Book Blueprint and Chapter Briefs

This blueprint records the editorial intent of AI Boss: When Artificial Intelligence Manages People in European Enterprises and Public Authorities, Book 7 of The Architecture as Code Series. It is the brief every chapter is written against. The chapter briefs below are generated from scripts/outline_data.py by python3 scripts/generate_outlines.py; edit the data file, not the generated section. The conclusions of the earlier volumes that each chapter builds on are recorded in SERIES_FOUNDATIONS.md (F1 to F80) together with the open questions the book takes up (Q1 to Q12).

Premise

The first six volumes of the series made architecture, governance, the AI platform and cyber defence executable, examined whether the AI threat is real, and described what has happened to the security architect. AI Threat asked whether the dystopias are coming true at every scale from the individual mind to the end of the world, and found the quiet harms at the scales of the mind and the organisation the most pressing: the erosion of skill, the automated state that could not be corrected, surveillance and profiling, and the gradual disempowerment that follows a long sequence of reasonable delegations. It did not ask the question this book asks: what happens when AI takes the manager's chair.

That question is no longer hypothetical. Algorithmic management is observable: platforms dispatch, rate and deactivate workers; warehouses measure pace and time off task; recruitment systems screen, score and interview; productivity scores feed appraisals and dismissals; rostering systems allocate shifts and pay; and, since 2024, AI agents in the office assign tasks, assess output and summarise colleagues, while software vendors tell every employee to become an "agent boss". Europe has answered in part: the AI Act classifies AI in employment and worker management as high-risk, prohibits emotion recognition at the workplace and obliges deployers to inform workers; the Platform Work Directive regulates algorithmic management for platform workers; data protection and co-determination law constrain what an employer may delegate to a machine. What no instrument settles is how an organisation should decide which of the manager's functions to delegate, to what degree, and how it answers for the result.

This book asks that question for European enterprises and public authorities. It answers through what the AI boss does today (Part II), what being managed by a machine does to people (Part III), what happens when people manage agents and agents manage agents (Part IV), what Europe has decided (Part V) and how to design an accountable AI boss (Part VI), framed by Part I and judged in Part VII.

Thesis

The thesis, stated in Chapter 1 and tested in Chapters 2 to 32: AI may direct work, but it may not, on its own, decide about people. Management splits into five functions. The functions that assign and pace work can be delegated at a degree matched to reversibility. The functions that carry consequences for a person, pay, promotion, discipline, dismissal and access to work, always have a named human decider who can be identified, contested and held to account. Every management decision that AI makes or shapes about a person is a record the person can see, contest and have corrected. What is delegated is itself recorded, so that the sum of delegations is visible before it becomes irreversible. And the manager's role is transformed, not abolished: from supervisor of tasks to accountable leader of people and conductor of agents.

Chapter 33 judges the thesis and states what would have shown it to be wrong: evidence that fully automated consequence decisions produce fairer and safer outcomes than human decisions with records and appeal, or evidence that algorithmic direction without human leadership sustains performance and wellbeing over years. The series' rule that Level 4 Autonomous is never used for irreversible actions becomes, in this book, the rule that Level 4 is never used for an M3 Consequence decision about a person.

Audience

Line managers, team leaders and heads of function whose work AI is beginning to do
Chief human resources officers, HR business partners and people-analytics leads
Executives and management bodies who decide what to delegate
Works councils, union representatives, safety representatives and employee-relations leads
Procurement and programme leaders in public authorities that buy worker-management systems
AI architects, platform and product teams that build or integrate management functions into AI systems
Risk, compliance, data-protection and internal-audit functions, and the supervisors who oversee them

Boundaries

About management, not a repeat of the earlier volumes. Platform construction, AI governance, AI security, AI-enabled defence, the AI threat and the security architect's role are summarised where this book needs them and referenced by chapter (see SERIES_FOUNDATIONS.md). This book refers to Books 1 to 6 and never to a later volume.
Defensive and constructive. The book contains no instructions for covert monitoring, for defeating works-council rights or for evading labour law. Monitoring is described at the level needed to assess and govern it.
Europe first. The AI Act, the GDPR, the Platform Work Directive, the Charter of Fundamental Rights, the occupational-safety and equal-treatment directives, national co-determination law and public-sector employment are the setting. Legal content is attributed to the instrument and article, verified against the live text, and framed as organisational and architectural implications, not legal advice.
No anthropomorphism. A system ranks, assigns, flags, scores or recommends; "decides" is used only for a system to which a decision was formally delegated, and the record then says who delegated it.
People described by what studies measured. Effects on autonomy, health, skill and voice are reported from the evidence, and nobody is called lazy, stupid or replaceable.
Findings and judgement kept apart. Every outlook is labelled observable, expected or speculation, and the author's own projections are kept to Chapters 29 and 33 and labelled.
No science-fiction tropes, no evidence ladder, no four shifts as red threads. Those belong to AI Threat and The Transformed Security Architect; this book refers to Book 6's Shift 4 where the conductor of an AI team is the subject, and otherwise uses its own red threads.

Red threads

The five management functions (Chapter 2), named exactly so in every chapter: M1 Direction (assigning, sequencing, pacing and instructing work); M2 Evaluation (measuring, rating and judging performance and conduct); M3 Consequence (hiring, pay, promotion, discipline, dismissal, deactivation and access to work); M4 Development (feedback, learning, coaching and career); M5 Care (safety, health, wellbeing, voice and the duty of care). Every chapter in Parts II to VI states which functions it concerns in its opening and its Summary; Part I defines them and Part VII concludes them.
Degrees of delegation (Chapter 4): the series' autonomy levels, Level 0 Advisory, Level 1 Assisted, Level 2 Supervised, Level 3 Bounded and Level 4 Autonomous (AI Red Team, Chapter 6), applied per management function, and recorded in a delegation profile (DLG-*) for each system or team. Level 4 is never used for M3 Consequence. The usage modes, Mode 0 Ambient AI to Mode 5 Autonomous production agents (AI Governance, Chapter 2; AI Platform and Security, Chapter 2), describe how the AI boss is deployed and whose identity it uses; every narrative chapter has an ### Across the usage modes or ### Across the degrees of delegation subsection.
The constructive debate (Chapters 11 to 15): the case for concern, the case for calm, where the positions meet, what can be done, then a verdict, as in AI Threat.
Management decisions as records (Chapter 25): the Management Decision Record (MDR-*; Nordvik Region MDR-NR-0001 onwards, Hansa Mutual MDR-HM-0001 onwards), the delegation register of management functions (management/delegation-register.yaml, extending the delegation register of AI Threat, Chapter 26), the oversight record and oversight-quality metric (Chapter 26), the contestation record (CNT-*; Chapter 27), the worker information notice (Chapter 28), the mandate manifest for agents a manager directs (MND-*; Chapter 16, reusing The Transformed Security Architect, Chapter 13), AI Decision Records in this book's range AIDR-0600 to AIDR-0699 (AIDR-0600 to AIDR-0649 Nordvik Region, AIDR-0650 to AIDR-0699 Hansa Mutual), Architecture Decision Records, and AI threat register entries (THR-*) where a chapter adds a management scenario.
Observable, expected, speculation (AI Governance, Chapter 22): every outlook is labelled, and the author's own projections are confined to Chapters 29 and 33.
The two organisations (EXAMPLE_ORGANISATIONS.md, people-management profile): Nordvik Region, about 6,000 staff running two local hospitals and home care for five municipalities, whose HR director Ingrid Halvorsen is commissioned to govern algorithmic management (AIDR-0600) and whose home-care unit manager Mikael Strand leads about 35 staff rostered by the procured ShiftWise system; and Hansa Mutual, whose claims-triage agent allocates work to adjusters, whose Hamburg claims team leader Jana Petersen directs about 20 handlers and a pilot set of agents, and whose Konzernbetriebsrat chair Thomas Brandt holds the co-determination rights under BetrVG; the executive board's request is recorded as AIDR-0650. Every narrative chapter has a ### Nordvik Region and Hansa Mutual subsection before its summary; Chapter 30 follows the four people through a year, Chapter 32 concludes for each organisation. The shared timeline, against which the subsections are checked, is recorded in the profile and fixed by Chapter 30. At Nordvik the year runs from the screening pause (AIDR-0601, 2026-09-18) and the commission (AIDR-0600, 2026-11-16) through the profile agreed with the unions on 2027-01-20, the first roster MDR-NR-0001 published on 2027-01-22, CNT-NR-0001 and its travel-time rule (2027-01-26 to 2027-03-25), the appeal path AIDR-0631 (2027-03-15), Ingrid Halvorsen's report of 2027-06-30 and version 1.1.0 of DLG-NR-ROS-01 from the review of 2027-08-01, to home-care wave 2 on 2027-09-13. At Hansa it runs from the board request (AIDR-0650, 2026-11-23) and the reclassification of allocation with DLG-HM-CTA-01 v1.0.0 for Germany only (2026-12-14) through MDR-HM-0001 (2027-01-12), the reversibility test of March, DLG-HM-REC-01 versions 1.4 and 1.5 (March and April), the first report and DLG-HM-CTA-01 v1.2.0 on 2027-04-30 and the role charter of 2027-07-01, to the extension of MND-HM-CLM-01 to 2028-06-30 (AIDR-0686, November 2027).
The open questions Q1 to Q12 (SERIES_FOUNDATIONS.md), raised in Chapter 1 and answered, answered in part or left open in Chapter 33.

Status

The book was started on 2026-11-06 from the engine of The Transformed Security Architect, the table of contents (scripts/toc_definition.py), these briefs, SERIES_FOUNDATIONS.md, the people-management profiles of the two organisations and a seed bibliography. All 33 chapters now exist as first drafts, reviewed for sources and for continuity (identifiers, dates, versions and the timeline of Chapter 30); the glossary and Appendix A are drafted. The status of each chapter is recorded in its brief below (outline, draft, complete). A chapter moves from outline to draft when its prose, diagram and worked examples exist and its tests pass, and to complete when it has been reviewed for sources, continuity and the red threads.

Length and weighting

Numbered narrative chapters are 3,600 to 5,500 words of prose, as in the other volumes. Chapters 2, 4, 20, 25 and 26 are expected to be the longest, because they carry the five functions, the degrees of delegation, the decision rights in the management chain, the records and the oversight on which the rest of the book depends. Chapters 30 and 32 are narrative and conclusion chapters and may be shorter. Any weighting decision taken during drafting is recorded in the brief.

Chapter briefs

Part I: The Question

Chapter 1: Introduction: Who Is the Boss?

