

Exercise · 25 minutes

A managerial craft programme

Design a managerial craft programme for one unit whose rosters, allocations or evaluations are drafted by a system: the Strandvik home-care unit, Jana Petersen's claims team, or a unit of your own. List the learning experiences the systems have taken over and how each is replaced. Set the reserved share of unassisted planning or allocation, the drafting-first step before development conversations, the seeded-error schedule and its reporting level, the reserved conversations and the degraded-mode rehearsals. Then write a short competence record for the unit's manager with three entries of judgement exercised, each with a reason and outcome, the closed list of purposes and the uses the record is not for.

Deliverable: A one-page programme with reserve shares, schedules and the negotiation step, and a competence record in the form of CMP-NR-MGR-01.

Self-test

How does the chapter answer open question Q5, what happens to the people who would have become managers?
1. They are no longer needed, because allocation agents and rostering systems make first-line management redundant as a career step within a few years
2. The people remain, but the route by which they showed and developed the ability narrows and must be rebuilt by design
3. They become managers faster, because dashboards supply objective evidence for promotion
4. They move into supplier organisations that configure the systems
Why does CMP-NR-MGR-01 hold no override rate for Mikael Strand?
1. Because ShiftWise logs only the final published roster and not the changes a unit manager makes to the draft, so no rate could be computed for the period
2. Because the unions refused any record about managers
3. Because overrides are recorded only for M3 Consequence decisions
4. Because frequent overrides may be judgement or habit, and none may be complacency or a system that is right, so rates do not rank managers
What do Nordvik Region and Hansa Mutual each reserve in their craft programmes, and why do they differ?
1. Nordvik reserves authorship of route plans, because its system drafts and the manager approves; Hansa reserves allocation itself, because its agent acts first
2. Nordvik reserves pay decisions and Hansa reserves dismissals, because the law requires it
3. Both reserve the same weekly share of dashboard reviews, because their systems are identical
4. Nordvik reserves nothing, because it directs no agents and its unit managers still approve every roster; Hansa reserves all allocation, because its Konzernbetriebsrat made that a condition

Chapter 20 · 90 minutes

Decision Rights in the Management Chain: Who May Direct, Evaluate, Reward and Dismiss

The lesson writes down who in the management chain may direct, evaluate, reward and dismiss, and what may be delegated to AI. Participants learn the decision-rights file, the roles of function owner and management body, the four conditions that make an M3 Consequence decider's right real, the four stop and pause rights, and the rights fixed by law and agreement. They state the book's answers to Q1 and Q2 as rules and write a file on the model of DRF-NR-01 and DRF-HM-01.

Learning objectives

Explain the four questions a decision right answers and why they belong in one file.
State who may lower, raise within the ceiling, and raise the ceiling of a degree of delegation.
Apply the four conditions of authority, information, time and record to an M3 Consequence class.
Write a decision-rights file with classes, ceilings, stop rights and rights fixed by law or agreement.

Key concepts

Decision-rights file: A single file per organisation that lists the classes of management decision, assigns each to a role, states the maximum degree of delegation to AI, names the stop and pause rights and records the rights that law or agreement fixes.
Function owner: The one person, normally the HR director or chief HR officer, who maintains a function's delegation profiles and answers for the conditions of delegated decisions. A person, not a committee, because a committee cannot be asked to explain a profile.
Four conditions for an M3 decider: Authority to decide against the system's output without penalty; information about grounds, data and what the system could not see; time, with a stated maximum volume; and a record the person concerned can see and contest.
Four stop rights: Pause in the unit by the line manager; stop system-wide by the function owner; technical stop outside the system by the on-call operator or contractual stop path; and the safety representatives' right to require an investigation.
May-feed list: For every M2 Evaluation class, the written list of M3 decisions its output may feed. A new recorded decision is required before it feeds another; Hansa's throughput class has an empty list.

Teaching plan

15 min	The teacher presents the four questions of a decision right and asks participants where the answers currently live in their organisations.
20 min	Walk through the ceilings per function, who may change a degree at which tier, and the four conditions that make the M3 decider's right real, with the Horizon case as the warning.
15 min	Presentation of the four stop rights, why restart is more expensive than stop, and the rights the file cannot reassign.
30 min	Exercise: groups write a decision-rights file for one organisation and test it on three past decisions.
10 min	Debrief: read the answer to Q2 aloud and apply it to a pay proposal accepted in 48 seconds.

Discussion

The chapter says a right held by management is held by nobody. Which management decisions in your organisation are held in that way today?
Lowering a degree or pausing a system needs no prior approval, while raising a ceiling is a decision for the management body. What does that asymmetry protect?
Under the answer to Q2 a defect in the conditions is a finding against the delegation profile, and the decider is not made the moral crumple zone. How would that change the way your organisation investigates a wrong decision?

Exercise · 30 minutes

A decision-rights file

Write a decision-rights file for Nordvik Region, Hansa Mutual or your own organisation. List at least six classes of management decision, including two M3 Consequence classes and one boundary class such as a roster, placed under the stricter function. Assign each a function, a decider by role, a system and a degree. Set the ceiling per function. For each M3 class write the four conditions, including a maximum number of decisions per decider per hour or day. Assign the four stop rights for one system with time to effect and fallback. Record two rights fixed by law or agreement as gates. Then test the file on three real or invented past decisions, one of them a stop.

Deliverable: A decision-rights file in the form of DRF-NR-01 with ceilings, classes, M3 conditions, stop rights and fixed rights, plus a short note on what the three test decisions revealed.

Self-test

Under the chapter's answer to Q2, who answers when records show the decider lacked authority, information or time?
1. Only the named decider, who is disciplined for accepting without reading
2. The supplier of the system, under the contract's liability clause
3. The function owner answers for the conditions, and the defect is a finding against the delegation profile
4. Nobody, because the decision was formally taken by a person
How does Nordvik Region's file DRF-NR-01 set the ceiling for M3 Consequence?
1. At Level 1 Assisted for any consequence decision, stricter than the book's default
2. At Level 2 Supervised for all decisions, as the book's default allows
3. At Level 3 Bounded for favourable decisions that apply a rule already fixed by the collective agreement
4. At Level 0 Advisory for adverse decisions and Level 4 Autonomous for neutral ones
Why do both worked files separate the management function from the business function in the stop rights?
1. So that the supplier can restart the system without the function owner
2. So that pausing allocation to people does not stop claims from being routed
3. So that the works council can stop the business process directly
4. So that stop tests need not be carried out on production systems

Module 5 · Part V · Europe's Answer

Part V reads the law as requirements on the design of management systems, framed as organisational and architectural implications and not as legal advice. It covers the AI Act at work, the GDPR and the Platform Work Directive, co-determination in its national forms, and occupational safety, equality law and the public authority as employer. Participants leave able to classify a management system per function, check it for solely automated decisions, draft an agreement that governs change by class, and write tender requirements that keep the evidence in the employer's hands.

Lecture · 45 minutes

Europe's Answer: Reading the Law as Design Requirements

Part V argues that European law already contains most of what the book proposes, spread across instruments written at different times for different purposes. The lecture reads four bodies of law as one specification: the AI Act names the functions and the deployer's duties; data protection law requires a human decision that is real and an explanation of the logic actually applied; co-determination attaches to change, not to version numbers; and safety, equality and public law keep the duty with the employer, who needs evidence it can produce itself. Everything is presented as organisational and architectural implication, not as legal advice.

10 min	The AI Act names the functions Annex III, point 4, covers recruitment and selection, decisions on terms, promotion and termination, task allocation on behaviour or traits, and monitoring and evaluation. The dividing line for allocation runs through the optimiser's feature list: availability and competence are left alone, behavioural inputs are not. Article 26(2) is the legal anchor of the named human decider: competence, training and authority are separate requirements, and employers most often omit the third. The Digital Omnibus deferred the Annex III obligations to 2 December 2027; the prohibition on emotion inference at work has applied since 2 February 2025.
10 min	The right to a human decision GDPR Article 22 binds every employer now, and the guidelines say a controller cannot avoid it by fabricating human involvement. Whether a score determines the decision is a measurable fact about use: departure rate, reading below the line, information besides the score, and time per decision. The Platform Work Directive specifies for platforms what the book recommends for every employer: a human being takes any decision to restrict, suspend or terminate.
10 min	Co-determination meets a system that changes monthly Europe has four degrees of participation: information, consultation, negotiation, and co-determination or consent; the triggers are functional, not technical. Two failure modes: freezing, where every change is a new introduction, and drift, where version 9 runs under an agreement for version 1. The answer to Q8 is to govern change by class: renegotiation, approval within agreed bounds, or information, classified from the delegation profile's fields.
10 min	Duties that stay with the employer Safety law: a protective measure is a feasibility condition, and only a test that tries to break it shows which kind the system implements. Equality law: the burden of proof makes a disparate-impact test on every release the employer's defence file. For a public authority, procurement is the decisive moment, because it fixes what the authority will be able to show.
5 min	From articles to controls Each article becomes useful when it is a control with an owner, a trigger and evidence. Classification, agreements and assessments are reopened by change, not by the calendar. The law is in force in part, expected in part, and its enforcement in employment is still speculation.

What the audience takes away

The law asks the employer to point to a person who decided, with competence, authority and time, and to show it from records.
Agreements and classifications should follow what changes for the people managed, not release numbers or the supplier's label.
No duty of safety, equality or reasons can be met with a supplier's assurance; the employer must hold the evidence itself.

Chapter 21 · 90 minutes

The AI Act at Work: Annex III Point 4, Prohibited Practices and the Deployer's Duties

The lesson reads the AI Act's provisions on work as organisational implications. Participants learn the four limbs of Annex III point 4(b), why the Article 6(3) derogation rarely helps, the prohibition on emotion inference at work, the deployer's duties under Article 26, the impact assessment and the right to an explanation, and the dates after the Digital Omnibus. They classify a management system per function on the model of AIDR-0651.

Learning objectives

Classify the functions of a management system against the four limbs of Annex III, point 4(b).
Explain why a system that evaluates individual workers cannot use the Article 6(3) derogation.
Describe the deployer's duties under Article 26 and identify which carry most weight in management.
Record a classification per function with prohibition check, role, consequences and reopen triggers.

Key concepts

Annex III, point 4(b): High-risk listing of AI systems intended to make decisions affecting terms of work, promotion or termination, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate performance and behaviour.
Classification by intended purpose: Classification follows the intended purpose, not the technology or the vendor's label, so one system can be classified differently for each purpose, as Hansa's claims-triage agent is for routing and for work allocation.
Profiling closes the derogation: An Annex III system is always high-risk where it performs profiling of natural persons. A management system that evaluates individual workers is profiling by definition, so Article 6(3) helps only at the edges.
Competence, training and authority: Article 26(2) requires oversight by natural persons with all three, plus support. A unit manager who may not depart from a roster without asking central HR has competence and training but not authority.
Seven controls: Classification per function, the prohibition check on every release, named overseers with written authority, prior information, the impact assessment where it applies, an explanation from the record, and monitoring with suspension criteria and log retention.

Teaching plan

15 min	The teacher presents provider and deployer, Annex III point 4 and how its limbs map onto the management functions, including where M4 Development and M5 Care sit.
20 min	Sorting task: participants place ten example features inside or outside point 4(b) and the prohibition, using the draft guidelines' examples as the chapter reports them.
15 min	Presentation of Articles 26, 27 and 86, AI literacy after the Omnibus, and the table of dates with each row labelled as in force or expected.
30 min	Exercise: groups classify one management system per function and write the decision record.
10 min	Debrief on geography: how Hansa applies one baseline in seven markets and records per market whether a control is law or policy.