File: 01_introduction.md
Status: draft
Purpose: Introduce the book on its own terms: algorithmic management is observable today, in platform dispatch and deactivation, warehouse rate monitoring, automated hiring screens, productivity scores, algorithmic scheduling and, since 2024, AI agents in the office that assign tasks and assess output, while people in turn are asked to become the managers of agents. State the question the earlier volumes did not ask, what happens when AI takes the manager's chair, and the working thesis: AI may direct work but may not, on its own, decide about people; every management decision AI makes or shapes is a record the person can see and contest; the manager's role is transformed, not abolished; and the quiet harms deserve attention before the loud ones. Name the five management functions, the degrees of delegation and the usage modes that carry the book, say what would show the thesis to be wrong, and explain why the question is a European organisational question in 2026, with the AI Act, the Platform Work Directive and co-determination law in force or pending.
Builds on: F70, F69, F66, F71, F75, F21, Q1, Q12
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M4 Development; M5 Care
Sections: Who is the boss?; The AI boss is not science fiction; The question the series had not asked; The working thesis; Five functions and five degrees; Why Europe, and why now; Rules of the book; Two organisations; How to read this book
Practical guides: A first inventory of the AI that already takes part in managing people in an organisation, including Mode 0 features nobody decided on
Worked examples: The HR director's commission at Nordvik Region recorded as AIDR-0600 and the Hansa Mutual executive board's request for a group policy on AI in management recorded as AIDR-0650, both written as decision records with reviewers and a report date
Nordvik Region and Hansa Mutual: Nordvik's HR director, Ingrid Halvorsen, is commissioned on 16 November 2026 to govern algorithmic management across the Region's about 6,000 staff (AIDR-0600), with the AI governance lead, the data protection officer and the chair of the central work-environment committee as reviewers and a report due by 30 June 2027. The commission follows the pause of the recruitment-screening pilot (AIDR-0601) and the planned go-live of ShiftWise routes in home care. A week later Hansa's executive board asks the chief HR officer and the chief risk officer for a group policy on AI in management, inviting the Konzernbetriebsrat to the review (AIDR-0650), because the claims-triage agent already allocates work to people and the agent-boss pilot starts in January. The chapter introduces Mikael Strand, Jana Petersen and Thomas Brandt, whose year Chapter 30 follows.
Usage modes and degrees of delegation: The chapter introduces the usage modes as the series uses them and the book's distinction between AI that assists a manager with the manager's identity (Modes 1 to 3), AI applications built or procured to perform management functions (Mode 4) and agents that direct work with identities of their own (Mode 5); it introduces the degrees of delegation as the autonomy levels applied per management function.
Observable, expected, speculation: Platform deactivation, warehouse rate systems, automated hiring screens and office agents that assign tasks are observable; the spread of AI-directed work across European knowledge work is expected; the thesis is the book's own claim, tested in Parts II to VI and judged in Chapters 29 and 33.
Figure: The AI boss at the centre of the book: the five management functions around a person, the degrees of delegation as rings, and the parts of the book that treat each function

Chapter 2: What a Manager Does: Five Management Functions and Who Performs Them

File: 02_what_a_manager_does.md
Status: draft
Purpose: Define what a manager does, as the frame on which every later chapter rests, and divide it into five management functions that can each be delegated, measured and governed separately: M1 Direction, M2 Evaluation, M3 Consequence, M4 Development and M5 Care. Draw on the classical and empirical literature of management, from Fayol's functions and Drucker's practice to Mintzberg's observation that managerial work is fragmented, interpersonal and informational, and on the job-design and psychological-safety research that explains why the manager matters to the people managed. Show who performs each function today in a hospital ward, a home-care unit and a claims team, which parts already sit in systems, and where the functions overlap, so that a reader can map any AI system onto the functions it touches. Explain why the split is not arbitrary: the functions differ in reversibility, in the law that governs them and in what a person loses when they go wrong.
Builds on: F9, F22, F17, F24, F73, Q1, Q11
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M4 Development; M5 Care
Sections: What managers were said to do; What managers were observed to do; Five management functions; Why the split matters; Where the functions overlap; Who performs each function today; The manager as a person in the system; Mapping an AI system onto the functions
Practical guides: Mapping a team's management functions: who performs each today, which parts sit in systems, and which an AI system now touches
Worked examples: A function map as code, management/function-map.yaml, for the Strandvik home-care unit at Nordvik Region and Jana Petersen's claims team at Hansa Mutual
Nordvik Region and Hansa Mutual: In the Strandvik home-care unit, Mikael Strand performs most of M1 Direction through visit schedules, M2 Evaluation through informal observation and the time-per-visit dashboard, M3 Consequence through recruitment and the annual pay review under the collective agreement, M4 Development through development conversations recorded without ratings, and M5 Care through the work-environment duties he shares with the safety representative. In Jana Petersen's Hamburg claims team of about 20 handlers, the claims-triage agent already performs much of M1 Direction by allocating claims within allocation rules, the throughput figures carry part of M2 Evaluation under KBV-IT-2024, and HR and the annual objectives cycle carry M3 Consequence. The two function maps show how differently the same five functions are already distributed between people and systems.
Usage modes and degrees of delegation: Mode 0 features in HR suites and Mode 2 office assistants touch the functions without anyone deciding it; Mode 4 systems such as ShiftWise and TalentLens are procured or built for a named function; Mode 5 agents such as the claims-triage agent perform a function with identities of their own. The function map records the mode for every system it lists.
Observable, expected, speculation: The management literature and the current distribution of functions are observable; the claim that the five functions are the right unit for delegation is the book's own and is tested in Chapters 4 and 20.
Figure: The five management functions as segments around the managed person, each annotated with its reversibility, its governing law and the systems that already touch it in a care unit and a claims team

Chapter 3: From Scientific Management to Algorithmic Management: A Century of Managing by Measurement

File: 03_from_taylor_to_the_algorithm.md
Status: draft
Purpose: Place the AI boss in a century of managing by measurement, so that the reader can tell what is new from what is old. Trace the line from Taylor's time-and-motion studies and the separation of planning from doing, through Braverman's deskilling thesis and Edwards's account of technical and bureaucratic control, the human-relations and sociotechnical counter-movements, call-centre and warehouse metrics, to the platform algorithm and the office agent. Show that algorithmic management inherits Taylorism's ambition to make work measurable and directable, and adds three things: direction and evaluation at the scale and speed of software, opacity to the people managed, and the possibility of consequence without a human decision. Use the measurement literature, from Campbell's law to the tyranny of metrics, to explain why what is measured comes to replace what was meant. End with the lessons the history offers and the ones it cannot.
Builds on: F32, F66, F69, F12, F21, Q4
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M4 Development; M5 Care
Sections: The one best way; Deskilling and control; The counter-movements; The control revolution; Call centres, warehouses and the metric; The algorithm as manager; What is new; When the measure becomes the target; What history teaches and what it cannot
Worked examples: A comparative table of control regimes, from foreman to agent, with what each measured, who decided and how a worker could contest it
Nordvik Region and Hansa Mutual: Nordvik's home-care service has measured visits for a decade through planned visit times agreed with the municipalities; the visit-registration app and the time-per-visit dashboard turn that measure into a figure per person and route, which is the step from bureaucratic to algorithmic control. Hansa's claims operation has long measured handling time and backlog per team; the claims-triage agent adds allocation by software and throughput figures per handler, which KBV-IT-2024 confines to team leaders. Both organisations can therefore see their own history in the chapter, and both can see where a familiar measure becomes a new form of control when an AI system computes it per person and acts on it.
Usage modes and degrees of delegation: Earlier control regimes had no usage modes; the chapter shows that Mode 4 systems reproduce bureaucratic control in software, while Mode 5 agents add direction with an identity of its own, which has no historical precedent beyond the automated assembly line.
Observable, expected, speculation: The history and the measurement literature are observable; the claim that algorithmic management is Taylorism plus three new properties is the book's interpretation and is tested in Part II.
Figure: A timeline of managing by measurement from the stopwatch to the agent, with the control regime, the measure and the worker's means of contest at each step

Chapter 4: The AI Boss Defined: Management Functions, Degrees of Delegation and Accountability

File: 04_the_ai_boss_defined.md
Status: draft
Purpose: Define the AI boss precisely and give the book its working instruments. An AI boss is any system or agent that performs part of a management function for or over a person; algorithmic management is the practice. Apply the series' five autonomy levels per management function as degrees of delegation, Level 0 Advisory to Level 4 Autonomous, introduce the delegation profile that states the level per function for a system or team, and state the rule that Level 4 is never used for M3 Consequence. Set out the accountability problem in the terms of the accountability literature, the problem of many hands and the moral crumple zone, and answer it with a named human decider for every consequence, a record for every management decision AI makes or shapes, and a delegation register that shows the sum of what has been delegated. Preview the artefacts of Part VI, so that Parts II to V can use them as a lens.
Builds on: F50, F51, F75, F69, F38, F8, F19, F63, Q1, Q2
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M4 Development; M5 Care
Sections: A definition; Degrees of delegation; Matching degree to function; The rule for consequences; The accountability problem; Delegation profiles; The delegation register; The artefact set; Reading Parts II to V through the instruments
Practical guides: Writing a first delegation profile for an AI system that touches management functions
Worked examples: The delegation profiles DLG-NR-ROS-01 for Nordvik's ShiftWise and DLG-HM-CTA-01 for Hansa's claims-triage allocation, as code, with level per function, named decider and review date
Nordvik Region and Hansa Mutual: Nordvik's DLG-NR-ROS-01 sets ShiftWise at Level 2 Supervised for M1 Direction, with unit managers approving every roster, one Level 3 Bounded class for offering open shifts to staff who opted in, and Level 1 Assisted for the M5 Care rest-rule checks; it is negotiated with the unions before go-live in home care and recorded as AIDR-0602. Hansa's DLG-HM-CTA-01 records the claims-triage agent's allocation of claims to handlers at Level 3 Bounded within allocation rules, with the team leader able to reassign, and throughput figures at Level 1 Assisted for M2 Evaluation. Writing the two profiles shows Nordvik that nothing in its estate reaches M3 Consequence once screening is paused, and shows Hansa that TalentLens and the sales score do, which Chapters 7 and 8 take up.
Usage modes and degrees of delegation: The degree of delegation is set per function, not per mode, but the mode constrains it: Mode 1 to Mode 3 tools act with the manager's identity and cannot exceed the manager's own authority; Mode 4 applications act within their configuration; Mode 5 agents act with identities and permissions of their own, which is where Level 3 Bounded and the rule against Level 4 for M3 Consequence matter most.
Observable, expected, speculation: The autonomy levels and delegation registers of the earlier volumes are observable practice in the two organisations; their application per management function is the book's proposal, tested in Chapters 20 and 25 to 27.
Figure: A matrix of the five management functions against the five degrees of delegation, with the cell for Level 4 in M3 Consequence closed, and the delegation profiles of the two organisations plotted on it

Part II: The AI Boss Today

Chapter 5: Platform Work: The Algorithm as Dispatcher

File: 05_platform_work.md
Status: draft
Purpose: Examine platform work, where the algorithm as dispatcher is most fully developed and best documented: ride-hailing, food delivery and on-demand services in which software allocates tasks, sets prices and pay, rates workers through customer ratings and deactivates accounts. Describe what the research on drivers and couriers measured, the information asymmetries, the soft control of nudges and ratings, and the wage-setting practices that have been called algorithmic wage discrimination. Set out what European courts and supervisory authorities found in the Foodinho, Deliveroo and Uber cases, from fines under the GDPR to the right to human intervention in deactivation, and what the Spanish Rider law and the Platform Work Directive now require. Draw the lessons for employers outside the platform economy, whose shift-offer and allocation systems increasingly work the same way, and for M3 Consequence when access to work is switched off by software.
Builds on: F21, F31, F57, F69, Q2, Q3
Management functions: M1 Direction; M2 Evaluation; M3 Consequence
Sections: The algorithm as dispatcher; What the research measured; Pay set by software; Ratings as evaluation; Deactivation; Courts and authorities; The Rider law and the Platform Work Directive; Lessons beyond the platform
Practical guides: Testing an internal shift-offer or allocation system against the Platform Work Directive's algorithmic-management chapter, even where it does not apply
Worked examples: An assessment record of Nordvik's open-shift offers in ShiftWise against the Directive's provisions on transparency, human monitoring and human review
Nordvik Region and Hansa Mutual: Neither organisation runs a platform, but both use platform logic. Nordvik's ShiftWise offers open shifts to hospital and home-care staff who opted in, at Level 3 Bounded, which resembles platform dispatch: whoever accepts first gets the shift, and acceptance history could become a hidden rating. The HR director, Ingrid Halvorsen, asks whether acceptance rates are stored and used, and the answer shapes the transparency notice of Chapter 28. Hansa's claims-triage agent allocates claims to handlers at Level 3 Bounded, much as a dispatcher allocates rides, and Jana Petersen's team asks whether its allocation rules favour fast handlers. Both organisations decide to test their systems against the Platform Work Directive's provisions as a benchmark, although the Directive applies to digital labour platforms rather than to them.
Usage modes and degrees of delegation: Platform dispatch is a Mode 4 or Mode 5 system acting at Level 3 Bounded or Level 4 Autonomous for M1 Direction; the cases show what happens when the same system also reaches M3 Consequence through deactivation without a human decision, which the Directive now requires for decisions to restrict, suspend or terminate.
Observable, expected, speculation: Platform dispatch, rating and deactivation and the decisions against them are observable; the transposition of the Platform Work Directive by 2 December 2026 and its effects are expected; the spread of platform logic into employment is observable in shift-offer systems and expected in allocation agents.
Figure: The platform management loop of allocation, pricing, rating and deactivation, with the points at which the Platform Work Directive requires transparency, human monitoring and human review