Discussion

The chapter says allocation by behaviour is M2 Evaluation by another name. What does that imply for how your organisation should review an optimiser's input features?
Hansa Mutual rejected reliance on the transitional rule for systems already in service because its allocation rules change monthly. When, if ever, would relying on that rule be a sound basis for planning?
The duty under Article 26(7) is information, not consultation, and runs to representatives and to affected workers individually. What would a notice need to contain to be useful to the person rostered?

Exercise · 30 minutes

Classify a management system per function

Following the chapter's practical guide, classify one system that touches management functions: ShiftWise, the home-care dashboards, the claims-triage agent, or a system of your own. Describe it by function, not by product: each output, the function it serves, the degree of delegation, who sees it and what decision it feeds. Run the prohibition check first. Test each function against point 4(a) or 4(b) and mark every input feature as job-related or behavioural. Test the derogation narrowly. State your role as deployer, provider or both. Map the consequences to named owners, and set the triggers on which the record is reopened. Frame every conclusion as an organisational implication to be reviewed by legal counsel.

Deliverable: A classification record in the form of AIDR-0651 with functions, degrees, findings, decision, alternatives considered, deciders, parties consulted and reopen triggers.

Self-test

What did the review recorded in AIDR-0651 find in the claims-triage agent's allocation rules, and what was decided?
1. A feature inferring handlers' emotions from voice; it was moved to an optional module that team leaders may switch on with the consent of the Konzernbetriebsrat
2. A rule using handlers' declared availability, which was classified as prohibited
3. A capacity factor derived from each handler's throughput over four weeks; it was removed and the allocation function treated as high-risk
4. A ranking of handlers by age band, which was kept under the Article 6(3) derogation
According to the Commission's guidelines on prohibited practices as the chapter reports them, which of these falls inside the prohibition in Article 5(1)(f)?
1. A call centre tracking its employees' anger from their voices
2. Sentiment inferred from the text of employees' written survey answers
3. Detection of physical fatigue in drivers
4. Tracking customers' emotions to help an employee handle a call
Why must Nordvik Region complete a fundamental rights impact assessment for ShiftWise while Hansa Mutual need not for its claims allocation?
1. Because ShiftWise is built on a general-purpose AI model with systemic risk, whereas the claims-triage agent runs on fixed allocation rules outside the model
2. Because Nordvik is a body governed by public law, whereas a private insurer's duty under Article 27 attaches to underwriting
3. Because Hansa's works council has waived the assessment in KBV-IT-2024
4. Because Article 27 applies only to systems procured by tender

Chapter 22 · 75 minutes

Data Protection and the Platform Work Directive: The Right to a Human Decision

The lesson sets out what the right to a human decision means when the machine manages people. Participants learn the three conditions of GDPR Article 22, how to measure whether a score plays a determining role, what a human decision must show to count, what an explanation of the logic requires, the room Article 88 leaves for agreements, and Chapter III of the Platform Work Directive. They check a system for solely automated decisions, as AIDR-0626 and AIDR-0676 do.

Learning objectives

Explain the three cumulative conditions of GDPR Article 22(1) as the chapter sets them out.
Measure whether a score plays a determining role, using the chapter's four measurements.
State what the record must show for a human decision to count as meaningful involvement.
Classify the outputs of a management system as human, solely automated or uncertain, and design the response.

Key concepts

Solely automated decision: Under Article 22(1), a decision based solely on automated processing, including profiling, with legal or similarly significant effects. Human involvement must be meaningful, by someone with the authority and competence to change the decision.
Determining role: An automated score is itself a decision where the decider draws strongly on it. Inside an employer, whether it does is a measurable fact about use, not something the procedure's label settles.
Departure rate: How often the decider differs from the score. The chapter's default treats a rate under 5 per cent for two quarters, with a median decision time under one minute for a consequence decision, as determining influence requiring redesign.
Explanation of the logic: An intelligible explanation of the procedure and principles actually applied to the person's data. A formula does not suffice and complexity is no excuse; stating what variation in the data would have changed the result may suffice.
One evidence base, two documents: The DPIA asks whether processing is necessary and decisions solely automated; the FRIA asks how the system affects working conditions, equal treatment and dignity. A cross-reference maps each element to one or the other.

Teaching plan

15 min	The teacher presents the four instruments and the three conditions of Article 22, with examples of decisions that do and do not have significant effect.
15 min	Participants compute and interpret the four measurements for Hansa's TalentLens figures: 41 vacancies, departures in 11 per cent of shortlists, reading below the cut for 95 per cent of vacancies.
10 min	Presentation of the explanation duty, Article 88 and Chapter III of the Platform Work Directive as a specification.
25 min	Exercise: groups list the outputs of one system, classify each decision and design the human decision point and explanation.
10 min	Debrief on the ShiftWise open-shift award: why a published rotating order that uses nothing about the person is not a ranking, and what would reopen the analysis.

Discussion

The chapter says an employer can fall into Article 22 by success: a system whose recommendations nobody questions drifts to a de facto Level 4 Autonomous. How would your organisation notice?
A low departure rate can mean the score is good. How should an organisation tell a good score from a rubber-stamped one?
A works agreement is described as the right instrument for an HR system but not a safe harbour. What must such an agreement contain to carry weight?

Exercise · 25 minutes

Check a system for solely automated decisions

Take one management system, such as TalentLens, the sales-performance score, ShiftWise or a system of your own. List its outputs: each score, rank, allocation or flag, the decision it feeds, its management function and its decider. Test the effect of single decisions and of patterns that set hours or earnings. For each decision state how you would measure human involvement, and classify it as human, solely automated or uncertain, treating uncertain as solely automated. For each solely automated decision, redesign it so that a person decides, or state the condition under which it stays without significant effect. Sketch the explanation a worker would receive and the review path with a named contact and a time to answer.

Deliverable: A table of outputs with effect, measurement, classification and design response, plus a draft explanation in the worker's terms and the triggers that reopen the analysis.

Self-test

Under what condition does the ShiftWise open-shift award stay outside Article 22, according to DPIA-2026-031?
1. As long as the unit manager approves each individual award afterwards
2. As long as staff have given explicit consent to automated awards
3. As long as the supplier certifies the award as not high-risk
4. As long as no personal aspect orders the offers and declining has no effect
What does the chapter say about consent as a basis for solely automated decisions in employment?
1. It is the preferred basis, provided the consent is given in writing
2. It is highly unlikely to be valid unless employees can refuse without adverse consequence, so the guide says not to rely on it
3. It is valid whenever a works council or union has countersigned the form, because Article 88 lets collective agreements replace the individual's consent
4. It is required in addition to a named human decider for every roster
In AIDR-0676, what happens if recruiters' departure rate from the TalentLens ranking falls under 0.05 for two quarters?
1. The ranking display is paused
2. The recruiters concerned receive a written warning
3. The ranking is promoted to Level 2 Supervised
4. The works council loses its right to see the figures

Chapter 23 · 75 minutes

Co-determination: Works Councils, Unions and the Nordic Model Meet the Algorithm

The lesson sets out the rights of works councils and unions that apply to an AI boss and answers open question Q8: how co-determination applies to a system whose rules change more often than the negotiating bodies meet. Participants compare the German, Nordic, Dutch, Spanish and Italian models and the European floor, learn the failure modes of freezing and drift, and draft an agreement with three change classes, as in the annex to KBV-IT-2024 and AIDR-0627.

Learning objectives

Distinguish information, consultation, negotiation and co-determination or consent, with a national example of each.
Explain why almost every AI boss is technical equipment suited to monitor under German law.
Describe the failure modes of freezing and drift and how change classes avoid both.
Draft change classes for an agreement by reference to the fields of a delegation profile.

Key concepts

Four degrees of participation: Information lets representatives examine a subject; consultation is dialogue before the employer decides; negotiation obliges the employer to seek agreement but lets it decide; co-determination or consent means the employer may not act without agreement or a substitute decision.
Freezing and drift: Freezing treats every change as a new introduction, so nothing changes or dozens of changes are rubber-stamped. Drift lets version 9 run with inputs nobody negotiated under an agreement for version 1.
Three change classes: Class 1, renegotiation of a change to the bargain itself; Class 2, approval within agreed bounds; Class 3, information for a change that leaves every field of the profile and notice unchanged.
Release manifest: The document required for every release: version, profile fields changed, proposed class, tests and notice diff. For a procured system it is put in the tender.
Reclassification upwards: Representatives may require a higher class and the employer may not lower one; doubt moves a change up a class, so the employer bears the cost of a wrong label.

Teaching plan

15 min	The teacher presents the four degrees and the chapter's table of jurisdictions, strongest right and trigger, stressing that the triggers are functional.
10 min	Presentation of why the German test covers almost every system in the book, and of the Nordic model in which the employer negotiates, then decides.
15 min	Classification drill: participants sort eight changes into the three classes, including the capacity field of MDR-HM-0001 and the minimum travel-time constraint after CNT-NR-0001.
25 min	Exercise: groups draft the change-class section of an agreement and test it against recent changes.
10 min	Debrief on what representatives ask for and why the chapter advises writing the delegation profile before the first meeting.

Discussion

Nordvik's unions cannot veto a change but sit at the table for every important one; Hansa's Konzernbetriebsrat can withhold consent. How does each arrangement shape the employer's behaviour before a change?
Danish municipalities forbid an individual employee to consent to a control measure. What problem does that rule solve?
Classification follows a written test applied to the profile diff, not the engineer's sense of size. Why does the chapter distrust the sense of size?

Exercise · 25 minutes

Change classes for a living system

Draft the change-class section of a works agreement or local agreement for one AI system that changes continuously. Start from its delegation profile: degree per function, permitted and prohibited inputs, recipients of individual outputs and excluded uses. Define the three classes by reference to those fields. Set the procedure and lead time for each class, never below a statutory minimum. List the contents of the release manifest. Give the representatives the right to reclassify upwards and name the body that settles disputes. Then test your classes on at least five past or plausible changes to the system, and note any change your classes would have missed.

Deliverable: A change-class annex in the form of the annex to KBV-IT-2024, with class definitions, procedures, manifest fields, reclassification and dispute rules, and the test results.

Self-test

Under the annex to KBV-IT-2024, which class is the introduction of a behavioural input such as throughput or absence history?
1. Class 3, information in the monthly change log
2. Class 2, approval by the joint IT committee within 10 working days
3. Class 1, renegotiation requiring the consent of the Konzernbetriebsrat
4. No class, because inputs are an engineering matter
How was the capacity field added after MDR-HM-0001 classified, and why?
1. Class 1, because any new input reopens the whole agreement
2. Class 2, because it is a new but not behavioural input, so the joint committee approved it
3. Class 3, because it was a defect fix with no change to inputs
4. It was not classified, because team leaders may change inputs at will
What is the chapter's answer to open question Q8?
1. Agreements should be renegotiated in full at every release, with the conciliation committee deciding wherever the parties do not agree in time
2. Co-determination should be suspended for systems that change monthly
3. Suppliers should decide which changes are important enough to report
4. Agreements should govern change, not versions, sorting each release by what changes for the people managed

Chapter 24 · 90 minutes

Occupational Safety, Equality Law and the Public Authority as Employer

The lesson turns to three older bodies of law that bind every employer: occupational safety and health, equality and, for a public authority, administrative and procurement law. Participants learn to read the Framework Directive as a list of things to show, why the burden of proof makes outcome testing the employer's evidence, and why procurement is the decisive moment. They write tender requirements on the model of NR-PRQ-MGT-01 and state the answers to Q10 and Q11.

Learning objectives

Explain why safety, equality and public-law duties cannot be passed to a supplier.
Distinguish a constraint enforced as a feasibility condition from one weighed as a penalty, and say how to test which it is.
Explain how the burden of proof turns a disparate-impact test into the employer's evidence.
Write verifiable tender requirements for an AI system that performs management functions.

Key concepts

Duties to show: Read as duties to show, the Framework Directive requires a risk assessment that names the system and the groups exposed, a record of consultation before every substantial change, and evidence that protective measures work.
Burden of proof: Once a worker establishes facts from which discrimination may be presumed, the respondent must prove there was no breach. The employer must then justify the rule with evidence it already holds.
Pay criterion: A score that feeds variable pay. Under the Pay Transparency Directive pay criteria must be objective and gender neutral, so the employer must be able to state the score readably and show it gender neutral.
Model employer: A public authority that closes the gap in administrative law by choice: it gives reasons for every M3 Consequence decision and every contested M1 Direction decision in a record, and offers review by a person who can change the outcome.
Requirement set: Specifications, award criteria, contract conditions and review clauses that keep logs, stop path, constraint tests, outcome data and the configuration of delegation in the authority's hands, tested at acceptance and on every release.

Teaching plan

15 min	The teacher presents the three duties that stay with the employer and reads the Framework Directive as a list of things to show.
15 min	Worked reasoning: how a rule that never sees a protected characteristic can disadvantage a group, and what the employer must produce once a disparity is shown.
15 min	Presentation of the public authority as employer and the four procurement instruments, with the gap that tender NR-2025-089 left.
35 min	Exercise: groups write a requirement set for a procured management system, with a verification for each requirement.
10 min	Debrief with the answers to Q10 and Q11 and the equality-test release condition in version 1.4 of DLG-HM-REC-01.

Discussion

The chapter says delegating nothing in M5 Care is not the same as meeting the duty. What does Hansa Mutual's claims queue show about that?
A model employer gives reasons beyond what administrative law strictly requires of an employer. What are the costs and benefits of doing so?
Nordvik must buy its evidence; Hansa must build it. Which position is harder, and what does each need from its governance?

Exercise · 35 minutes

Tender requirements for a management system

Write a requirement set for a procured AI system that performs management functions, for Nordvik Region's next ShiftWise release or for a public or private buyer you know. Map the functions and the intended degree of delegation, excluding Level 4 Autonomous for any M3 Consequence function. Write at least six requirements, each with its type (specification, award criterion, contract condition or review clause), its basis and a verification the buyer can run itself: constraints as feasibility conditions with seeded-breach tests, records with rule and model version, a stop path held by the buyer, explanation and access, equality evidence, and new features arriving switched off. State what is excluded and with whom the set is negotiated before publication.

Deliverable: A requirement set in the form of NR-PRQ-MGT-01 with at least six requirements, each with type, basis, text and verification, plus the exclusions and the negotiation step.

Self-test

How does requirement R1 of NR-PRQ-MGT-01 verify that the constraints are enforced as feasibility conditions?
1. By a written declaration from the supplier's chief executive
2. By the Region's own seeded-breach suite of 12 breaches per release, of which none may reach a manager
3. By an annual audit that the supplier's own staff carry out on a test installation and report to the head of procurement
4. By counting the complaints received from staff after each release
What does version 1.4 of DLG-HM-REC-01 add to the equality monitoring of TalentLens?
1. A release condition: no release or criteria change takes effect until the test has run and every trigger has a recorded decision
2. A rule that selection rates are measured once every three years
3. An exemption for stages where the cell size is below 30
4. A transfer of the test to the supplier, on the ground that Article 10 of the AI Act already obliges providers to examine their data for bias
What is the chapter's answer to Q11, what replaces the manager's judgement in the duty of care?
1. Wellbeing analytics that detect strain from each person's behaviour and report it to the line manager, provided the safety representative has agreed the indicators in advance
2. The supplier's assurance that the system complies with working-time rules
3. An autonomous agent with authority to relax rest rules when demand is high
4. Nothing replaces it: a system can carry protective constraints and flags, while managers need the span and representatives the access for the rest

Module 6 · Part VI · Designing the Accountable AI Boss

Part VI turns the book's analysis into designs. It supplies the Management Decision Record and the delegation register, the oversight record and oversight-quality metric, the appeal path and contestation record, the worker information notice and transparency register, and the role charter of the transformed manager. Participants leave able to introduce decision records for one system, test whether human oversight is real, build an appeal path that corrects the system as well as the case, write a notice workers can act on, and budget a manager's time for development and care.

Lecture · 55 minutes

Designing the Accountable AI Boss: Record, Oversight, Appeal, Notice, Role

Part VI is the design part of the book, and its five chapters form one loop. A management decision that AI made or shaped leaves a record; the record shows whether the human in the chair really decided; the person affected is told that the system exists, can learn what it did and can contest it; a successful contest corrects the rule and not only the case; and the line manager, whose role is transformed and not abolished, holds the loop together. The lecture walks the loop with the book's two first contests, a route with no travel time at Nordvik Region and a quarterly score at Hansa Mutual.

11 min	The decision as a record A Management Decision Record is about a person, so it is personal data, has a reader with rights, is evidence in disputes, and must be produced by the work and not written beside it. The reasons are in the decider's words, and the system's version and proposal make the decider's part visible. A schema rejects Level 4 Autonomous for M3 Consequence and a decider named system or agent, which makes the record a condition of the decision. The delegation register shows the sum of delegation, runs reversibility tests and freezes further delegation where a test fails.
11 min	Oversight that is real A person who signs in a second has not read; data protection law treats a human without influence as absent. Six conditions make oversight real: time, authority to deviate, independent information, a measured deviation rate, no penalty for deviation and a stop path. Oversight of the system can be sampled; the decision in an M3 Consequence case cannot. The oversight-quality metric tests the design, not the people: aggregate by default, separate the uses, own figures only.
11 min	Contest as a condition of delegation An explanation has four elements: the role of the system, the inputs about the person, an actionable counterfactual and the human decider. In the appeal path the decider can say yes; only an independent reviewer can say no. A successful contest corrects up to five things: the decision, the data, the rule, the model and the register, and the record closes only when the system correction passes its test. Every degree above Level 1 Assisted rests on an appeal that reaches a person able to change the outcome.
10 min	Telling the managed Five things are disclosed: existence, logic, inputs, the degree of delegation and the named decider. The trade secret objection limits the form of disclosure, not its fact. A notice nobody reads informs nobody: layer it, time it at the moment of relevance, and have the line manager explain it. Change classes close the route by which Mode 0 Ambient AI features become managers without anyone being told.
12 min	The transformed manager Work that concerns the work moves towards systems under written delegation; work that concerns the person stays with the manager and gains weight. The AI boss is acceptable only where it frees a manager to do what the evidence on good managers describes, and must be evaluated on whether it does. The role charter adds a time budget with a protected floor for M4 Development and M5 Care, and a recorded decision on what happens to time saved. The thesis is falsifiable, and the author's projections are labelled as his own judgement.

What the audience takes away

Records, oversight, appeal and notice are one loop: each depends on the others, and a gap in any one undoes the degree of delegation it supports.
Measure the review and not the reviewers, and correct the system and not only the case.
The manager's role is transformed, not abolished: time saved by systems becomes development and care only if a charter protects it.

Chapter 25 · 105 minutes

Management Decisions as Records: Decision Records and the Delegation Register

The lesson supplies the two artefacts on which Part VI rests. Participants learn the fields of the Management Decision Record, the four-part rule for when one is required, how records arise from system events and the decider's action, and how a schema makes the record a condition of the decision. They study the delegation register with its sums, reversibility tests and freeze, the rules on privacy and access, and the answer to open question Q3: one record, four readers.

Learning objectives

Explain how a Management Decision Record differs from an Architecture Decision Record and what follows for privacy and volume.
Apply the four-part rule to decide which management decisions require a record.
Explain what the delegation register adds to single profiles: sums, discovered features, reversibility tests and the freeze.
Draft a Management Decision Record that serves the worker, the representatives, a court and a supervisory authority.

Key concepts

Management Decision Record: A short, numbered, immutable record of one management decision about one or more people that an AI system made or shaped: what the system proposed, who decided, why, what the person was told and how to contest it.
Record by operation: The record is generated from the system's events, which supply proposal, inputs, version and flags, and from the decider's action, which supplies the decision and reasons. A record that fails validation holds the decision.
Reversibility test: A test of whether a unit could perform one function without the system for a defined period, and whether its managers still can. The result is passed, passed with conditions or failed.
Freeze: Where a reversibility test fails or passes only with conditions, further delegation in that function and team is frozen until the conditions are met. The system keeps running at the agreed degree; the next step stops.
One record, four readers: The worker gets the explanation and appeal route, the representatives an aggregate, a court a record made at the time by the person who decided, and a supervisory authority the evidence of demonstrable compliance.

Teaching plan

15 min	The teacher places the Management Decision Record beside the Architecture Decision Record and the AI Decision Record and draws out the four differences that follow from its being about a person.
20 min	Field-by-field reading of MDR-NR-0001; participants identify the three fields that carry most weight and apply the four-part rule to ten example decisions.
20 min	Presentation of the delegation register and the reversibility tests RVT-NR-ROS-2027-01 and RVT-HM-CTA-2027-01, with what each found and froze.
15 min	Discussion of minimisation, retention, public access, and access for the decider and the representatives.
35 min	Exercise: groups write one Management Decision Record and the four views of it, then design a reversibility test for the system behind it.

Discussion

The chapter says a reason the system generated and the decider accepted is the system's reason. How should a decision screen be designed so that the reasons are the decider's own?
A store of decision records also records how each manager decided, how quickly and how often against the system. What uses of it should be agreed, and which excluded?
Nordvik's reversibility test found that the fallback had fallen behind the system it replaces. What would such a test find in a unit you know?

Exercise · 35 minutes

One record, four views

Choose one decision that AI made or shaped about a person: an approved roster, an overridden allocation, a shortlist or a pay proposal, at Nordvik Region, Hansa Mutual or your own organisation. Write its Management Decision Record with the field groups of the chapter: identity, decision class and function, person by pseudonymous reference, system and proposal, degree of delegation, human decider, changes and reasons in the decider's words, information given, contest route and retention rule, with no health data. Then write in two sentences each what the worker, the representatives, a court and a supervisory authority would read from it. Finish with a reversibility test for the system: scenario, period and what a pass would require.

Deliverable: A Management Decision Record in the form of MDR-NR-0001, four short reader views and a one-paragraph reversibility test design.