Chapter 6: The Warehouse and the Shop Floor: Pace, Rate and Monitoring

File: 06_warehouse_and_shop_floor.md
Status: draft
Purpose: Examine algorithmic management on the warehouse floor and in other paced work: scanner-driven rates, time-off-task measures, automated warnings and the link between pace and injury. Describe what was reported and measured: the automated productivity terminations reported at Amazon fulfilment centres in 2019, the injury analyses of 2021 and the United States Senate committee report of 2024, the French data protection authority's 2024 fine for excessively intrusive monitoring of warehouse staff, and California's 2021 law on warehouse quotas. Connect the evidence to the job-demand and control model and to the employer's duty under the Framework Directive on safety and health at work to adapt work to the individual. Show how the same pattern appears in care visits and back-office queues, and how M1 Direction becomes M5 Care when the system sets the pace.
Builds on: F52, F53, F21, F66, Q11, Q4
Management functions: M1 Direction; M2 Evaluation; M5 Care
Sections: Rate and time off task; What was reported; The French decision; Quotas made visible; Pace and health; The employer's duty; The pattern elsewhere; Designing pace with people
Practical guides: Assessing a pace or rate system as a work-environment risk with safety representatives
Worked examples: A work-environment risk assessment of the visit-time targets in Nordvik's home-care routes, linked to the delegation profile DLG-NR-HCD-01
Nordvik Region and Hansa Mutual: In the Strandvik home-care unit, visit times planned in ShiftWise routes and registered on staff phones play the part of the rate: an assistant nurse's contestation CNT-NR-0001 of a route that left no time between two visits is the unit's first sign of pace pressure. Mikael Strand and the safety representative assess the planned visit times as a work-environment risk, and the dashboards stay at Level 0 Advisory under DLG-NR-HCD-01. At Hansa, the Polish service centre, with about 1,400 of the group's claims staff, works in queues that resemble a warehouse more than an office; the claims-triage agent paces the queue, and throughput figures go to team leaders only under the group works agreement KBV-IT-2024. Neither organisation uses automated warnings, and both record that decision.
Usage modes and degrees of delegation: Rate systems are usually Mode 4 applications at Level 1 Assisted or Level 2 Supervised for M2 Evaluation; the danger arises when warnings or terminations follow automatically, which is Level 4 in M3 Consequence and excluded by the book's rule.
Observable, expected, speculation: Rate monitoring, the reported terminations, the CNIL fine and the California law are observable; the link between monitored pace and injury is reported in analyses whose methods the chapter states; extension of rate systems into care and office work is observable in visit-time systems and expected elsewhere.
Figure: The pace loop: target, measurement, warning and consequence, with the points where a safety representative, a human reviewer and the work-environment risk assessment intervene

Chapter 7: Hiring by Machine: Screening, Scoring and Automated Interviews

File: 07_hiring_by_machine.md
Status: draft
Purpose: Examine hiring by machine: targeted job advertising, CV screening and ranking, assessments, chatbots and automated video interviews. Describe what is documented about bias in automated hiring, from Amazon's abandoned recruiting tool to the iTutorGroup settlement and the Workday litigation, against the older field evidence of discrimination by human recruiters and the evidence that algorithms can sometimes outperform human discretion. Explain why the AI Act lists recruitment and selection in Annex III point 4(a), why emotion recognition in interviews is prohibited by Article 5(1)(f), how New York City's audit law compares, and what co-determination over selection guidelines under BetrVG § 95 means for a screening model. Treat hiring as M3 Consequence, because access to work is decided, and show how a recruiter keeps the decision and the record rather than signing a ranking.
Builds on: F16, F31, F38, F44, F25, Q2, Q3
Management functions: M2 Evaluation; M3 Consequence
Sections: The automated funnel; What went wrong, documented; Human discretion as the baseline; Audits and their limits; Annex III point 4(a); No emotion recognition; Selection guidelines and co-determination; Keeping the decision with the recruiter
Practical guides: Introducing, pausing or retiring a screening model with a record a candidate, a union and a supervisor can read
Worked examples: The delegation profile DLG-HM-REC-01 for TalentLens with affect and voice analysis switched off, and Nordvik's pause record AIDR-0601 with the questions the union asked
Nordvik Region and Hansa Mutual: Nordvik piloted a screening module of its applicant-tracking system for assistant-nurse and care-assistant vacancies in spring 2026 and paused it in September when the union for care staff asked how the ranking was made and the supplier could not show it (AIDR-0601). The chapter uses the pause as a decision worth recording: the Region could not explain a high-risk system it deployed, so it stopped. Hansa uses TalentLens in Poland for service-centre roles and in Germany for graduate and apprentice hiring, at Level 1 Assisted for M3 Consequence under DLG-HM-REC-01; the supplier's affect and voice analysis is switched off group-wide, and the group works agreement KBV-IT-2024 covers it. Thomas Brandt asks for the selection criteria as a § 95 guideline, which turns the configuration into a co-determined document.
Usage modes and degrees of delegation: Hiring systems are typically procured Mode 4 applications at Level 1 Assisted; chat-based screening and agentic sourcing tools push towards Mode 5, and any system that rejects candidates without a recruiter reaching a decision is Level 4 in M3 Consequence and excluded.
Observable, expected, speculation: Automated screening, the documented cases and the AI Act's classification are observable; the application date of the high-risk obligations, as amended by the Digital Omnibus, is to be verified by the writer; agentic sourcing and interviewing at scale are expected.
Figure: The hiring funnel with each stage marked by function, degree of delegation and the record a recruiter keeps, and the prohibited affect-analysis stage removed

Chapter 8: Performance Scores, Productivity Metrics and Dismissal by Algorithm

File: 08_performance_and_dismissal.md
Status: draft
Purpose: Examine performance scores, productivity metrics and dismissal by algorithm: individual scores computed from activity data, rankings used in pay and promotion, automated warnings and terminations. Describe what was reported in the Amazon Flex deactivations, the Amsterdam judgments on automated dismissal of drivers, the backlash against Microsoft's productivity score in 2020, and, as a cautionary case of trust in a system over people, the Post Office Horizon litigation and inquiry. Use the feedback and measurement literature to explain why a score that drives consequences invites gaming and misreads work, and the CJEU's SCHUFA judgment to explain when a score that decisively shapes a decision falls under GDPR Article 22. Show how M2 Evaluation becomes M3 Consequence through a pay input, a warning or a list, and how a score is kept advisory in practice and not only on paper.
Builds on: F2, F21, F31, F52, F38, Q2, Q3
Management functions: M2 Evaluation; M3 Consequence
Sections: The score; What scores miss; Dismissal by algorithm, reported; The productivity-score backlash; Trusting the system over the person; Scores and Article 22; From evaluation to consequence; Keeping a score advisory
Practical guides: Testing whether a performance score is decisive in practice: overrides, agreement rates and the time managers spend before acting
Worked examples: The delegation profile DLG-HM-SPS-01 for Hansa's sales-performance score as one input to variable pay, and the contestation record CNT-HM-0001 of a quarterly score
Nordvik Region and Hansa Mutual: Hansa's sales-performance score rates sales staff on conversion, retention and complaint indicators, and sales managers use it as one input to variable pay at Level 0 Advisory for M3 Consequence under DLG-HM-SPS-01. A sales employee contests a quarterly score (CNT-HM-0001), and the chapter uses the case to test whether the score is advisory in practice: how often managers depart from it, and whether they can say why. Nordvik's home-care dashboards show planned against actual time per visit; individual figures are visible to the unit manager, and the HR director decides that they may not be used in the annual pay review or in any disciplinary matter, recorded as a condition in DLG-NR-HCD-01. Neither organisation uses AI in discipline or dismissal, and both record that as a decision rather than an absence.
Usage modes and degrees of delegation: Scores are usually computed by Mode 4 applications or Mode 0 analytics in licensed suites; the degree of delegation is set by use, not by the software: a score read at Level 0 Advisory becomes Level 3 or Level 4 in practice when managers act on it without judgement, which the chapter shows how to detect.
Observable, expected, speculation: Individual productivity scores, the reported deactivations and the judgments are observable; the extension of scoring into knowledge work through office suites is observable and growing; regulatory limits on scores used in pay are expected as the AI Act's high-risk obligations apply.
Figure: How a score travels from M2 Evaluation to M3 Consequence: data, score, threshold, manager and decision, with the checkpoints that keep it advisory

Chapter 9: Scheduling, Pay and the Allocation of Work

File: 09_scheduling_pay_and_allocation.md
Status: draft
Purpose: Examine scheduling, pay and the allocation of work, where algorithmic management is most common outside platforms and least visible: rosters drafted from demand forecasts, open shifts offered by software, routes for mobile workers, and work items allocated to people by queue or agent. Describe what research measured about unstable schedules and worker health and wellbeing, the evidence on pay set or shaped by software, and the reporting on just-in-time scheduling. Set out the European constraints: the Working Time Directive and the CJEU's ruling that working time must be recorded, the Pay Transparency Directive, the AI Act's listing of task allocation based on individual behaviour or traits in Annex III point 4(b), and co-determination over working-time arrangements. Show how a roster or an allocation, though reversible, shapes earnings, family life and health, so that M1 Direction carries M3 Consequence and M5 Care inside it.
Builds on: F50, F19, F63, F69, Q1, Q8
Management functions: M1 Direction; M3 Consequence; M5 Care
Sections: Rosters and routes by forecast; Unstable schedules; Allocation by queue and agent; Pay shaped by software; Working time, recorded; Task allocation as high-risk; Pay transparency; The right degree of delegation
Practical guides: Negotiating a delegation profile for a rostering or allocation system with unions or a works council
Worked examples: DLG-NR-ROS-01 for ShiftWise and the first approved roster MDR-NR-0001 of the Strandvik unit; DLG-HM-CTA-01 and the first reassignment MDR-HM-0001 by a Hamburg team leader overriding the claims-triage agent
Nordvik Region and Hansa Mutual: ShiftWise drafts four-week rosters for Nordvik's two hospitals and visit schedules and routes for home care, at Level 2 Supervised, with one Level 3 Bounded class for open-shift offers to staff who opted in; its rest-rule checks flag breaches at Level 1 Assisted. The profile DLG-NR-ROS-01 is negotiated with the unions under the co-determination act before go-live in home care (AIDR-0602), and MDR-NR-0001, Mikael Strand's first approved roster, records what he changed and why. At Hansa, the claims-triage agent allocates claims to handlers at Level 3 Bounded; Jana Petersen's reassignment MDR-HM-0001 is the first record of a team leader overriding it. The chapter asks both organisations whether allocation affects earnings: at Nordvik through shift premiums, at Hansa only indirectly, through the objectives cycle.
Usage modes and degrees of delegation: Rostering is typically a Mode 4 application at Level 2 Supervised; allocation by a Mode 5 agent at Level 3 Bounded is reversible per item but cumulative in effect, which is why the profile records reassignment rights and a review of allocation patterns by person.
Observable, expected, speculation: Forecast-driven rostering, open-shift offers and agent allocation are observable in both organisations; agents that also negotiate swaps and premiums are expected; fully autonomous scheduling without approval is speculation in the regulated care setting.
Figure: The rostering and allocation loop from forecast to approved roster and allocated claim, with the decider, the degree of delegation and the record at each step