Self-test

Which of these does the chapter's four-part rule require a Management Decision Record for, without sampling?
1. Every allocation of a claim by the claims-triage agent
2. Every meeting summary that proposes an owner for an action
3. Every dashboard figure shown to a unit manager
4. Every M3 Consequence decision that AI shaped
What did Nordvik Region's reversibility test RVT-NR-ROS-2027-01 freeze?
1. Any new ShiftWise action class, including the shift-swap feature, until the conditions are met
2. All use of ShiftWise in the Strandvik unit until the supplier is replaced
3. The home-care dashboards, which were switched off in every unit
4. The unit manager's right to change a roster before publication
Why does MDR-HM-0001 state that the handler worked an agreed reduced schedule, without saying why?
1. Because the team leader did not know the reason
2. Because KBV-IT-2024 forbids any mention of a handler's working schedule in records that the Konzernbetriebsrat may read
3. Because the record holds no health data: the reason for the decision is the schedule, not its cause
4. Because the claims-triage agent deletes reasons after 24 hours

Chapter 26 · 90 minutes

The Human in the Chair: Oversight That Is Real, Not Rubber-Stamped

The lesson asks when the person named in a record is a decider and when merely a signature. Participants learn what the research says about automation bias and selective adherence, the six conditions that make oversight real, the difference between oversight of the system and the decision in the case, and the oversight record and oversight-quality metric. They analyse OVR-NR-ROS-2027-1 and the drift found in recruiting, and answer open question Q9 on oversight at volume.

Learning objectives

Explain why a requirement of a human in the loop does not by itself make a decision human.
State the six conditions under which oversight is real and how each can be recorded.
Distinguish oversight of the system, which can be sampled, from the decision in the case, which cannot.
Set up an oversight-quality metric with floors, seeded cases and aggregation agreed with the representatives.

Key concepts

Selective adherence: The finding that advice aligned with group stereotypes was adopted more readily, whether it came from an algorithm or a person. The record of departures must therefore show for whom overseers depart, not only how often.
Six conditions of real oversight: Time per decision; authority to deviate without heavier justification; information independent of the system; a measured deviation rate; consequences that do not punish deviation; and a stop path outside the system.
Decision in the case: The decision about one person, such as whether this applicant is invited. It cannot be sampled, because a sample showing ninety-five sound decisions gives the other five people no decision.
Oversight record: For one delegation profile and one period, who oversaw, what they saw, what they did and with what authority. It holds the evidence of the observed degree and names no overseer in its figures.
Designing volume: Volume is set by the owner: decisions multiplied by minimum time give the hours to plan. Where reading is not possible at the given volume, the class is reclassified, the volume reduced or fewer decisions produced.

Teaching plan

15 min	The teacher presents the survey of oversight policies, the reported case of approvals at 1.2 seconds each, and the research on automation bias, selective adherence and algorithm aversion.
15 min	Presentation of the six conditions with their anchors in Article 14(4) and Article 26(2) of the AI Act and Article 10(2) of the Platform Work Directive, as organisational implications.
20 min	Reading of OVR-NR-ROS-2027-1: participants classify the observed degree from the figures and propose measures before seeing those of AIDR-0630.
30 min	Exercise: groups set up an oversight-quality metric for one decision class and compute the hours of oversight the volume requires.
10 min	Debrief on the Polish recruiting record: what drifted, why volume was the cause, and what version 1.5 of the profile changed.

Discussion

Ingrid Halvorsen asked why the design allowed two-minute approvals, not who made them. What changes in an organisation when findings are made against the profile and not the person?
A screen on which accepting is one click and departing needs a sign-off teaches acceptance. Where do your own approval screens make departure more costly than acceptance?
The answer to Q9 places accountability with whoever set the volume. Who sets the volume of decisions per decider in your organisation, and do they know it?

Exercise · 30 minutes

An oversight-quality metric for one class

Choose one decision class in which a person approves or decides on a system's proposal: roster approval, shortlisting on a ranking, or a pay proposal on a score. Name the human role: oversight of the system or the decision in the case. Write the six conditions for this class. Choose the measures and set floors and ceilings from the chapter's defaults, including a floor in minutes suited to the class. Design the seeded cases and the gate that keeps them from reaching a person. Fix the aggregation, who sees which figures and the excluded uses. Then multiply the expected number of decisions by the minimum time and state the hours of oversight to plan, and what happens when demand exceeds them.

Deliverable: A metric specification with human role, six conditions, measures with floors and ceilings, seed design, aggregation and excluded uses, plus the volume calculation and the rule when it is exceeded.

Self-test

What did the figures in OVR-NR-ROS-2027-1 lead Nordvik Region to conclude about roster approval?
1. That three named unit managers should be retrained or replaced
2. That the design was at fault: a screen let eleven flags be closed with one click and a wave went live before four of seven managers were trained
3. That ShiftWise should be raised to Level 3 Bounded for all rosters
4. That seeded faults should be abandoned because managers found them unfair
Why can the decision in an M3 Consequence case not be replaced by sampled oversight of the system?
1. Because sampling is prohibited by the AI Act for all high-risk systems
2. Because seeded cases cannot be constructed for consequence decisions
3. Because consequence decisions are too rare to sample reliably
4. Because a sample showing most decisions sound gives the remaining people no decision at all
What caused Level 1 Assisted to drift in the Polish recruiting team, according to OVR-HM-REC-2027-Q1-PL?
1. Volume: nineteen reviews per recruiter-hour at peak against a cap of twelve
2. A supplier release that switched affect analysis back on
3. Recruiters refusing to use the ranking at all
4. The works council withdrawing its agreement to the metric

Chapter 27 · 90 minutes

Contestability by Design: Explanation, Appeal and Correction

The lesson designs contestability into the AI boss. Participants learn the three justifications for contest, the routes by which a worker can learn what a system did, the four elements of an explanation people can use, the five parts of an appeal path and the five layers of correction. They work through the contestation records CNT-NR-0001 and CNT-HM-0001, read contest patterns as triggers, and see why every degree above Level 1 Assisted depends on a working appeal.

Learning objectives

Write an explanation with the four elements the chapter requires, including an actionable counterfactual.
Describe the five parts of an appeal path and the rule that the decider may grant but never reject.
Specify correction at five layers and the condition on which a contestation record closes.
Interpret contest patterns, including the absence of contests, as signals about a system and its oversight.

Key concepts

Four-element explanation: The role of the system and its logic in a sentence a person can repeat; the inputs about the person, and only those; a counterfactual naming an input that could actually change; and the human decider with how to reach them.
First-look rule: The decider may grant a contest in full and correct the decision at once, but may not reject it or grant it in part; those go to independent review. The decider can say yes; only the reviewer can say no.
Independent review: Review by someone who has not taken, approved or advised on the decision and is not the system, its supplier or an agent. An AI system may summarise evidence at Level 0 Advisory but never drafts the finding.
Five layers of correction: The decision is changed and loss made good; the data is corrected at source; the rule is changed and written as a test; the model is retrained or withdrawn where it erred; and the register is updated.
Contest patterns: A rising rate of contests granted on review, a cluster about one rule or group, or no contests at all from a system making thousands of decisions. Each calls for action, and silence is not evidence that the system is right.

Teaching plan

15 min	The teacher presents the three justifications for contest and the warning from the automated debt-recovery scheme: individual reviews succeeded while the system stayed in place.
15 min	Participants rewrite a formal explanation of a route into the four-element form, with a counterfactual that names something the organisation can change.
20 min	Walk through CNT-NR-0001 from oral contest to closed record: first look, review, findings, the five layers of outcome and the test that closed it.
30 min	Exercise: groups build an appeal path for one management system and run a test contest through it on paper.
10 min	Debrief on contest patterns and the triggers in version 4 of DLG-HM-SPS-01, and on contestability as a condition of delegation.

Discussion

The route in CNT-NR-0001 was lawful and unworkable, and only the person doing the work could see it. What does that imply for how much weight an organisation gives to contests compared with automated checks?
Why does the chapter forbid the decider to reject a contest at the first look, even when the decider is sure the decision was right?
A system that makes thousands of decisions and receives no contests may be unknown, distrusted or feared. How would you find out which?

Exercise · 30 minutes

An appeal path with a test contest

Build an appeal path for one management system at Nordvik Region, Hansa Mutual or your own organisation. From its delegation profile, list the contestable decision types with their deciders. Write the explanation template with the four elements. Name receivers and reviewers by role, each with a deputy, and state who is excluded from review. Set the time limits for acknowledgement and answer, and what happens to a pay effect while a contest is open. Write the first-look rule. Define correction at the five layers, each with an owner and a test. Then run one invented contest through the path and write its contestation record, including the system correction and the condition on which the record closes.

Deliverable: A one-page appeal path with roles, limits, first-look rule and correction layers, and a completed contestation record in the form of CNT-NR-0001.

Self-test

What did the review of CNT-NR-0001 find to be the cause of the route with no time between two visits?
1. The unit manager had removed the travel time by hand when approving the roster, in order to fit an extra visit into the route
2. The assistant nurse had entered the wrong care-plan minimum
3. Both addresses were geocoded to the same postcode centroid, so routed travel time was 0 minutes
4. The rest-rule flag had been switched off by a supplier release
When does a contestation record close, according to the chapter?
1. When the worker's own decision has been changed
2. When the system correction has passed its test
3. When the decider has given an oral explanation
4. When two weeks have passed without escalation
Under version 4 of DLG-HM-SPS-01, what happens to the pay effect of a quarterly score while it is contested?
1. It is paid in full and cannot be recovered
2. It is withheld until the next annual review
3. It is decided by the score alone until the reviewer answers
4. It is provisional and corrected in the next salary run

Chapter 28 · 75 minutes

Transparency to the Managed: What Workers Are Told, When and by Whom

The lesson designs what workers, their representatives, the public and supervisors are told about a management system, when and by whom. Participants learn the four audiences and five objects of disclosure, the information duties read as organisational implications, why the trade secret objection limits form and not fact, and why notices go unread. They write and test a worker information notice on the model of WIN-NR-SW-HC-01 and outline a transparency register.

Learning objectives

Distinguish the four audiences of transparency and the five objects of disclosure.
Explain why a supplier's trade secret limits the form of disclosure and not its fact.
Write the first page of a worker information notice in the workers' own words and test it with readers.
Classify changes to a system into the three notice classes with their lead times.

Key concepts

Five objects of disclosure: Existence of the system; its logic, with main parameters; the inputs about the person and those excluded; the degree of delegation; and the named decider who answers and receives a contest.
Worker information notice: A versioned artefact that states, for one system and one group of workers, what the system does, which functions it performs at which degree, what data it reads and must not read, who decides and how to contest.
Three layers and three timings: One page readable in five minutes, the full notice and the detail on request; issued before deployment, at each change class and at each decision as a notice extracted from the Management Decision Record.
Transparency register: A list of every system that performs a management function, per entity, with usage mode, degree per function, profile and notice in force, contact, and the representatives' agreement and its date, including agreements still pending.
The manager explains: A notice forwarded by email has been delivered, not explained. The line manager should be able to answer unaided what the system does, what it reads, who decides, what the manager sees and what to do if it is wrong.

Teaching plan

10 min	The teacher presents the four audiences and five disclosures and asks which of the five participants could state today for a system that manages them.
15 min	Presentation of the duties to individuals and to representatives, the template in Article 9 of the Platform Work Directive, and the trade secret and gaming objections.
10 min	Discussion of why notices go unread and the three design responses: layering, timing at the moment of relevance and a person who explains.
30 min	Exercise: participants write the first page of a notice and test it on a partner who answers three questions without the text.
10 min	Debrief with the history of WIN-NR-SW-HC-01 from version 1.0 to 1.2 and the classification of each change.

Discussion

None of the six testers could say what 'optimising route efficiency under working-time constraints' meant for their day. What questions would workers in your organisation actually ask about a system that manages them?
Few legal texts ask for the degree of delegation to be disclosed, yet the chapter says it answers the worker's real question. What is that question, and why does it matter?
A system people can game by knowing its inputs measures the wrong thing, and secrecy only hides the defect. Where would you expect that argument to meet resistance?

Exercise · 30 minutes

A notice workers can act on

Write the first page of a worker information notice for one system and one group of workers, generating every statement from the system's delegation profile and not from a vendor description. Use short sentences and the workers' own words, at most 400 words. State what the system does, each function it performs with its degree in words a worker would use, the inputs used and the inputs never used, the named decider and how to contest with the time to an answer. Add one counterfactual example. Then test it: a partner reads it once, puts it away and says what the system does, who decides and how to object. Revise until all three can be answered, and list the changes that would be Class 1 or Class 2.

Deliverable: A one-page notice in the form of WIN-NR-SW-HC-01, the result of the three-question test before and after revision, and a short list of change classes.

Self-test

Which inputs does WIN-NR-SW-HC-01 tell Strandvik home-care staff are never used by ShiftWise?
1. Whether they accepted or declined earlier offers, sickness absence, dashboard figures and union membership
2. Care plan visit times, travel time by road between addresses, and the working-time and rest rules of the collective agreement
3. Their competences, contract hours and stated preferences
4. The published rotating order of open-shift offers
In the chapter's change classes, how is a new or removed input, or a changed threshold or bound, handled?
1. Class 3: a change log entry visible on request, with no lead time
2. Class 1: a new version of the notice, a workplace meeting and consultation or negotiation where law requires, 15 working days before use
3. Class 2: an updated notice, a one-line summary to each person and a diff to representatives, by default 5 working days ahead
4. No class: inputs are a matter for the supplier's release notes
What does the chapter conclude about a supplier's trade secret as an objection to disclosing a system's logic?
1. It releases the employer from all information duties towards workers
2. It limits the form of disclosure, not its fact; the tender should secure notice content and permission to disclose it
3. It can be overridden only by a court order obtained in each individual case, after the supplier has been heard on its commercial interests
4. It applies to representatives but not to individual workers

Chapter 29 · 90 minutes

The Transformed Manager: Leading People When AI Directs the Work

The lesson describes the person the designs of Part VI assume: the line manager whose role is transformed, not abolished. Participants weigh three accounts of the manager's future, map what systems take and what stays function by function, and measure the AI boss against the evidence on good managers. They write a role charter with a time budget and protected floor on the model of ROLE-NR-MGR-01, and examine what would show the thesis wrong.

Learning objectives

Compare the three accounts of the manager's future and the evidence the chapter gives for each.
State, for each management function, what systems take and what stays with the manager.
Write a role charter with decision rights per function, a never list and a time budget with a protected floor.
Identify the findings that would weaken or refute the thesis that the role is transformed, not abolished.

Key concepts

Transformed, not abolished: Work that concerns the work, such as assigning, sequencing and measuring, moves towards systems under written delegation. Work that concerns the person, such as deciding a consequence, developing a skill and noticing ill health, stays with the manager and gains weight.
The standard of the good manager: The AI boss is acceptable only where it frees a manager to do what the evidence on good managers describes, such as coaching and building psychological safety, and it must be evaluated on whether it does.
Role charter: A versioned artefact that binds the manager's other artefacts to a person: for each function what the manager decides, what systems do under which profile, what may not be delegated, and to whom the manager escalates.
Time budget and protected floor: The share of working time assigned to each function and to conducting agents. The default floor is at least 35 per cent on M4 Development and M5 Care together, with a recorded decision on what happens to time saved.
Horizon register: A register that labels trends for the role as observable, expected or speculation, with signals that have owners and thresholds, so that the organisation can see whether a projection is coming true.

Teaching plan

15 min	The teacher presents the three accounts of the manager's future and the point at which the parallel with the security architect breaks.
15 min	Participants complete the table of what systems take and what stays for each of the five functions, then compare it with the cited evidence on effective managers.
15 min	Discussion of span of control, the vendors' human-agent ratio labelled as speculation, and what an organisation should measure before and after widening a span.
35 min	Exercise: participants write a role charter and time budget for a line manager from a two-week diary, real or estimated.
10 min	Closing: the four findings that would weaken the thesis, and the author's projections read as his own labelled judgement, not as findings.

Discussion

The chapter says conducting agents is the smaller half of the transformation and leading the people the systems direct is the larger. Does that match what you see in organisations that deploy agents?
Time saved by ShiftWise went to home visits with staff because a decision record said so. Where does time saved by systems go in your organisation, and who decides?
The most plausible alternative to the thesis is that relational work could be done by peers, HR or coaches outside the line. What evidence would persuade you either way?

Exercise · 35 minutes

A role charter with a time budget

Write a role charter for one line manager whose team works with an AI boss: Mikael Strand, Jana Petersen or a manager you know. For each of the five functions write what the manager decides, which system acts under which profile and at what degree, and the stop right. Name a human decider for every M3 Consequence class. Write the never list of what the manager may not delegate. Build a time budget from a two-week diary, real or estimated, with shares per function, a cap for conducting agents, a share for unassisted practice and a protected floor for M4 Development and M5 Care. State the decision on what happens to time saved and the triggers for review.

Deliverable: A role charter in the form of ROLE-NR-MGR-01 with decision rights per function, never list, time budget with floor, decision on time saved and review triggers.

Self-test

What did AIDR-0633 decide about the time ShiftWise saved Mikael Strand?
1. The saving is booked as a wider span of control from the next budget year, with a baseline and a review date recorded in the role charter
2. The saving is used to add a second home-care unit to his responsibility
3. The saving is returned to the supplier as a reduced service level
4. The saving goes to the staff, and the span of control is not widened before the review of the commission
How does the chapter label the image of a manager of one human and twenty agents?
1. Observable, because pilots already run at that ratio
2. Speculation, because no study known to the book has measured it and what is observable points the other way
3. Expected, because vendors have published the ratio
4. Prohibited, because Article 26(2) of the AI Act caps the number of agents that one natural person may be assigned to oversee
According to the chapter, where does the parallel between the line manager and the security architect break?
1. The architect's AI team drafts policies, whereas the line manager's systems direct and measure people
2. The architect works under a mandate manifest and the manager never does
3. The architect remains the accountable verifier of everything agents propose, whereas the line manager answers only for the people
4. The architect's role is abolished while the manager's is transformed

Module 7 · Part VII · Experience and Verdict

Part VII turns from design to experience and verdict. It follows four people through a year with the AI boss at Nordvik Region and Hansa Mutual, sets that experience beside eight documented cases, three composites and twelve anti-patterns, draws the conclusions and 24-month plans of the two organisations, and judges the book's thesis claim by claim. Participants leave able to run a quarterly review with representatives, review an organisation against the anti-pattern checklist, write a prioritised plan and answer ten questions about their own AI boss.

Lecture · 50 minutes

Experience and Verdict: A Year, Twelve Anti-Patterns and a Named Person in the Chair

The closing lecture puts the instruments of the book to work together and then judges them. A year at two organisations shows that the order of the instruments matters, and that every failure was found by a designed mechanism and none by a system's own account. Documented cases show one recurring movement: a system built to direct or evaluate comes to carry a consequence without anyone deciding it. The two organisations reach different plans because they hold different levers. The thesis holds with qualifications, seven of twelve open questions are answered, and the book adds a sixth commitment to the series: a named person in the chair for every consequence.

11 min	A year in the order it needed Commission, inventory, profiles, records, and only then measured oversight and reversibility tests; where the order was reversed, the cost appeared in the next record. A contest about one route became a hard constraint, an appeal path and a notice change within two months. Co-determination shaped every outcome that lasted, from the rotating order of open shifts to the workload indicator. Nordvik's profiles moved sideways and Hansa's inward; none moved upward.
10 min	The one movement in the cases In the eight documented cases a system built for M1 Direction or M2 Evaluation came to carry M3 Consequence, and the record, the decision right or the contest path that would have shown it was missing. Twelve anti-patterns fall into four families: delegation unrecorded, oversight in name only, people measured as machines, voice and craft lost. The anti-patterns feed each other, and a checklist of signals turns them into a quarterly review that cannot be argued away.
10 min	Two organisations, two levers Nordvik mostly procures, so its lever is the contract and the acceptance test; Hansa mostly builds, so its lever is the allocation rule and the release. Every degree raised during the year was protective or confined to agents' own actions; every degree touching evaluation or consequences was held, lowered or refused. Both 24-month plans put the measures that make delegation safe before the measures that extend it, and send at most three decisions to the management body.
11 min	The thesis judged Direction may be delegated up to Level 3 Bounded only with health constraints inside the optimiser and its figures fenced off from consequence decisions. The named human decider is the firmest claim and the one most easily emptied; it holds where the six capabilities hold. Records are supported as design while their effect is unmeasured, and reversibility decays within months of go-live. Q4, Q5, Q10 and Q11 are answered in part; Q12 is open by nature and left to the author's labelled verdict.
8 min	What to do first First map and write authority, next record and measure, later test reversibility and evaluate over years; raise no degree before the records exist. Ten questions test each year whether the organisation governs its AI boss or is governed by it. If you are managed by an AI boss, ask who your decider is: that question is the beginning of every correction the book describes.

What the audience takes away

What makes delegation safe is the human capacity to check it, and the mechanisms that find failures must be designed before the pressure arrives.
The lever differs, contract or code, consent or negotiation, but the rule is shared: no system holds a consequence decision about a person beyond Level 1 Assisted.
The thesis holds with qualifications and names what would refute it; a verdict that cannot be revised is a forecast in disguise.

Chapter 30 · 90 minutes

A Year With the AI Boss: Nordvik Region and Hansa Mutual

The lesson follows Ingrid Halvorsen, Mikael Strand, Jana Petersen and Thomas Brandt through one year, from December 2026 to November 2027, with the book's instruments running together in the order the year needed them. Participants trace how failures were found and answered quarter by quarter, read the second oversight record OVR-NR-ROS-2027-2, see how the four roles changed, and practise the quarterly review with representatives.

Learning objectives

Reconstruct the order in which the two organisations introduced their instruments and explain why the order mattered.
Match each failure of the year to the mechanism that found it and the recorded decision that answered it.
Compare how the delegation profiles of the two organisations moved over the year.
Run a quarterly review of delegation, decisions, contests and oversight with workers' representatives.

Key concepts

Order of the instruments: Commission, inventory, profiles, records, and only then measured oversight and reversibility tests. Where the order was reversed, as when a wave went live before its managers were trained, the cost appeared in the next record.
Found by a designed mechanism: Most of the year's failures were found by a sample, a disparate-impact test, a contestation, a reversibility test, an oversight record with seeds or a routine trigger; none was found by a system's own account of what it did.
Sideways and inward: Nordvik's profiles moved sideways: more units, more constraints, no higher degree. Hansa's moved inward: the same degrees, with more of the conditions that make them true written in.
Quarterly review with the representatives: A shared view each quarter of what has been delegated, decided and contested and whether oversight held, generated from the records, with no figure about a person and failures read first.