Chapter 10: Knowledge Work: The AI Manager Enters the Office

File: 10_the_ai_manager_in_the_office.md
Status: draft
Purpose: Examine the arrival of the AI manager in knowledge work: office assistants that summarise meetings and assign action items, workflow agents that allocate tickets and cases, analytics that report on collaboration and focus time, and the vision of every employee as an agent boss. Describe what was measured, the productivity studies of AI-assisted work, the usage data of AI providers and the evidence on labour-market effects so far, and what was promised, separating the two. Explain the transparency paradox, in which observed workers hide improvements, and the controversy over individual productivity figures in office suites. Show how Mode 0 and Mode 2 features quietly perform M1 Direction and M2 Evaluation, why works councils in Germany treat them as technical equipment suited to monitor, and how an organisation decides which office features may assign, summarise or report about people.
Builds on: F15, F49, F57, F66, F12, Q6
Management functions: M1 Direction; M2 Evaluation; M4 Development
Sections: The assistant that assigns; Workflow agents; Analytics about people; The agent boss vision; What was measured; The transparency paradox; Technical equipment suited to monitor; Deciding about office features
Practical guides: Building a feature register for office assistants and analytics: which features assign, summarise or report about people, at what degree, and with whose agreement
Worked examples: Hansa's decision to keep the wellbeing and collaboration-analytics feature switched off group-wide pending works-council agreement, recorded as a decision record in the AIDR-0650 to AIDR-0699 range, with Nordvik's staff-assistant action lists at Level 0 Advisory as the contrast
Nordvik Region and Hansa Mutual: Nordvik's internal staff assistant and the pilot of the embedded workplace assistant summarise meetings and propose action lists with named owners at Level 0 Advisory; the meeting chair assigns. The HR director asks whether anyone reads the proposed owners as instructions, and a sample of minutes answers it. Hansa's embedded office assistants do the same across the group, and the wellbeing and collaboration-analytics feature in the licensed office suite (Mode 0) is switched off pending works-council agreement. The agent-boss pilot HM-AB-PILOT-01 starts in January 2027 in Jana Petersen's claims team and an IT service-desk team, and the chapter sets up the question Part IV answers: whether people managing agents is the same activity as agents managing people.
Usage modes and degrees of delegation: In knowledge work the AI boss arrives through Mode 0 features nobody decided on, Mode 2 assistants with the user's identity and Mode 3 or Mode 5 workflow agents; the feature register records the mode and the degree for each feature that performs a management function.
Observable, expected, speculation: Assistants that assign action items and office analytics are observable; workflow agents that allocate knowledge work are observable in pilots and expected at scale; the claim that every employee becomes an agent boss is a vendor projection, labelled speculation until measured.
Figure: An office day in which assistant features assign, summarise and report, with each feature placed on the five functions and its degree of delegation

Part III: What It Does to People

Chapter 11: Autonomy, Dignity and Control: Being Managed by a Machine

File: 11_autonomy_dignity_control.md
Status: draft
Purpose: Stage the first debate of Part III: what being managed by a machine does to autonomy, dignity and the experience of control at work. Present what studies of platform workers, warehouse staff and office workers measured about autonomy, meaningfulness, isolation and perceived fairness, and connect it to self-determination theory, the job-characteristics model and the job-demand and control model. Give the case for concern at its strongest, that algorithmic direction removes discretion and the human relationship through which people are recognised, and the case for calm at its strongest, that many workers prefer consistent rules to a capricious supervisor and that algorithms can be designed to widen choice. Show where the positions meet, on the design of discretion and recourse, and what individuals, managers and organisations can do. End with a verdict labelled observable, expected or speculation, and with the European anchors of dignity at work in the Charter and the Council of Europe convention.
Builds on: F66, F53, F69, F21, Q11, Q4
Management functions: M1 Direction; M2 Evaluation; M5 Care
Sections: Being managed by a machine; Autonomy, competence and relatedness; Demand and control; Algorithm aversion and appreciation; The case for concern; The case for calm; Where the positions meet; What can be done; Verdict
Constructive debate: Concern: Algorithmic direction removes the discretion through which people exercise skill and the relationship through which they are recognised, and studies of platform and warehouse workers report lower autonomy and meaningfulness. Calm: Many workers report that consistent rules are fairer than a capricious supervisor, some studies find algorithm appreciation, and systems can be configured to widen choice of shifts and tasks. Where they meet: Both positions agree that the effect depends on how much discretion the design leaves, whether reasons are given and whether a person can be reached. What can be done: Individuals: ask for reasons and use preference settings. Managers: keep discretion and the conversation. Organisations: design for choice and recourse and measure autonomy. Representatives: negotiate discretion into the delegation profile.
Practical guides: Measuring autonomy and perceived fairness before and after an AI system takes over part of M1 Direction
Worked examples: A short autonomy and fairness survey, agreed with the representatives, run in the Strandvik home-care unit before and after ShiftWise routes, with results reported at unit level only
Nordvik Region and Hansa Mutual: In the Strandvik unit, ShiftWise routes reduce the planning Mikael Strand used to do by hand, and assistant nurses gain the ability to state preferences and accept open shifts, but some lose the informal swaps they arranged with each other. The unit runs a short survey on autonomy and perceived fairness before and after go-live, agreed with the union and reported at unit level only. At Hansa, claims handlers in Jana Petersen's team describe the claims-triage agent's allocation as fairer than the old practice of taking claims from a shared queue, but experienced handlers miss choosing complex claims they enjoy. The team agrees a bounded preference rule within the allocation rules of DLG-HM-CTA-01, which the Konzernbetriebsrat welcomes as an example of discretion negotiated into the profile.
Usage modes and degrees of delegation: Autonomy is affected most by Mode 4 and Mode 5 systems at Level 3 Bounded for M1 Direction, where no person mediates the instruction; at Level 2 Supervised the manager can restore discretion, which the chapter shows is a design choice rather than a by-product.
Observable, expected, speculation: The measured effects on autonomy in platform and warehouse work are observable; effects in knowledge work under office agents are expected and under-researched; long-term effects on dignity and identity at work are speculation until measured over years.
Figure: The constructive debate on autonomy and dignity: the two cases, the common ground and the actions, with the verdict and its label

Chapter 12: Fairness, Bias and the Appeal That Never Comes

File: 12_fairness_bias_and_contestability.md
Status: draft
Purpose: Stage the debate on fairness, bias and contestability in algorithmic management. Present what is documented about bias in systems that evaluate and decide about people, from facial-analysis error rates and a health algorithm that used cost as a proxy for need to the Bologna ruling on a delivery platform's ranking and the gender gap measured among rideshare drivers, and what is documented about the appeal that never comes: decisions without reasons, appeals that reach no person, and public-sector cases such as SyRI and the Dutch childcare-benefits affair that show what happens when contest fails at scale. Give the case for concern, that algorithmic management scales unfairness and removes the person to appeal to, and the case for calm, that human managers are biased too and algorithms can be audited and corrected in ways people cannot. Show where they meet and what can be done, with explanation rights under GDPR Article 22 and AI Act Article 86, and end with a verdict.
Builds on: F31, F16, F44, F65, F38, Q3, Q2
Management functions: M2 Evaluation; M3 Consequence
Sections: Documented bias; Why bias enters; The appeal that never comes; When contest fails at scale; The case for concern; The case for calm; Where the positions meet; What can be done; Verdict
Constructive debate: Concern: Algorithmic management can reproduce historical bias at scale, through proxies nobody chose, and replaces the manager a person could appeal to with a process that gives no reasons. Calm: Human managers are documented to discriminate, algorithms can be tested for disparate impact and corrected, and a consistent rule can be fairer than many inconsistent ones. Where they meet: Both positions agree that fairness depends on measuring outcomes by group, giving reasons, and keeping a person who can be reached and who can correct the decision. What can be done: Individuals: use the right to an explanation and to contest. Organisations: test disparate impact on every release and keep contestation records. Representatives: ask for outcome data by group. Supervisors: enforce Articles 22 and 86.
Practical guides: Running a disparate-impact test on a management system with the representatives, and publishing what it found
Worked examples: A disparate-impact test of claims allocation by handler group at Hansa, reported to the Konzernbetriebsrat, and a fairness review of open-shift offers by contract type at Nordvik
Nordvik Region and Hansa Mutual: At Nordvik, the HR director commissions a fairness review of ShiftWise open-shift offers by contract type, because part-time and hourly staff in home care depend on extra shifts for their income and the offers are made at Level 3 Bounded; the review uses the Region's existing disparate-impact practice from the benefits-triage system. The paused screening pilot (AIDR-0601) is the Region's own example of an appeal that could not be answered. At Hansa, a disparate-impact test of claims allocation by handler group, age band and part-time status is reported to the Konzernbetriebsrat under KBV-IT-2024, and Thomas Brandt asks for the same test on the sales-performance score, after CNT-HM-0001 shows that the complaints indicator penalises staff in markets with higher complaint rates.
Usage modes and degrees of delegation: Bias and contestability matter most where M2 Evaluation and M3 Consequence are delegated beyond Level 1 Assisted; in Mode 0 analytics nobody may know a score exists, which makes contest impossible until the feature is registered.
Observable, expected, speculation: Documented bias cases and the public-sector failures are observable; enforcement of the right to an explanation under Article 86 is expected once the high-risk obligations apply; the claim that audited algorithms will become fairer than managers is speculation.
Figure: The constructive debate on fairness and contestability, with the contest path from decision to reason, review and correction and the points where it fails

Chapter 13: Surveillance at Work and the Chilling Effect

File: 13_surveillance_at_work.md
Status: draft
Purpose: Stage the debate on surveillance at work and its chilling effect. Describe the monitoring that algorithmic management requires and invites, from activity tracking and location data to keystroke and screen capture, and what was decided against it: the Hamburg data protection authority's 2020 fine for profiling of service-centre staff, the CNIL's 2024 warehouse decision, the CJEU's 2023 judgment on Article 88 and national employee-data rules, and the long German case law on technical equipment suited to monitor. Present what research measured about monitoring, from reduced theft and improved productivity in one restaurant study to the transparency paradox and the effect on voice and psychological safety. Give the case for concern, that pervasive monitoring chills voice and turns every act into evidence, and the case for calm, that proportionate, transparent monitoring is lawful, can protect workers and is negotiated in Europe. Show where they meet and what can be done, and end with a verdict.
Builds on: F52, F57, F31, F15, Q11, Q8
Management functions: M2 Evaluation; M5 Care
Sections: What the AI boss needs to see; What authorities decided; Article 88 and national law; Suited to monitor; What research measured; The case for concern; The case for calm; Where the positions meet and what can be done; Verdict
Constructive debate: Concern: Pervasive monitoring makes every act potential evidence, chills voice and experimentation, and gives the employer a profile of the person that no manager had before. Calm: Monitoring that is proportionate, transparent and negotiated is lawful in Europe, can protect workers' safety and pay claims, and studies find productivity gains without harm in some settings. Where they meet: Both positions agree that purpose, proportionality, aggregation and the workers' representatives' agreement decide whether monitoring protects or chills. What can be done: Individuals: know the notice and use access rights. Organisations: aggregate by default, limit purpose and never infer emotions. Representatives: negotiate access and retention. Supervisors: test proportionality.
Practical guides: A proportionality review of monitoring data used by a management system: purpose, aggregation, retention and access
Worked examples: The proportionality review of the visit-registration data behind Nordvik's home-care dashboards, and the conditions in Hansa's KBV-IT-2024 that confine individual throughput figures to team leaders
Nordvik Region and Hansa Mutual: Nordvik's visit-registration app on staff phones records when each home-care visit starts and ends, and the time-per-visit dashboard turns it into figures per route and per person; the data protection officer and the safety representative run a proportionality review, and the Region removes location traces between visits from the data the dashboard may use. Hansa's group works agreement KBV-IT-2024 confines individual throughput figures from the claims-triage agent to team leaders, and the wellbeing and collaboration analytics in the office suite stay switched off. Thomas Brandt asks how the Swiss entity, where the ordinance on health at work prohibits systems that monitor behaviour, and the Polish service centre, under the Labour Code's monitoring provisions, apply the same principle under different law.
Usage modes and degrees of delegation: Mode 0 features in licensed suites are the main source of unnoticed monitoring; Mode 4 and Mode 5 systems collect monitoring data as a by-product of directing work, so purpose limitation must be designed into the system that directs, not added afterwards.
Observable, expected, speculation: Workplace monitoring and the decisions against it are observable; tighter limits under the Platform Work Directive and the AI Act's prohibition of emotion recognition are observable in law and expected in enforcement; long-term effects on voice are under-researched and labelled expected where studies point that way.
Figure: The constructive debate on surveillance at work, with a monitoring data flow showing purpose limitation, aggregation and negotiated access as controls