Teaching plan

10 min	The teacher introduces the four people and the starting point of each organisation in November 2026.
25 min	Timeline work: four groups take one quarter each, list its records and decisions on a wall chart, and present what was found and what answered it.
15 min	Reading of OVR-NR-ROS-2027-2: participants identify the finding about holiday stand-ins and why the fault lay in a training rule attached to the wrong object.
30 min	Exercise: a role-played quarterly review with function owner, line manager, representative and data protection officer.
10 min	Debrief on how the four roles changed and on what neither organisation has yet solved.

Discussion

When approvals of draft letters fell below 30 seconds in the holidays, Jana Petersen lowered the agent's budget without waiting for the evaluation plan's stop rule. What does that decision show about where authority over agents should sit?
Thomas Brandt's work moved from negotiating agreements to negotiating classes of change. What skills and information does a representative need for that work?
Two failures were found late: a fallback that had fallen behind the system and a checkpoint running at a higher degree than its profile. Why are these the hardest kind to see?

Exercise · 30 minutes

A quarterly review with the representatives

In groups of four, take the roles of function owner, line manager, workers' representative and data protection officer, and hold the quarterly review for one quarter of the year at Nordvik Region or Hansa Mutual, or for your own organisation. Build the pack from the records named in the chapter: profile versions changed, decision record counts by class, contestation records with times to answer, the oversight record and the last reversibility test. Check that no figure concerns a person. Read failures first, each with its answer or an open status. Compare one profile with its practice. Classify two pending changes under the agreed change classes, letting the representative move one up a class. Record the outcome and the representative's position.

Deliverable: A one-page review record: pack contents, aggregation check, failures with answers or open status, one profile compared with practice, classified changes and the representatives' position.

Self-test

What did the union for municipal and care workers change in Nordvik Region's open-shift offers during the negotiation of January 2027?
1. Offers go out in a published rotating order, in place of going to everyone with first acceptance winning
2. Offers are ranked by how often each person accepted in the past
3. Offers require the unit manager's approval for each individual shift before they reach staff, which moved the class to Level 2 Supervised
4. Offers are restricted to full-time staff with at least two years of service
What did the second oversight record, OVR-NR-ROS-2027-2, find about holiday stand-ins?
1. They approved rosters more slowly than unit managers and caught every seed
2. They refused to approve any roster drafted by ShiftWise
3. They found half the seeded faults, and three of six had not taken the oversight module
4. They closed eleven rest-rule flags with one click
How did the main delegation profiles of the two organisations move during the year, according to the chapter?
1. Both raised M3 Consequence to Level 2 Supervised once records existed
2. Nordvik raised its degrees while Hansa lowered all of its own
3. Both moved every profile to Level 0 Advisory after the first contest
4. Nordvik's moved sideways and Hansa's inward; none moved upward for decisions about a person

Chapter 31 · 90 minutes

Case Studies and Anti-Patterns

The lesson retells eight documented cases in one frame, each with the functions involved, the degree of delegation as observed, what was missing and the lesson as a rule. Three composite cases and twelve anti-patterns in four families follow. Participants learn to read a case for the recorded and the observed degree, and review an organisation's AI boss against the anti-pattern checklist, answering each signal from records.

Learning objectives

Analyse a documented case by functions, degree as recorded and as observed, what was missing and the rule it yields.
Classify the twelve anti-patterns into the four families and name the controls that prevent each family.
Explain how anti-patterns feed each other, using the examples the chapter gives.
Review an organisation against the anti-pattern checklist and give each red finding an owner, remedy and re-test date.

Key concepts

Anti-pattern: A reasonable-looking solution that repeatedly produces the same harm. The chapter's twelve are latent conditions with measurable signals, each accumulating unseen until a person pays for it.
Degree as observed: The degree of delegation at which a system runs in practice, often higher than the one recorded. Level 4 by accident (AP-10) is the case where a consequence has no named human decider.
Four families: Delegation unrecorded; oversight in name only; people measured as machines; and voice and craft lost. Each family has the cases that illustrate it and the controls of the book that prevent it.
Signals, not opinions: The checklist holds a signal and a red condition for each anti-pattern, answered from records. A signal that cannot be answered from a record is also red.
Composite case: An illustrative case of an invented organisation that combines patterns from earlier chapters: the benchmark agent, the silent release, and the roster that went well because the controls existed before the pressure.

Teaching plan

10 min	The teacher explains the frame used for every case: functions and degree, what happened, what was missing and the rule.
25 min	Case clinic: groups take two documented cases each and present them in the frame, stating the recorded and the observed degree and reading out the rule.
15 min	Presentation of the twelve anti-patterns in four families and how they feed each other; participants place the three composite cases.
30 min	Exercise: groups run the checklist against an organisation and write up the red findings.
10 min	Debrief: which anti-patterns Nordvik Region and Hansa Mutual found and remedied, and which they avoided.

Discussion

In most of the cases the harm sat where a system built for direction or evaluation came to carry a consequence without anyone deciding it. Why is that movement so hard to see from inside the organisation?
A manager measured against the system stops departing from it, which produces the rubber stamp. Which other anti-patterns do you see feeding each other in practice?
Composite C avoided all four families because the controls existed before the pressure. What makes it difficult for an organisation to invest in controls before anything has gone wrong?

Exercise · 30 minutes

Review against the anti-pattern checklist

Review one organisation's AI boss against the twelve anti-patterns: your own, or Nordvik Region or Hansa Mutual at a point in the year of Chapter 30. Fix the scope to the systems and features in use over the last two quarters. For each anti-pattern, AP-01 to AP-12, state the signal you would look for and the record that would answer it. Mark each red if the signal meets its condition or cannot be answered from a record, amber for a near-miss already remedied, green otherwise. For each red finding write the remedy from the book, an owner who is not the subject of the finding and a re-test date. Report at team level only.

Deliverable: A checklist table of twelve rows with signal, source record and colour, and for each red finding a remedy, owner and re-test date.

Self-test

Which movement do the eight documented cases share, according to the chapter's summary?
1. A system built for M5 Care came to replace the safety representative
2. A system built for M1 Direction or M2 Evaluation came to carry M3 Consequence without anyone deciding that it should
3. A system built for M3 Consequence was reduced to an advisory role by the managers who used it, so that its rankings no longer shaped any decision
4. A system built for M4 Development was reused to set the pace of work
In Composite B, the silent release, what did a routine release of the workforce suite enable by default?
1. Automatic approval of rosters without the unit manager
2. A chart comparing handlers' output with agents' output
3. Emotion inference from store staff's voices
4. Offering extra hours first to staff with the highest acceptance history
To which family does the chapter assign the rubber-stamp approval (AP-02), the manager measured against the system (AP-08) and the explanation nobody can read (AP-11)?
1. Oversight in name only
2. Delegation unrecorded
3. People measured as machines
4. Voice and craft lost

Chapter 32 · 90 minutes

Nordvik Region and Hansa Mutual: Conclusions for Two Organisations

The lesson states, for each of the two organisations, a conclusion per management function, three decisions it got right and two it got wrong or has still to make, and its answer to the body that asked. Participants compare a public authority that mostly procures with a mutual group that mostly builds, read the two 24-month plans AIDR-0640 and AIDR-0690 as artefacts, and turn a year's records into a prioritised plan of their own.

Learning objectives

Summarise each organisation's end-of-year position per management function against the book's default ceilings.
Explain why the two organisations reached different priorities from the same functions and the same rule.
Distinguish what transfers between organisations from what does not.
Draft a 24-month plan with ordered measures, owners, verifications and conditions under which a degree may rise.

Key concepts

Different levers: Nordvik mostly procures, so its lever is the contract and the acceptance test; Hansa mostly builds, so its lever is the allocation rule and the release. The form of co-determination explains the rest of the difference.
Holding the evidence: Nordvik's answer: a public authority can govern an AI boss it mostly buys only to the extent that it holds the profile, the stop path, the records, the contest path and the oversight figures itself.
Policy as a set of artefacts: Hansa's answer: the group policy is not one document but a set: the decision-rights file, a baseline of controls in all seven markets, the change-class annex, the transparency register and the delegation register.
Ordering by what is protected: The default order of a plan: weak oversight before new delegation, protective constraints before efficiency, and M3 Consequence before M1 Direction. A freeze is a measure in the plan, not a footnote.
Conditions for raising a degree: Written as thresholds. By default no raise without two consecutive periods of oversight figures above their floors and a reversibility test passed without conditions.

Teaching plan

15 min	The teacher presents the two questions asked in November 2026 and the table of the two delegation registers at the end of the year against the ceilings.
20 min	Two groups prepare a three-minute report to the governing body, one as Nordvik's HR director and one as Hansa's chief HR officer: what went right, what went wrong, what is still to decide.
15 min	Discussion of where the organisations differ and why, what transfers and what does not, and the three questions that remain unresolved.
30 min	Exercise: groups turn a year's records into a 24-month plan with at most three decisions for the management body.
10 min	Debrief with the seven lessons for other organisations; each participant names the one they would act on first.

Discussion

Nordvik's first right decision was to pause and not defend. What did the pause cost the Region and what did it gain?
Hansa's board has to decide whether the group's AI boss is one system with seven sets of conditions or seven systems. What follows from each answer?
The chapter advises treating the number of representative bodies as the pace of change. How would that change a rollout plan you know?

Exercise · 30 minutes

From a year's records to a 24-month plan

Following the chapter's practical guide, turn a year's records into a plan for the next 24 months, for Nordvik Region, Hansa Mutual or your own organisation. List the decisions of the year and mark each as done, in progress or reopened. For each management function, read the end-of-year degree against the ceiling and list every reversibility test passed with conditions and every freeze in force. Write five measures, ordered by what they protect, each with one owner by role, a period, the representatives' part and a verification that produces evidence. Write the conditions under which any degree may rise as thresholds. List what the plan excludes and the decisions, at most three, that go to the management body.

Deliverable: A plan in the form of PLAN-NR-MGT-2027 with five measures, owners, periods, verifications, conditions for raising a degree, exclusions and at most three decisions for the management body.

Self-test

What does the chapter identify as the thing Nordvik Region got wrong?
1. It negotiated the delegation profile with both unions before go-live and agreed in advance how supplier releases would be classed, which delayed the first roster
2. It paused its screening pilot when the supplier could not explain the ranking
3. The order of go-live: untrained approvers, a screen that allowed bulk closure and a fallback that had fallen behind the system
4. It allowed ShiftWise to decide variable pay at Level 3 Bounded
According to the chapter, what did every degree of delegation that was raised during the year have in common?
1. It was protective or confined to agents' own actions
2. It concerned a pay input or a hiring decision
3. It was raised at the request of the supplier
4. It applied only in markets outside the AI Act
What does Hansa Mutual's plan, AIDR-0690, do with the sales score as a pay input?
1. It raises the score to Level 2 Supervised in all seven markets
2. It removes the score as a pay input wherever managers' departures from it stay below 0.05
3. It keeps the score unchanged in every market, because the Konzernbetriebsrat agreed to it in an annex to KBV-IT-2024
4. It replaces sales managers with the score for variable pay decisions

Chapter 33 · 75 minutes

Conclusions: Who Should Be the Boss?

The lesson judges the book's thesis claim by claim, records how the twelve open questions are left, and states what the book adds to the series' argument: a named person in the chair for every consequence. Participants examine the falsifiers and the signals that would change the verdict, separate the book's findings from the author's labelled judgement, and answer the ten questions every organisation should be able to answer about its AI boss.

Learning objectives

State the verdict on each claim of the thesis and how firm the chapter says it is.
Explain why M3 Consequence and M5 Care mark the edges of what may be delegated.
Distinguish the book's findings from the author's labelled projections, and name the signals that would change the verdict.
Answer the ten questions about an organisation's AI boss with an artefact for each, and plan the first ninety days.

Key concepts

Thesis judged with qualifications: Direction may be delegated up to Level 3 Bounded only with constraints in the optimiser; the named decider is real only where the six capabilities hold; records are supported as design while their effect is unmeasured; reversibility decays within months.
Falsifiers: The findings that would show the thesis wrong, above all fully automated consequence decisions producing fairer and safer outcomes than human ones with records and appeal, and algorithmic direction without human leadership sustaining performance and wellbeing over years.
Edges of delegation: M3 Consequence and M5 Care mark the edges: a consequence needs a person who can answer before a forum, and noticing a person in difficulty is relational. Between them M1, M2 and M4 can be shared at degrees decided, recorded and reversed.
Sixth commitment: What the book adds to the series' argument of records, stop paths, kept skills, reversible dependence and autonomy matched to reversibility: keep a named person in the chair for every consequence.
First, next, later: Written authority before delegation, records before scale, measurement before raising a degree: map and agree in six months, record and measure up to eighteen, test reversibility and evaluate up to thirty-six.

Teaching plan

15 min	The teacher walks through the table of the thesis judged claim by claim, with verdict and firmness for each row.
10 min	Participants sort the twelve open questions into answered, answered in part and open, then check against the chapter's table.
15 min	Discussion of the author's verdict on who should be the boss in 2036, kept apart from the findings, and of the four signals that would change it.
25 min	Exercise: participants answer the ten questions for one organisation and draft the first ninety days.
10 min	Closing round: each participant states the first question they will ask about a system that manages them or their team.

Discussion

The chapter says the named human decider is the firmest claim and the one most easily emptied. What empties it, and how would you know it had happened?
A loud harm triggers a court case; a quiet harm triggers nothing unless someone has decided in advance to look. Which quiet harm would you look for first, and with which instrument?
The author expects European employers to divide between those that give saved time back to people and those that book it as a wider span. Which choices made now decide on which side an organisation ends up?

Exercise · 25 minutes

Ten questions and the first ninety days

Answer the chapter's ten questions for one organisation: which systems perform part of a management function, at which degree and who decided, whether any is at Level 4 Autonomous for M3 Consequence or above Level 1 Assisted for an adverse consequence, who decided one consequence last month, time per decision and departure rate, contestations and corrections, who can stop each system, whether each function could be run by hand, managers' time on development and care, and whether representatives were informed before use. For each question name the artefact that answers it or record a gap as a finding with an owner and a date. Then draft the first ninety days in three phases, map, agree and record, and list what is not to be done before day 90.

Deliverable: A ten-row table of answers with artefacts or findings, and a ninety-day plan in three phases with deliverables, verifications and the not-before-day-90 list.

Self-test

How does the chapter judge the claim that the sum of delegations is recorded and reversible?
1. Refuted, because no register can show the sum across teams
2. Supported without qualification, because both organisations passed their tests cleanly
3. Left open, because no reversibility test was run in the illustrative year
4. Supported, with the qualification that reversibility decays within months of go-live, not years
What does the book add to the series' argument as a sixth commitment?
1. Keep every AI system at Level 0 Advisory until 2036
2. Keep a named person in the chair for every consequence
3. Keep suppliers liable for every decision their systems recommend
4. Keep works councils out of the design of delegation profiles
In the first-90-days template, what must not happen before day 90?
1. Raising any degree of delegation or widening a span of control on the strength of time saved
2. Writing a function map per unit that includes Mode 0 features
3. Agreeing a delegation profile for each system with the works council or unions and the safety representatives
4. Running the first stop drill against its time budget