Chapter 14: Health, Pace and Burnout: Occupational Safety Under Algorithmic Direction

File: 14_health_pace_and_burnout.md
Status: draft
Purpose: Stage the debate on health, pace and burnout under algorithmic direction. Present what was measured: the job-demand and control and effort-reward evidence, analyses of injury rates and pace in monitored warehouses, research on schedule instability and health, the WHO and ILO estimate of deaths attributable to long working hours, and European surveys of psychosocial risk and digital technology at work. Set out the employer's duty under the Framework Directive 89/391/EEC, national work-environment law such as the Swedish rules on organisational and social work environment, ISO 45003 and the Platform Work Directive's provisions on safety and health in algorithmic management. Give the case for concern, that systems that set pace optimise output and externalise strain, and the case for calm, that algorithms can enforce rest rules, balance workload and detect overload earlier than a manager. Show where they meet and what can be done, then give a verdict on M5 Care under algorithmic direction.
Builds on: F52, F53, F66, F19, Q11, Q4
Management functions: M1 Direction; M5 Care
Sections: Pace, strain and recovery; Monitored pace and injury; Schedules and health; The employer's duty; National rules and standards; The case for concern; The case for calm; Where the positions meet and what can be done; Verdict
Constructive debate: Concern: Systems that set pace optimise for output and push the cost of strain onto workers' bodies and minds, and the injury analyses and scheduling studies point the same way. Calm: Algorithms can enforce rest rules a busy manager overlooks, balance workload across a team and detect overload earlier, and nothing in the evidence shows harm is inherent in the software. Where they meet: Both positions agree that health outcomes depend on whether health constraints are inside the optimiser, whether pace is negotiated and whether someone with authority watches strain. What can be done: Individuals: report strain and use the right to stop work in immediate danger. Organisations: put rest and workload limits into the system and run risk assessments on every change. Representatives: review pace parameters. Authorities: inspect algorithmic pace as a work-environment risk.
Practical guides: Writing health and rest constraints into an optimiser, and testing that they hold
Worked examples: A work-environment risk assessment for ShiftWise routes in home care, with rest and travel-time constraints written as tests, reviewed by the central work-environment committee
Nordvik Region and Hansa Mutual: Nordvik's ShiftWise flags breaches of rest rules at Level 1 Assisted, and Mikael Strand resolves them; the contestation CNT-NR-0001 of a route with no time between visits leads to a minimum travel-time constraint, written as a test and reviewed by the central work-environment committee, whose chair is a reviewer of AIDR-0600. Home care, with the highest share of part-time and hourly staff and the highest turnover, is where the Region watches strain most closely. Hansa delegates nothing in M5 Care to AI and keeps the wellbeing analytics switched off; the chapter asks whether that is caution or neglect, since the claims-triage agent sets the pace of the queue in Hamburg and at the Polish service centre. Jana Petersen proposes a team-level workload indicator, which Thomas Brandt accepts only at team level.
Usage modes and degrees of delegation: Health is affected by Mode 4 and Mode 5 systems that set pace at Level 2 Supervised or above; the chapter argues that M5 Care itself should rarely be delegated beyond Level 1 Assisted, while health constraints must be inside every system that performs M1 Direction.
Observable, expected, speculation: The measured links between strain, schedules and ill health are observable; the effect of algorithmic pace specifically is observable in some sectors and expected in others; the effect of office agents on burnout is speculation until measured.
Figure: The constructive debate on health and pace, with the optimiser shown with and without health constraints and the safety representative's place in the loop

Chapter 15: Does It Work? Productivity, Quality and the Evidence for Algorithmic Management

File: 15_does_it_work.md
Status: draft
Purpose: Stage the debate on whether algorithmic management works. Present the evidence on productivity and quality: monitoring and theft in restaurants, algorithmic and human discretion in hiring, AI assistance in customer support and consulting, the randomised study of experienced developers that found a slowdown, the Danish labour-market evidence of small effects, employer surveys on the use and perceived benefits of algorithmic management, and the measurement literature on gaming and on metrics that crowd out quality. Distinguish short-run output from long-run performance, retention, safety and learning, and ask what would count as evidence over years rather than quarters. Give the case for concern, that algorithmic management raises measured output while eroding quality, trust and the pipeline of skill, and the case for calm, that it improves consistency, matching and speed. Show where they meet and what can be done, and end with a verdict on Q4.
Builds on: F66, F32, F53, F2, Q4, Q5
Management functions: M1 Direction; M2 Evaluation; M4 Development
Sections: What counts as working; Evidence from monitoring; Evidence from hiring; Evidence from AI-assisted work; Employer surveys; Gaming and the crowding out of quality; The case for concern and the case for calm; Where the positions meet and what can be done; Verdict
Constructive debate: Concern: Algorithmic management raises what is measured and erodes what is not: quality, trust, discretion and the pipeline of skill, and the long-run costs appear after the quarter in which gains are reported. Calm: Controlled studies show gains in consistency, matching and speed, largest for less experienced workers, and employers report better coordination and decisions. Where they meet: Both positions agree that most evidence is short-run, that outcomes depend on design and that balanced, long-run measures agreed in advance are missing. What can be done: Organisations: define success over years, include quality, retention and wellbeing, and run controlled roll-outs. Representatives: agree the evaluation design. Researchers and authorities: fund longitudinal studies.
Practical guides: Designing a multi-year evaluation of an AI system that performs management functions, with outcomes agreed with representatives
Worked examples: The evaluation plan for ShiftWise in home care with staggered unit roll-out, and the evaluation plan for the agent-boss pilot HM-AB-PILOT-01 with quality, retention and wellbeing outcomes
Nordvik Region and Hansa Mutual: Nordvik stages ShiftWise routes unit by unit in home care, starting with Strandvik, so that the units not yet using it serve as a comparison, in the way the Region measured its staff assistant against a later rollout wave; outcomes include missed visits, overtime, sick leave, turnover and staff-rated fairness, over at least two years. Hansa's agent-boss pilot HM-AB-PILOT-01 is evaluated against comparable claims teams on throughput, reopened claims, complaint rates, handler turnover and a team-level wellbeing measure agreed with the Konzernbetriebsrat. The chapter shows that both organisations had first proposed quarterly throughput as the measure of success, and why both changed it.
Usage modes and degrees of delegation: The evidence differs by mode: Mode 1 to Mode 3 assistance has controlled studies; Mode 4 management systems mainly employer surveys and case studies; Mode 5 agents that direct people almost no evidence yet, which is why the chapter asks for evaluation designs before scale.
Observable, expected, speculation: Short-run gains in assisted work and the gaming of metrics are observable; long-run effects of algorithmic management on performance and skill are expected to be mixed but are not yet measured; claims of large organisation-wide gains are speculation.
Figure: The constructive debate on whether algorithmic management works, with outcomes plotted over time: short-run output against long-run quality, retention and skill

Part IV: Managing Machines

Chapter 16: The Agent Boss: When the Team Is Software

File: 16_the_agent_boss.md
Status: draft
Purpose: Turn the question around: what happens when people become the managers of AI agents, the agent boss of the Work Trend Index. Explain what transfers from managing people to managing agents, giving instructions, setting scope, checking output, and what does not, motivation, development, care and the duty owed to a person. Show that an agent team needs a written mandate, a degree of delegation matched to reversibility, containment that holds whatever the agent does, and a manager who remains the accountable verifier of everything the agents produce, reusing the mandate manifest of The Transformed Security Architect, Chapter 13. Describe what is documented about agents acting beyond their mandate, from the Replit database deletion to the agentic-misalignment experiments, and what an experiment in letting an agent run a small shop showed about agents in management roles. Give every team leader a way to tell whether they are managing an agent team or being managed by it.
Builds on: F12, F27, F45, F49, F50, F75, F76, F67, Q6
Management functions: M1 Direction; M2 Evaluation; M4 Development
Sections: The agent boss; What transfers; What does not transfer; The mandate manifest; Agents beyond their mandate; Containment for a team's agents; Verifying what agents produce; Managing or being managed
Practical guides: Writing a mandate manifest for the agents in a team, and a weekly verification routine for the team leader
Worked examples: The mandate manifest MND-HM-CLM-01 for the document-intake agent and the recovery agent in Jana Petersen's claims team, with the delegation profile DLG-HM-AB-01
Nordvik Region and Hansa Mutual: From January 2027 Jana Petersen directs a document-intake agent and a recovery agent that drafts subrogation letters, alongside her about 20 claims handlers, under the agent-boss pilot HM-AB-PILOT-01 (AIDR-0652). The agents act at Level 2 Supervised under MND-HM-CLM-01 and DLG-HM-AB-01; every subrogation letter is approved by a handler before it is sent, and Jana is accountable for the agents' output. The pilot's rule is that people manage agents and agents do not manage people. Nordvik runs no agents that a manager directs, since Mode 3 tools are limited to the small platform team and the Region has no Mode 5 agents; Ingrid Halvorsen records that Nordvik's line managers will not be asked to direct agents until the Region has the supervision capacity, as a decision record in the AIDR-0600 to AIDR-0649 range.
Usage modes and degrees of delegation: Managing agents is a Mode 3 or Mode 5 activity: Mode 3 agentic work tools act with the manager's identity and Mode 5 agents with their own. The degree of delegation per agent action is set in the mandate manifest, and no agent a person directs is given M3 Consequence over a person.
Observable, expected, speculation: Agent teams in software development and service operations are observable; people directing agents as part of ordinary line management is observable in pilots and expected to spread; the claim that every employee becomes an agent boss is speculation.
Figure: A team leader with people and agents: the mandate manifest, the containment boundary and the verification loop around the agents, and the functions that remain with the leader for the people

Chapter 17: Orchestrators: AI Managing AI Agents and the Chain of Accountability

File: 17_orchestrators.md
Status: draft
Purpose: Examine orchestration: AI agents that assign tasks to other agents, check their output and decide what happens next, and the chain of accountability that results. Describe how orchestrator and worker agents are built, through multi-agent research systems and protocols for agent-to-agent communication and tool access, and what is documented about how multi-agent systems fail. Apply the accountability literature, the problem of many hands, the moral crumple zone and the distinction between answerability and liability, to a chain in which a person directs an orchestrator that directs agents whose output reaches people. Show where the chain must end, in a named person who owns the orchestrator's mandate and who can stop the whole chain, and why an orchestrator may perform M1 Direction and M2 Evaluation of agents but never M3 Consequence for a person. Answer Q7 with a rule that can be tested.
Builds on: F13, F19, F51, F63, F67, F68, F45, Q7
Management functions: M1 Direction; M2 Evaluation; M3 Consequence
Sections: Agents managing agents; Protocols and plumbing; How multi-agent systems fail; Many hands; The moral crumple zone; Where the chain ends; Making the chain visible; A rule for Q7
Practical guides: Drawing the chain of accountability for a multi-agent workflow and testing the stop path
Worked examples: The accountability chain for Hansa's claims workflow, from the claims-triage agent through the document-intake and recovery agents to the handler, with the owner of each mandate and the single stop path
Nordvik Region and Hansa Mutual: At Hansa, the claims-triage agent routes a claim and allocates it to a handler; in Jana Petersen's pilot team the document-intake agent prepares the file and the recovery agent drafts a subrogation letter. That is already a chain, and the chapter draws it: who owns each mandate, which agent's output another agent consumes, and who can stop all of it. Thomas Brandt's question is whether the throughput figures produced along the chain could be combined into an evaluation of a handler that nobody decided on. At Nordvik the chain runs through a supplier: the ShiftWise supplier announces an optimisation component that negotiates gaps with staffing agencies' systems, and the Region uses the chapter's test to decide what it requires by contract before any such component is enabled, under tender NR-2025-089.
Usage modes and degrees of delegation: Orchestration is a Mode 5 pattern even when it is reached through Mode 3 tools; each agent in the chain has its own degree of delegation, but the chain's effective degree is set by the weakest checkpoint, which the chain record makes visible.
Observable, expected, speculation: Multi-agent systems in development and research are observable; orchestrators directing work that reaches people are observable in pilots and expected to spread; long chains of agents across organisations are speculation in management, though expected in procurement and logistics.
Figure: A chain of accountability from a person through an orchestrator and worker agents to the people affected, with the owner of each mandate, the single stop path and the boundary below which no M3 Consequence may be taken

Chapter 18: The Mixed Team: People and Agents Under the Same Manager

File: 18_the_mixed_team.md
Status: draft
Purpose: Examine the mixed team, in which people and agents work under the same manager, and the questions that arise when they share work, figures and attention. Describe how work is divided, how figures for agents and people are kept apart so that agents' throughput does not become the benchmark for people, how handovers between agents and people are designed, and how the team's sense of competence and recognition survives when software does the visible part of the work. Draw on team design, psychological safety and the evidence on what happens to people who work beside more capable tools. Set out what co-determination means for a mixed team, where a works council has rights over systems that can monitor people but not over agents as such, and what the manager owes the people in M4 Development and M5 Care that no agent needs.
Builds on: F10, F22, F24, F53, F75, F66, Q6, Q11
Management functions: M1 Direction; M2 Evaluation; M4 Development; M5 Care
Sections: One manager, two kinds of worker; Dividing the work; Two sets of figures; Handovers; Recognition and competence; Development and care; Co-determination in a mixed team; Team design patterns
Practical guides: A team charter for a mixed team: division of work, separated figures, handovers and reserved learning tasks
Worked examples: The team charter for Jana Petersen's mixed claims team, with the rule that agents' throughput is reported separately and never compared with handlers' figures
Nordvik Region and Hansa Mutual: Jana Petersen's claims team is the book's mixed team: about 20 handlers, the claims-triage agent that allocates their work, and the pilot's document-intake and recovery agents. The team charter keeps complex claims and a share of routine claims for handlers to learn on, reports the agents' throughput separately from the handlers' and never compares them, and records each handover from agent to handler. Thomas Brandt asks for the charter to be attached to the pilot's terms (AIDR-0652) so that its separation of figures is binding. At Nordvik, the nearest equivalent is the home-care unit in which ShiftWise routes the work and the staff assistant drafts action lists; Mikael Strand uses the charter's questions to decide which planning tasks his senior assistant nurses keep, so that they can still plan routes by hand when the system is down.
Usage modes and degrees of delegation: Mixed teams combine Mode 2 assistants used by people, Mode 4 systems that direct people and Mode 5 agents directed by the manager; the team charter states which mode and degree applies to each kind of work.
Observable, expected, speculation: Mixed teams of people and agents are observable in pilots; their spread in claims, service and back-office work is expected; effects on team cohesion and learning over years are speculation until measured.
Figure: A mixed team under one manager, with people and agents, the separated figures, the handovers and the reserved learning tasks

Chapter 19: Skill, Judgement and Who Becomes a Manager When AI Does the Managing

File: 19_who_becomes_a_manager.md
Status: draft
Purpose: Ask who becomes a manager when AI does much of the managing, and what happens to managerial skill and judgement. Describe how people have traditionally learned to manage, by leading small teams, scheduling, giving feedback and handling conflict, and how first-line management work such as rostering, allocation and reporting is now being automated. Use the literature on becoming a manager, on expertise and intuitive judgement, on automation and skill decay and on the evidence that promotion often rewards performance in the job rather than managerial ability. Show the risk that managers lose the judgement they need to oversee the AI boss and that the route into management narrows, and the opportunity to free managers for the relational work that systems cannot do. Propose how organisations keep managerial craft, reserve learning experiences and evidence competence in records of judgement exercised, answering Q5.
Builds on: F11, F26, F53, F66, F77, F79, Q5
Management functions: M2 Evaluation; M4 Development
Sections: Learning to manage; The automated junior management work; Judgement and its decay; Who gets promoted; The narrowing route; Freeing managers for relational work; Keeping the craft; Evidencing competence
Practical guides: A managerial craft programme: unassisted rostering and allocation, reserved conflict and feedback experiences, and competence records
Worked examples: A competence record for a first-line manager that records judgement exercised against the AI boss, for example rosters changed and allocations overridden, with reasons
Nordvik Region and Hansa Mutual: Nordvik's home-care units have traditionally grown their unit managers from senior assistant nurses who planned routes and covered absences; with ShiftWise drafting routes, that apprenticeship narrows. Ingrid Halvorsen and Mikael Strand agree that deputies in each unit plan one week a month by hand, and that the unit manager's development conversation records the rosters he or she changed and why. Hansa's team leaders have come from experienced handlers, and the claims-triage agent now does the allocation they once learned on. Jana Petersen, with twelve years in claims, can judge the agent's allocations; her deputy has two. Hansa adds managerial judgement to its competence framework, with records of allocations overridden and reasons, so that the next team leader is chosen on evidence of judgement and not only throughput.
Usage modes and degrees of delegation: Managerial skill decays fastest where M1 Direction and M2 Evaluation are delegated at Level 3 Bounded or above, since the manager no longer drafts; at Level 2 Supervised the manager still edits, which keeps judgement only if the edits are real.
Observable, expected, speculation: The automation of first-line management tasks is observable; erosion of managerial judgement is expected, by analogy with the measured skill decay in other assisted work; the disappearance of first-line management as a career step is speculation.
Figure: The route into management before and after the AI boss, with the learning experiences that are automated, those that are reserved and the competence record that evidences judgement

Chapter 20: Decision Rights in the Management Chain: Who May Direct, Evaluate, Reward and Dismiss

File: 20_decision_rights_in_the_management_chain.md
Status: draft
Purpose: Set out decision rights in the management chain: who may direct, evaluate, reward and dismiss, and what may be delegated to AI. Turn the five management functions and the degrees of delegation into a written allocation of decision rights per role, per function and per system, in the manner of the decision tiers and stop rights of the earlier volumes. Show that consequence decisions about a person always have a named human decider with the authority and information to decide otherwise, that stop and pause rights over a management system are assigned before they are needed, and that co-determination and the law fix some rights that management cannot reassign. Address the accountability of management bodies, which cannot be delegated to a system. Answer Q1 and Q2 as rules: which functions may be delegated, to what degree, by whom, and who answers for a consequence decision an AI recommended and a manager accepted.
Builds on: F17, F18, F19, F51, F63, F74, F78, Q1, Q2
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M4 Development; M5 Care
Sections: Decision rights as an artefact; Who may delegate; Direction and evaluation; Consequence; Development and care; Stop and pause; Rights fixed by law and agreement; Q1 and Q2 as rules
Practical guides: Writing a decision-rights file for management functions and reviewing it with the representatives
Worked examples: management/decision-rights.yaml for both organisations, with the decider per function, the maximum degree of delegation, the stop right and the representatives' rights
Nordvik Region and Hansa Mutual: Nordvik's decision-rights file gives unit managers the right to approve rosters and to pause ShiftWise offers in their unit, the HR director the right to set or change any delegation profile after negotiation with the unions under the co-determination act, and safety representatives the right to require an investigation of a work-environment risk. No consequence right is delegated. Hansa's file gives team leaders the right to reassign any claim the agent allocated, recruiters the hiring decision, sales managers the pay decision, and the chief HR officer the right to stop any management system group-wide; it records that changes to systems suited to monitor require the agreement of the Konzernbetriebsrat or the competent local body under BetrVG § 87(1) no. 6. The chief HR officer asks how the executive board's own accountability is shown, and the file names it.
Usage modes and degrees of delegation: Decision rights are written per function, but the mode decides where they are enforced: Mode 1 to Mode 3 tools inherit the manager's rights, Mode 4 applications enforce them in configuration and Mode 5 agents through their mandate and permissions.
Observable, expected, speculation: Decision tiers, stop rights and co-determination rights are observable practice; their written allocation per management function is the book's proposal; its uptake by organisations is expected as the AI Act's deployer duties apply.
Figure: The management chain as a decision-rights matrix: roles against the five functions, with the maximum degree of delegation, the human decider for consequences and the stop rights

Part V: Europe's Answer

Chapter 21: The AI Act at Work: Annex III Point 4, Prohibited Practices and the Deployer's Duties

File: 21_the_ai_act_at_work.md
Status: draft
Purpose: Explain what the AI Act requires of AI used to manage people, as organisational and architectural implications rather than legal advice. Cover Annex III point 4, which lists AI for recruitment and selection and for decisions affecting work-related relationships, promotion and termination, task allocation based on individual behaviour or traits, and monitoring and evaluating performance and behaviour; the derogation in Article 6(3) and its limits; the prohibition of emotion recognition at the workplace in Article 5(1)(f) and the Commission's guidelines; the deployer's duties in Article 26, including human oversight by competent people and the duty in Article 26(7) to inform workers' representatives and affected workers before a high-risk system is used at the workplace; the fundamental rights impact assessment of Article 27 for public bodies; the right to an explanation in Article 86; and AI literacy under Article 4. Address the application dates as amended by the Digital Omnibus, which the writer verifies against the live text.
Builds on: F15, F16, F25, F31, F78, F38, Q3, Q10
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M5 Care
Sections: Employment in the AI Act; Annex III point 4; The Article 6(3) derogation; Prohibited at work; The deployer's duties; Impact assessment and explanation; AI literacy for managers; Dates and the Digital Omnibus; From articles to controls
Practical guides: Classifying an AI system that touches management functions under the AI Act, and recording the result
Worked examples: Hansa's reassessment AIDR-0651 of the claims-triage agent's work-allocation function under Annex III point 4(b), and Nordvik's FRIA-2026-007 for ShiftWise
Nordvik Region and Hansa Mutual: Nordvik deploys ShiftWise as a high-risk system under Annex III point 4(b); as a public body it completes the fundamental rights impact assessment FRIA-2026-007, informs the unions and staff under Article 26(7) before go-live in home care, and assigns oversight to unit managers who receive training under Article 4. Whether its home-care dashboards fall under point 4(b) is left open in the profile, and the chapter shows how the Region decides it. Hansa reopens the classification of the claims-triage agent, made as not high-risk for its insurance function (AIDR-0024), for its work-allocation function (AIDR-0651), and keeps TalentLens as high-risk under point 4(a) with affect and voice analysis switched off under Article 5(1)(f). Its UK and Swiss entities sit outside the Act, and the chapter shows how the group applies one policy across seven markets.
Usage modes and degrees of delegation: The Act classifies by intended purpose, not by mode; a Mode 0 feature in an HR suite can be high-risk if it evaluates performance, and a Mode 5 agent that allocates tasks on individual behaviour falls within point 4(b) whatever its architecture.
Observable, expected, speculation: The Act's text and the prohibition are observable law; the dates on which the high-risk obligations apply, as amended by the Digital Omnibus, and the content of harmonised standards are expected and must be verified; enforcement practice in employment is speculation until supervisors act.
Figure: The AI Act's provisions on work mapped onto the five management functions, with the high-risk listing, the prohibition, the deployer's duties and the worker's rights

Chapter 22: Data Protection and the Platform Work Directive: The Right to a Human Decision