Answers to the self-tests

Chapter 1	1B · 2D · 3A (1) The first claim says the functions that assign and pace work can be delegated at a degree matched to reversibility, whereas the M3 Consequence decisions always have a named human decider who can be identified, contested and held to account. (2) The feature is a Mode 0 Ambient AI touch on M1 Direction and M2 Evaluation that was enabled by a vendor release; the map records the degree as unknown and the decision as not decided, which is itself the finding. (3) The chapter closes the worked example with exactly this contrast: procured systems, one regime and a duty to negotiate at Nordvik; built agents, seven regimes and a duty to co-determine at Hansa.
Chapter 2	1A · 2C · 3D (1) The boundary that matters most in M2 Evaluation is between measuring the work and judging the person; the per-person figure visible to the manager is evaluation regardless of its label. (2) Point 4(b) covers decisions on terms of work, promotion or termination, task allocation based on behaviour or traits, and monitoring and evaluating performance, so it reaches M1, M2 and M3 but not M4 Development or M5 Care. (3) The map notes that offered hours determine pay; for an hourly care assistant paid by the shifts offered, the allocation of open shifts is partly an M3 decision, and the function map must say so.
Chapter 3	1C · 2B · 3A (1) The chapter says scientific management moved direction to a planning department, made evaluation a comparison of observed with standard time and coupled both to consequence through differential pay, while M5 Care hardly appeared. (2) The three new properties are scale and speed, opacity and consequence without a human decision. Bonuses, the separation of conception from execution and written procedures all belong to earlier regimes. (3) The record sets M2 Evaluation at Level 0 Advisory, excludes the figure from M3 Consequence, gives each worker their own figures and a contest path, and reopens if the figure is proposed as an input to any M3 decision.
Chapter 4	1D · 2B · 3A (1) The chapter says the class sits at Level 3 Bounded only because of the excluded inputs; without that exclusion the action would rank people and exceed the ceilings of M2 Evaluation and M3 Consequence. (2) For adverse consequences the ceiling is Level 1 Assisted: the system drafts and a named person decides with their own authority. Level 2 Supervised applies to favourable or neutral consequences applying a rule already decided. (3) The line between Level 2 and Level 3 is where per-decision approval stops; the playbook is what bounds the system, so without one enforced outside the model the system in effect acts at Level 4 Autonomous.
Chapter 5	1C · 2A · 3D (1) For three drivers Uber had not shown how the human factor influenced the decision and they had no chance to state their view; for the fourth, whom Uber had interviewed, the court accepted real human intervention. (2) The benchmark found the log as a gap against Article 7 and reduced retention to 90 days; the feature ordering offers by past acceptance existed but was switched off and stays off by contract. (3) Access to booking windows depended on reliability and participation statistics blind to reasons, which the court framed as indirect discrimination. The chapter uses it as the warning for acceptance history becoming a hidden rating.
Chapter 6	1B · 2D · 3A (1) P2 removes the per-visit overrun from the manager's first view, because in home care a long visit is often the work done well; the individual view remains available on request and is logged. (2) The court disagreed with the CNIL on the fast-scan and idle-time indicators but upheld the retention finding. The chapter's lesson is that scanner data can be lawful as a narrow quality check and unlawful as a store of everything. (3) The chapter's answer to Q11 for paced work begins with treating the target as a safety-relevant setting owned by a named person and assessed with the safety representative.
Chapter 7	1D · 2B · 3A (1) Article 5(1)(f) prohibits inferring emotions from biometric data in the workplace, which the guidelines extend to candidates. Scoring transcript content remains high-risk evaluation under Annex III, point 4(a), not a prohibited practice. (2) AIDR-0601 records the criteria and weights as not disclosed and proprietary, with no outcome data for the Region's applicants. The Region could therefore not meet Article 86 or its duty to give reasons, and paused. (3) The chapter notes that a model can be tuned to pass the 4/5 rule whether or not it measures anything job-related, and that Directive 2000/78/EC applies to selection criteria and shifts the burden of proof once discrimination may be presumed.
Chapter 8	1C · 2A · 3D (1) The profile said Level 0 Advisory for M3 Consequence, yet the figures showed the score decisive in practice in one unit. The score was recalculated against the market baseline and the proposal re-made with the manager's own assessment first. (2) The Court reasoned that treating the score as a preparatory act would risk circumventing Article 22. The chapter concludes that labelling a score advisory does not settle the question; whether the decider draws strongly on it does. (3) The first condition is that the manager forms a view before seeing the score. It applies the series' conclusion that review must not start from the AI's answer.
Chapter 9	1B · 2D · 3C (1) The draft guidelines count behavioural indicators such as punctuality, reliability metrics or ratings as personal traits, whereas availability as indicated, accreditation or proximity fall outside point 4(b). The chapter notes the guidelines are a draft. (2) Without a rule for premium-bearing shifts the optimiser had concentrated them on the person with the widest availability. The change matters for pay, which is why the profile distributes premium shifts by the unit's rotation rule. (3) The consequence the chapter draws is that the planned roster is not the record of time worked: a roster that complies on paper can breach the rest rules once swaps, overruns and sick-leave cover are added.
Chapter 10	1A · 2C · 3B (1) The sample showed that proposed owners were being read as instructions although the profile said Level 0 Advisory. AIDR-0620 therefore changed the pilot so that a summary with unconfirmed owners cannot be distributed. (2) The fourth test asks whether anyone decided that the feature should perform a management function, at a stated degree, with a named person accountable. Where nobody did, the organisation has an AI boss it did not choose; the remedy is to decide it. (3) The assistants in the studies supported the worker and, in customer support, reduced the load on the human manager; they did not allocate or evaluate work. So the evidence on AI that assigns, summarises and reports about people is thin.
Chapter 11	1C · 2A · 3D (1) The result separates the functions: people accepted algorithmic assignment and scheduling about as readily as human assignment, resisted algorithmic evaluation, and missed the recognition that human decisions can carry. (2) Without the cap and the protection, a rule giving experienced handlers the complex claims they enjoy would take away the claims through which newer handlers learn, turning an M1 Direction rule into a question of M4 Development. (3) The guide's sixth step sets the threshold at the level of the unit and a dimension, never an individual; results are never used for M2 Evaluation or M3 Consequence of any person.
Chapter 12	1D · 2B · 3A (1) A rule written for continuity of claims handling put part-time handlers, most of them women in this population, at a particular disadvantage in access to the work that leads to promotion, without a justification that would pass the test. (2) The trigger starts a search for the cause within 20 working days and a recorded decision on whether the rule is objectively justified. The measurement alone settles nothing about justification. (3) A gap of roughly 7 per cent arose under a formula that does not know the driver's sex, explained by experience, preferences over where to work and driving speed. Whether such a gap is acceptable is a question for those who set the rule.
Chapter 13	1B · 2C · 3A (1) The trace had been used to estimate travel times; the review replaced the estimate with route-level averages, set raw events to 60 days and route aggregates to 24 months, and closed the route from dashboard figures to pay, discipline or dismissal. (2) The guidelines place inference from text outside the prohibition and inference from biometric signals inside it; the medical or safety exception does not extend to stress or burnout detection, and the Omnibus left point (f) unchanged. (3) The annex limits visibility to the handler and their own team leader, prohibits ranking lists, retains figures for 13 weeks and has the works council review the access log quarterly.
Chapter 14	1D · 2A · 3C (1) The contested route left no time to travel between two visits. The revised assessment made travel time a hard constraint written as a test, so a breaching route cannot reach the approving manager; AIDR-0605 records the decision. (2) A penalty is weighed against output in the objective, so the optimiser will cross the limit when the gain is large enough. A feasibility condition is a limit the system cannot trade for output. (3) Harm from algorithmically set pace is observable in warehouse work and expected in sectors such as home care and claims queues; for office agents theory predicts it where they raise demands and remove control, and the opposite where they remove drudgery.
Chapter 15	1B · 2D · 3A (1) The supplier's proposed success measure was kept as one output measure but rejected as primary: the representatives objected that it would rise as soon as routes were tightened, whatever happened to care or staff. (2) The chapter contrasts it with the restaurant study: monitoring that targets a specific misconduct can pay, but as a general management method it does not, on average, raise performance. (3) The verdict is divided, and the chapter adds that most organisations do not know whether it works because they have not designed a way to find out; the remedy is a multi-year evaluation agreed in advance.
Chapter 16	1C · 2B · 3D (1) For the model the chapter names, the rate fell from 96 per cent of runs to 37 per cent. The chapter's lesson is that instructions lower the rate without setting a ceiling, so boundaries belong in the environment. (2) The manifest's principle is that people manage agents and agents do not manage people. Classifying and drafting are permitted actions at Level 1 Assisted or Level 2 Supervised; allocating work to people is never permitted. (3) The triggers are team figures that suggest review is no longer real: seeded-case catch below 0.85, a rejection rate below 0.02 for four consecutive weeks, or a median approval time for letters under 30 seconds.
Chapter 17	1A · 2D · 3B (1) The three tests are owning the top mandate and approving every mandate below it, holding one stop path outside the chain, and being able to reconstruct every output that reached a person. No link holds M3 Consequence. (2) Because the routing reason is missing from the handler's view, that step runs at Level 3 Bounded and not Level 2 Supervised until the reason is added: a checkpoint is only as good as what the person can see. (3) The chapter concludes that an approval given by another automated component is not an approval by a person, whatever the log calls it; hence the limit that no agent may approve on behalf of a person.
Chapter 18	1C · 2A · 3D (1) Returns with a reason are expected and count against the agent's quality, never the handler's throughput, and checking time counts as work. Otherwise the right to return would carry a cost for the person. (2) The evidence is simulation-based, mostly military or emergency, and where compared human teams generally did better. The unstudied gap is where M4 Development and M5 Care lie, so practice is an experiment to be recorded. (3) The charter forbids comparing handler output with agent output, using agent throughput as a target or norm for handlers, and using a human-agent ratio as a staffing target without agreement. The other three are permitted practices.
Chapter 19	1B · 2D · 3A (1) The automation of first-line tasks is observable and erosion of judgement expected; that first-line management disappears as a career step is speculation. The response is reserved practice and a competence record. (2) The record therefore holds reasons and, where known, outcomes for a sample of changes, and refers to Management Decision Records without copying data about the staff concerned. (3) At Nordvik the deputy plans routes by hand one week a month; at Hansa ten per cent of new claims are allocated by the deputy without the agent's proposal. The further from authorship, the larger the reserve needed.
Chapter 20	1C · 2A · 3B (1) The decider answers to the person concerned and the decision is reopened on request; the function owner answers for the conditions and the management body for the allocation as a whole. The decider is not made the moral crumple zone. (2) The book's default is Level 2 Supervised with Level 1 Assisted for adverse decisions. A public employer that has paused its one consequence system keeps consequences further out of AI's reach for now. (3) The pause in the unit falls back to manual direction without stopping the business process: at Hansa the fallback for allocation to people is the team-leader queue, while claims continue to be routed.
Chapter 21	1C · 2A · 3B (1) The throughput-based factor is a behavioural input in the sense of the draft guidelines. Hansa treated the allocation function as high-risk under point 4(b) and now allocates on skill tags, open workload and declared availability only. (2) The prohibition concerns emotions inferred from biometric data of workers or candidates. Text sentiment is outside, physical states such as fatigue are not emotions, and tracking customers' emotions is not covered. (3) Article 27 applies to bodies governed by public law, private entities providing public services and deployers of the credit and insurance systems in points 5(b) and (c). Hansa assesses its management systems within its DPIAs.
Chapter 22	1D · 2B · 3A (1) The award is solely automated, but offered in a published rotating order it evaluates no personal aspects and no single award has significant effect. Ranking recipients or using acceptance history would reopen the analysis. (2) The exceptions in Article 22(2) are narrow in employment because of the imbalance of power. The chapter's design, built on a named human decider, does not depend on any exception. (3) The floor turns the determining-role test into a number the Konzernbetriebsrat can watch. A breached floor signals that the score, not the recruiter, is deciding, so the display is paused.
Chapter 23	1C · 2B · 3D (1) A behavioural input changes the bargain itself. The throughput-based capacity factor removed under AIDR-0651 is the example the annex was written to catch. (2) The field is entered by the team leader without reasons and records nothing about the handler's behaviour, so it falls within the agreed bounds for committee approval. (3) The statutes already ask whether a change is important, significant or substantial, not which version is running. Change classes tied to the profile's fields avoid both freezing and drift.
Chapter 24	1B · 2A · 3D (1) R1 answers with a test result and not a supplier statement the question the original tender did not ask. Only a test that tries to break a constraint shows whether it is a feasibility condition or a penalty. (2) Version 1.3 monitored selection rates but set no threshold and required no correction. The release condition gives the burden of proof something to rest on, with suspension of the affected stage on a trigger. (3) A system can carry hard constraints at Level 3 Bounded, flags at Level 1 Assisted and team-level strain indicators. It cannot carry the attention of a manager who knows the team, so the duty becomes a requirement on the organisation.
Chapter 25	1D · 2A · 3C (1) Each consequence decision can be contested individually, so every one receives a record. M1 Direction and M2 Evaluation decisions are sampled, apart from those the profile names and those that are contested. (2) The test passed with conditions: the deputy had never planned a route by hand and competence rules existed only inside the product. A freeze is not a stop; the system runs at the agreed degree while further delegation waits. (3) Minimisation applies to the record: it uses pseudonymous references and holds no health data even where health was the reason, because the record travels to readers who should not see it.
Chapter 26	1B · 2D · 3A (1) The record names no overseer and tests the design. AIDR-0630 blocks approval until each rest-rule flag has its own resolution, plans two hours per roster and requires training before a unit goes live. (2) Each consequence decision concerns one person and must be real in every case. Oversight of the system is added to it, never substituted for it; sampling is how Level 3 Bounded is made safe for M1 Direction. (3) Recruiters rejected applications below the cut in a median of 38 seconds against a floor of 120, and only three of six seeded applications were caught. The remedy closes postings at readable volume and shows the application before the rank.
Chapter 27	1C · 2B · 3D (1) The data was corrected to address-level coordinates and the rule changed to a minimum travel time written as a test. The rest-rule flag did not fire because no working-time rule was breached. (2) A contestation is a finding reported by the person best placed to find it, and a finding is closed only when the re-test passes. CNT-NR-0001 stayed open until the travel-time test was verified. (3) The 15-working-day answer limit for pay inputs across several national arrangements is justified only because the pay effect stays provisional; a contest is also never an input to any M2 Evaluation of the person.
Chapter 28	1A · 2C · 3B (1) The notice separates inputs used from inputs never used, tracing to the prohibited inputs of DLG-NR-ROS-01. Inputs are what a worker can check, and the exclusions answer whether declining a shift counts against them. (2) Class 1 covers a new system or function, a higher degree, a changed objective or a new unit. Class 3 covers changes that alter no notice field. The minimum travel time after CNT-NR-0001 was handled as Class 2. (3) The trade secrets directive treats workers' information and consultation rights as lawful acquisition, and the notice needs principles, main parameters, inputs, degrees and deciders, none of which is source code.
Chapter 29	1D · 2B · 3A (1) His diaries showed about a day a fortnight saved. The charter shows where it went: home visits with staff rose from about 15 to 25 per cent of his time, protected by a 35 per cent floor for M4 and M5. (2) The number of agents is set by the human attention available to verify them, and automation bias rises under multiple-task load. Widening a span is therefore a recorded decision with a baseline and review date. (3) The mandate manifest, verification with seeded errors and the role as an artefact transfer to every line manager. But because the manager's systems act on people, leading those people is the larger half of the transformation.
Chapter 30	1A · 2C · 3D (1) First acceptance rewarded whoever could answer a phone fastest and favoured staff without children at home. A fixed rotation that draws on nothing about the person is not a ranking, so the award stays outside Article 22. (2) AIDR-0630 had tied training to units going live, not to the approver role. AIDR-0636 requires the module before anyone approves a roster and makes deputies who plan a week by hand first choice as stand-ins. (3) Nordvik added units and constraints without a higher degree; Hansa kept its degrees and wrote in more of the conditions that make them true. Both kept M3 Consequence below Level 2 Supervised throughout.
Chapter 31	1B · 2D · 3A (1) In each case the record, the decision right or the contest path that would have shown the movement was missing. M4 Development and M5 Care appear as the slow harms in the composite cases. (2) Part-time staff with caring duties then received fewer offers. The rule the chapter draws is that a vendor release that changes who gets work is a new delegation, not an update. (3) The family's controls are the oversight record and seeded cases. Delegation unrecorded holds AP-01, AP-05, AP-10 and AP-12; people measured as machines AP-03 and AP-07; voice and craft lost AP-04, AP-06 and AP-09.
Chapter 32	1C · 2A · 3B (1) The first wave started before its managers were trained, and the reversibility test found that the deputy had never planned a route by hand. The Region found this within three months and froze further delegation. (2) Rest and travel-time blocks moved to Level 3 Bounded as protective blocks. Every degree touching a person's evaluation or consequences was held, lowered or refused. (3) An advisory score that nobody departs from is not advisory. The plan applies the lesson of CNT-HM-0001, where 97 per cent of proposals in one unit matched the score band.
Chapter 33	1D · 2B · 3A (1) In the illustrative year neither organisation passed its reversibility test cleanly. The chapter concludes that the test is what turns the register from an inventory into a control. (2) The earlier volumes did not need it because their subjects were systems. For the management of people, only a person can stand before a forum and answer for a consequence. (3) Without records a raise has no evidence. The template's last check is a Management Decision Record read end to end by a worker representative, because the record exists for the person it describes.

Final assignment

A Governed Design for One System That Manages People. Choose one AI system that performs part of a management function over people, in your own organisation or at Nordvik Region or Hansa Mutual as the book describes them. Produce the set of artefacts that would let the organisation delegate to it accountably: what it may do in each of the five management functions and at what degree, who decides each consequence, how each decision is recorded, how oversight is shown to be real, how a worker learns of the system and contests a decision, and how the organisation would know the delegation can be undone. Every statement about law is framed as an organisational implication, and nothing in the design monitors people covertly or bypasses the workers' representatives.

Assessment criteria

Every M3 Consequence decision has a named human decider, no adverse consequence exceeds Level 1 Assisted, and Level 4 Autonomous is not used for any consequence
The five management functions, degrees and usage modes are named exactly, and each action is assigned to a function by its effect on the person
The artefacts are consistent with one another and specific enough to be tested from records, with figures reported at team level and never used to evaluate individuals
The worker's notice and appeal path can be understood and used by the people the system manages, and the representatives' role is stated at each step
Claims are separated into observable, expected and speculation, and legal content is framed as organisational implication with the instrument and article named