File: 22_data_protection_and_platform_work_directive.md
Status: draft
Purpose: Explain the right to a human decision in European data protection law and in the Platform Work Directive, as organisational and architectural implications. Cover GDPR Article 22 on decisions based solely on automated processing, the SCHUFA judgment on scores that play a determining role, the Dun & Bradstreet judgment on meaningful information about the logic involved, Article 88 and the CJEU's 2023 judgment on national employee-data rules, and the guidance of the Article 29 Working Party on automated decisions and on data processing at work. Then set out Chapter III of Directive (EU) 2024/2831 on algorithmic management in platform work: limits on processing, transparency, human monitoring, human review and the rule that decisions to restrict, suspend or terminate are taken by a human being, with transposition by 2 December 2026. Show what the Amsterdam Uber and Ola judgments and the Italian decisions add, and how these rights become a design requirement for every employer.
Builds on: F31, F38, F21, F19, Q3, Q2
Management functions: M2 Evaluation; M3 Consequence
Sections: Solely automated decisions; Scores that decide; Explaining the logic; Employment data; The guidance; The Platform Work Directive, Chapter III; What the courts added; From rights to design
Practical guides: Checking a management system for solely automated decisions and building the human decision point and the explanation into it
Worked examples: An Article 22 assessment of ShiftWise open-shift offers at Level 3 Bounded, and an explanation template for Hansa's sales-performance score
Nordvik Region and Hansa Mutual: Nordvik's open-shift offers are the only management decision in its estate taken without a person in the individual case; the data protection officer assesses whether an offer is a decision with legal or similarly significant effect under Article 22 and concludes that it is not while offers go to all staff who opted in, but would be if the system began to rank them, which the Region writes into DLG-NR-ROS-01 as a condition. Hansa's sales-performance score is an input to variable pay; following the SCHUFA reasoning, the group tests whether managers depart from it often enough to show it is not determining, and builds an explanation of each score's indicators, which Article 88 rules in Germany and the Netherlands and the group works agreement KBV-IT-2024 require it to share with the employee.
Usage modes and degrees of delegation: Article 22 turns on whether a person is meaningfully involved, not on the mode; but Mode 4 and Mode 5 systems at Level 3 Bounded or Level 4 Autonomous are where solely automated decisions arise, and Mode 0 features can make them without anyone noticing.
Observable, expected, speculation: Article 22, the judgments and the Directive's text are observable; national transposition of the Directive and its application to employers that are not platforms through national law are expected; convergence of employment data-protection rules across member states is speculation.
Figure: The right to a human decision: a management decision path with the Article 22 test, the explanation service and the Platform Work Directive's human review, marked where each applies

Chapter 23: Co-determination: Works Councils, Unions and the Nordic Model Meet the Algorithm

File: 23_co_determination.md
Status: draft
Purpose: Explain how co-determination, works councils, unions and the Nordic model meet the algorithm, and answer Q8: how do co-determination and collective agreements apply to a system that changes daily? Cover German co-determination under BetrVG § 87(1) no. 6 on technical equipment suited to monitor and §§ 94 and 95 on questionnaires and selection guidelines, the 2021 reform that gave works councils an expert on AI and rights over AI in selection guidelines, and the case law that treats almost any software that records behaviour as suited to monitor; the Swedish co-determination act and collective agreements; the Dutch works councils' consent right; the information rights for works councils in Spain's Rider law and in Italy's transparency decree; the European Works Council and information and consultation directives; and the European social partners' framework agreement on digitalisation. Show how rights designed for systems installed once can apply to systems retrained, reconfigured and redeployed continuously, through agreements that govern change rather than versions.
Builds on: F52, F18, F16, F38, Q8
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M5 Care
Sections: Co-determination as a European constant; Germany; Sweden and the Nordic model; The Netherlands, Spain and Italy; The European level; Negotiating the algorithm; A system that changes daily; The delegation profile as a negotiated document
Practical guides: Drafting a works agreement or local agreement for an AI system that changes continuously: change classes, triggers, data access and review
Worked examples: The change-class annex to Hansa's KBV-IT-2024 for the claims-triage agent, and Nordvik's negotiation record with the unions for DLG-NR-ROS-01
Nordvik Region and Hansa Mutual: Nordvik negotiates DLG-NR-ROS-01 with the public-sector unions under the co-determination act before ShiftWise goes live in home care (AIDR-0602), and agrees that changes to the optimisation objectives or to the rules for open-shift offers are important changes that require renewed negotiation, while bug fixes and forecast refreshes are notified. Hansa's group works agreement KBV-IT-2024, negotiated by Thomas Brandt, governs the claims-triage agent's figures and TalentLens; the chapter follows its revision to add change classes for model updates and allocation-rule changes, and the rollout gate per entity in the Netherlands, Sweden, Norway, Poland, the United Kingdom and Switzerland, each under its own arrangement. The European Works Council is informed of the agent-boss pilot because it spans more than one member state in its next phase.
Usage modes and degrees of delegation: Co-determination attaches to systems suited to monitor, whatever their mode; Mode 0 features in licensed suites are where representatives most often learn of monitoring late, and Mode 5 agents that change their behaviour through updates are where change classes matter most.
Observable, expected, speculation: The national rights and the 2021 German reform are observable law; negotiated agreements on algorithmic management are observable in some sectors and expected to spread; a European framework for co-determination over AI is speculation.
Figure: Co-determination over an AI boss across the life of the system: introduction, change classes, triggers and review, with the representatives' rights at each step in Germany and Sweden

Chapter 24: Occupational Safety, Equality Law and the Public Authority as Employer

File: 24_safety_equality_and_public_employers.md
Status: draft
Purpose: Bring together three bodies of law that constrain the AI boss beyond the AI Act and data protection, and the special position of the public authority as employer. Cover occupational safety and health under the Framework Directive 89/391/EEC and national work-environment law, with psychosocial risk, ISO 45003 and the Swedish rules on organisational and social work environment as examples; equality law under the race, framework and recast gender directives and national acts, with indirect discrimination as the main risk of proxies; and pay transparency under Directive (EU) 2023/970. Then examine the public authority as employer: its duties under administrative law and the Charter, its accountability to elected bodies, procurement as the moment when the AI boss is chosen, and the lessons of SyRI and the childcare-benefits affair about automated decisions in the public sector. Answer Q10 and Q11.
Builds on: F31, F57, F47, F19, Q10, Q11
Management functions: M3 Consequence; M5 Care
Sections: Safety and health; Standards for psychosocial risk; Equality law; Pay transparency; The public authority as employer; Lessons from the public sector; Procurement as the decisive moment; Q10 and Q11
Practical guides: Tender requirements for an AI system that performs management functions in a public authority
Worked examples: The management-function annex Nordvik adds to its procurement requirements after NR-2025-089, and an equality impact check of Hansa's sales-performance score
Nordvik Region and Hansa Mutual: Nordvik is a public employer accountable to elected politicians, and its ShiftWise contract under tender NR-2025-089 predates this book's questions; the Region drafts a management-function annex for future tenders requiring explanation, data access for the unions, configurable degrees of delegation and a stop path outside the supplier's system. Its central work-environment committee treats ShiftWise routes as a work-environment change under the Swedish-model act. Hansa, a private employer, applies equality law to the sales-performance score, finding that the complaints indicator may disadvantage part-time staff in some markets, and prepares pay-transparency reporting for its EU entities under Directive (EU) 2023/970. Both organisations face Q11 in different forms: Nordvik in care work where M5 Care is the job, Hansa where it has delegated nothing in M5 Care to AI.
Usage modes and degrees of delegation: Safety, equality and pay-transparency law apply to outcomes, whatever the mode; procurement is where a public authority sets the mode and the degree of delegation of a Mode 4 system it will deploy, and where it must require them to be configurable.
Observable, expected, speculation: The directives and national acts are observable law; transposition of the Pay Transparency Directive by June 2026 is observable or pending by member state and must be verified; inspections of algorithmic management by work-environment authorities are expected.
Figure: The legal frame around the AI boss beyond the AI Act: safety, equality and pay law around the five functions, with the public employer's additional duties and procurement as the point of control

Part VI: Designing the Accountable AI Boss

Chapter 25: Management Decisions as Records: Decision Records and the Delegation Register

File: 25_management_decisions_as_records.md
Status: draft
Purpose: Introduce the book's central artefacts: the Management Decision Record and the delegation register of management functions. A Management Decision Record (MDR) records a management decision that AI made or shaped about a person: the function, the system and its version, the degree of delegation, what the system proposed, what the human decider decided and why, the information given to the person, and the route to contest; it is the sibling of the ADR and the AIDR. The delegation register, management/delegation-register.yaml, holds a delegation profile (DLG-*) for each system or team, stating the degree per function, the named decider, the representatives' agreement and the review triggers, and so extends the delegation register of AI Threat, Chapter 26. Show how both are generated as by-products of normal operation, validated against a schema, protected for privacy and access, and readable by the worker, the works council, a court and a supervisor, answering Q3.
Builds on: F8, F14, F38, F2, F4, F20, F69, Q1, Q3
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M4 Development; M5 Care
Sections: Why records; The Management Decision Record; Records by operation; The delegation profile; The delegation register; Schema, validation and change; Privacy and access; One record, four readers; How the records fit together
Practical guides: Introducing Management Decision Records for one system: scope, schema, generation and access
Worked examples: MDR-NR-0001, Mikael Strand's first approved roster for the Strandvik unit, and MDR-HM-0001, Jana Petersen's reassignment of claims overriding the claims-triage agent, with the delegation register management/delegation-register.yaml for both organisations
Nordvik Region and Hansa Mutual: At Nordvik, MDR-NR-0001 records the first roster for the Strandvik unit approved under DLG-NR-ROS-01: what ShiftWise proposed, the three shifts Mikael Strand changed and why, and the notice given to staff. The Region's register holds DLG-NR-ROS-01 and DLG-NR-HCD-01 and records the paused screening pilot, and the HR director reports it with AIDR-0600. At Hansa, MDR-HM-0001 records Jana Petersen's reassignment of claims that the claims-triage agent had allocated, with her reason; the register holds DLG-HM-CTA-01, DLG-HM-REC-01, DLG-HM-SPS-01 and DLG-HM-AB-01. Thomas Brandt agrees that MDRs are generated for every override and every consequence decision, but not for every allocation, and that the Konzernbetriebsrat sees them aggregated by team.
Usage modes and degrees of delegation: MDRs are generated differently by mode: Mode 4 and Mode 5 systems emit events that become records; decisions shaped by Mode 1 to Mode 3 assistants need the manager to record the assistant's part, which is why the register records which modes touch each function.
Observable, expected, speculation: Decision records for architecture and AI are observable practice in the series; Management Decision Records are the book's proposal; their adoption is expected where explanation rights and co-determination require evidence of human decisions.
Figure: The Management Decision Record and the delegation register: a decision flowing from system proposal through the human decider to the record, with the register above it and the four readers around it

Chapter 26: The Human in the Chair: Oversight That Is Real, Not Rubber-Stamped

File: 26_the_human_in_the_chair.md
Status: draft
Purpose: Design human oversight of the AI boss that is real rather than rubber-stamped, and answer Q9: can oversight be real when the overseer has hundreds of decisions an hour? Draw on the evidence on automation bias, selective adherence to algorithmic advice in the public sector and the flaws of policies that merely require a human in the loop, and on the oversight-quality metric and review designs of AI Threat, Chapter 14, and The Transformed Security Architect, Chapter 14. Distinguish oversight of the system, which can be sampled and measured, from the human decision in an individual consequence case, which cannot be delegated. Introduce the oversight record and an oversight-quality metric for management systems: time to decision, rate of departure from the proposal, reversals on contest and seeded cases, reported at team level where representatives require it. Show how the number of decisions per overseer, their competence and their authority to decide otherwise are designed, and what happens when oversight fails.
Builds on: F21, F52, F53, F76, F66, F19, Q2, Q9
Management functions: M2 Evaluation; M3 Consequence
Sections: The human in the loop, examined; Automation bias and selective adherence; Oversight of the system and decisions in the case; The oversight record; Measuring oversight quality; Designing volume; Competence and authority; When oversight fails
Practical guides: Setting up an oversight-quality metric for a management system with the representatives' agreement
Worked examples: An oversight record and metric for team leaders overseeing claims allocation at Hansa, extending OQM-HM-CLM-01 and reported at team level only, and Nordvik's roster-approval oversight record for unit managers
Nordvik Region and Hansa Mutual: Hansa's oversight-quality metric OQM-HM-CLM-01 already reports at team level only for the claims-triage agent; the chapter extends it to team leaders' oversight of allocation, with reassignments, the reasons given and seeded allocations that a careful team leader should catch, and Thomas Brandt agrees on condition that no individual team leader's figure is shown. The same design applies to recruiters reviewing TalentLens rankings, where the time spent before rejecting a candidate is the signal. At Nordvik, unit managers approve every ShiftWise roster; Ingrid Halvorsen learns from the oversight records that some approve a four-week roster in under two minutes, and the Region decides that a roster cannot be approved until its flagged rest-rule breaches have been resolved, and that unit managers receive the AI literacy training under Article 4 before their units go live.
Usage modes and degrees of delegation: Oversight design differs by degree: at Level 2 Supervised every case passes a person, so volume and time are the design variables; at Level 3 Bounded oversight moves to the rules and the sample; for M3 Consequence the human decision in each case is required and cannot be replaced by sampling.
Observable, expected, speculation: Automation bias and rubber-stamping are observable in the research and in the organisations' own records; oversight metrics for management systems are the book's proposal; their acceptance by works councils is expected where reported at team level.
Figure: The oversight loop for a management system: proposals, the human decider, seeded cases, the oversight record and the metric, with the escalation path back to a lower degree of delegation

Chapter 27: Contestability by Design: Explanation, Appeal and Correction

File: 27_contestability_by_design.md
Status: draft
Purpose: Design contestability into the AI boss: explanation, appeal and correction that a worker can use without a lawyer. Draw on the right to an explanation in AI Act Article 86, GDPR Articles 15 and 22 and the Dun & Bradstreet judgment, the Platform Work Directive's right to review, the research on counterfactual explanations and on how people perceive procedural justice in algorithmic decisions, and the courts' handling of platform deactivations. Introduce the contestation record (CNT-*) and the appeal path: who receives a contest, the time to answer, the independent reviewer, the correction of the decision and of the system, and the report of contest patterns to the representatives and the delegation register. Show that contestation is also the best signal that a system or its oversight is failing, and that an appeal that reaches a person able to change the outcome is the condition on which every degree of delegation above Level 1 depends.
Builds on: F19, F31, F38, F60, F21, Q3
Management functions: M2 Evaluation; M3 Consequence
Sections: Why contestability; What a worker can learn; Explanations people can use; The appeal path; The contestation record; Correcting the decision and the system; Contest patterns; Contestability as a condition
Practical guides: Building an appeal path for a management system, with time limits, reviewer and correction route
Worked examples: CNT-NR-0001, an assistant nurse's contestation of a home-care route that left no time between two visits, and CNT-HM-0001, a sales employee's contestation of a quarterly score, each from receipt to correction
Nordvik Region and Hansa Mutual: At Nordvik, CNT-NR-0001 is an assistant nurse's contestation of a route that left no time between two visits; Mikael Strand corrects the route within a day, and the correction becomes a minimum travel-time constraint in ShiftWise and a report to the central work-environment committee. The Region's appeal path routes contests to the unit manager first and to HR if the manager made the decision. At Hansa, CNT-HM-0001 is a sales employee's contestation of a quarterly score; the reviewer, a sales manager from another region, finds that complaints were attributed to the wrong person after a portfolio transfer, corrects the score and the pay input, and the data rule is fixed. Thomas Brandt asks for a quarterly report of contests by system and outcome, which becomes a trigger in DLG-HM-SPS-01.
Usage modes and degrees of delegation: Contestability is designed into the system whatever its mode, but Mode 0 features and Mode 5 agents need the most deliberate design, because the person affected may not know the system exists or which agent produced the result.
Observable, expected, speculation: Rights to explanation and review are observable in law; effective appeal paths in algorithmic management are observable in few organisations; contest data as a routine governance signal is expected.
Figure: The contestation path from decision to explanation, appeal, independent review, correction of the decision and correction of the system, with the contestation record and the report to representatives

Chapter 28: Transparency to the Managed: What Workers Are Told, When and by Whom

File: 28_transparency_to_the_managed.md
Status: draft
Purpose: Design transparency to the people managed: what workers are told about the AI boss, when and by whom. Cover the information duties of AI Act Article 26(7) towards workers' representatives and affected workers and Article 50 where people interact with AI, GDPR Articles 13 to 15, the Platform Work Directive's transparency provisions on automated monitoring and decision-making systems, Italy's transparency decree and Spain's Rider law, and the co-determination information rights of Chapter 23. Introduce the worker information notice, written for the person, layered from a one-page summary to the full logic, and the transparency register that lists every system that performs a management function, its degree of delegation and the person to contact. Address timing, before introduction, on change and on request, the manager's role in explaining rather than forwarding a notice, and the limits of transparency, including the transparency paradox and trade secrets.
Builds on: F31, F15, F57, F38, Q3, Q8
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M5 Care
Sections: The duty to inform; National examples; The worker information notice; The transparency register; Timing; The manager explains; Limits of transparency; Transparency for representatives
Practical guides: Writing a worker information notice for a management system and testing it with the people it is for
Worked examples: The worker information notice for ShiftWise in Nordvik's home care and the transparency register for Hansa's EU entities, listing the claims-triage agent's allocation, TalentLens and the sales-performance score
Nordvik Region and Hansa Mutual: Nordvik writes a one-page worker information notice for ShiftWise in home care, tested with assistant nurses in the Strandvik unit, stating what the system does, the degree of delegation per function, what data it uses, whether acceptance of open shifts is recorded and how to contest a route; Mikael Strand explains it at a workplace meeting before go-live, as required under Article 26(7) and agreed with the unions. Hansa publishes a transparency register on its intranet for the EU entities, listing every system that performs a management function, and adapts it for the UK and Swiss entities, which sit outside the AI Act. Thomas Brandt asks that the register show the Konzernbetriebsrat's agreement and its date for each system, which the group accepts.
Usage modes and degrees of delegation: Transparency is hardest for Mode 0 features nobody registered and for Mode 5 agents whose behaviour changes with updates; the register records mode and degree, and a change of either triggers a new notice.
Observable, expected, speculation: The information duties are observable law; worker information notices for algorithmic management are observable in platform work and expected elsewhere; transparency registers for employers are the book's proposal, by analogy with public algorithm registers.
Figure: Layers of transparency to the managed: the notice, the register, the manager's explanation and the representatives' view, with the triggers for renewed information

Chapter 29: The Transformed Manager: Leading People When AI Directs the Work

File: 29_the_transformed_manager.md
Status: draft
Purpose: Describe the transformed manager: leading people when AI directs the work, and conducting agents as part of the job. Bring together the evidence of the book on what managers do, what the AI boss can and cannot do and what people need from a manager, and the parallel with the security architect's Shift 4 in The Transformed Security Architect. Argue that the manager's role is transformed, not abolished: from supervisor of tasks to accountable leader of people and conductor of agents, holding the named decisions in M3 Consequence, the relationship in M4 Development and the duty in M5 Care, while M1 Direction and much of M2 Evaluation are shared with systems under written delegation. Set out the competences, the role charter and the time budget of the transformed manager, and what organisations must change in selection, training and span of control. Close with the author's own projections for the manager's role, labelled as such and kept apart from the findings.
Builds on: F71, F72, F73, F75, F77, F79, F80, Q5, Q6, Q12
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M4 Development; M5 Care
Sections: Transformed, not abolished; The parallel with the security architect; The manager in each function; Leading people; Conducting agents; The manager's charter and time budget; Competences; What organisations must change; The author's projections
Practical guides: Writing a role charter and time budget for a line manager whose team includes an AI boss and agents
Worked examples: A role charter for a first-line manager, with decision rights per function, the agents directed under mandate and a time budget that protects M4 Development and M5 Care
Nordvik Region and Hansa Mutual: Mikael Strand's role charter keeps roster approval, contested routes, development conversations and the work environment with him, and records that ShiftWise saves him about a day a fortnight that he spends on home visits with staff and on development conversations; the Region decides not to widen unit managers' span of control on the strength of the saving. Jana Petersen's charter adds the direction of the pilot's agents under MND-HM-CLM-01, the reassignment of claims and the verification routine, and protects time for coaching handlers whose routine work the agents now do. Ingrid Halvorsen and Hansa's chief HR officer reach the same conclusion by different routes: the transformed manager is chosen and trained for judgement and care, not for throughput.
Usage modes and degrees of delegation: The transformed manager uses Mode 1 to Mode 3 assistance with their own identity, oversees Mode 4 systems that direct their team and directs Mode 5 agents under mandate; the charter states the degree for each and keeps M3 Consequence and M5 Care with the manager.
Observable, expected, speculation: The automation of parts of M1 Direction and M2 Evaluation is observable; the shift of managers' time to relational work is expected only where organisations protect it; the author's projections for the role in 2036 are labelled as such and kept to their own section, consistent with Chapter 33.
Figure: The transformed manager: the five functions with what the manager holds, shares and conducts, the agents under mandate and the protected time for development and care

Part VII: Experience and Verdict

Chapter 30: A Year With the AI Boss: Nordvik Region and Hansa Mutual

File: 30_a_year_with_the_ai_boss.md
Status: draft
Purpose: Follow four people through a year with the AI boss, from November 2026 to November 2027, as an illustrative narrative that applies the book's instruments in two organisations. At Nordvik Region, the HR director answers her commission and the manager of a home-care unit takes ShiftWise routes into daily practice; at Hansa Mutual, a claims team leader runs the agent-boss pilot with people and agents and the chair of the Konzernbetriebsrat negotiates the agreements that govern it. Show the decisions they take quarter by quarter, the records they produce, what goes wrong, what is contested and corrected, how co-determination shapes the outcome and how their own roles change. Keep every fact consistent with the people-management profile and mark the narrative as illustrative, so that the reader can follow the instruments of Parts IV to VI in use rather than in principle.
Builds on: F69, F52, F75, F53, F21, Q8, Q9, Q6
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M4 Development; M5 Care
Sections: Four people, two organisations; Quarter one: commissions and agreements; Quarter two: go-live and the first contests; Quarter three: oversight under pressure; Quarter four: review and renegotiation; What went wrong; How the roles changed; What the year shows
Practical guides: A quarterly review of the delegation register, the MDRs, the contestation records and the oversight records with the representatives
Worked examples: A year's timeline as code for each organisation, linking each quarter's decision records in AIDR-0600 to AIDR-0649 and AIDR-0650 to AIDR-0699, the MDRs, contestation records and delegation profiles they produced
Nordvik Region and Hansa Mutual: The chapter follows Ingrid Halvorsen, Nordvik's HR director since 2023, who owns AIDR-0600 and reports to the regional executive committee by 30 June 2027; Mikael Strand, manager of the Strandvik home-care unit with about 35 staff, a third part-time or hourly, the first unit to use ShiftWise routes and the time-per-visit dashboard; Jana Petersen, a Hamburg claims team leader with twelve years in motor and property claims, about 20 handlers and, from January 2027, the pilot's document-intake and recovery agents under MND-HM-CLM-01; and Thomas Brandt, chair of the Konzernbetriebsrat since 2022, a former underwriter who negotiated KBV-IT-2024 and sits on the European Works Council. Their year begins with AIDR-0600 and AIDR-0650 and ends with the revised delegation register in each organisation.
Usage modes and degrees of delegation: The narrative shows the modes in use: Nordvik with a Mode 4 rostering system, Mode 2 and Mode 4 assistants and no agents directed by managers; Hansa with a Mode 5 claims-triage agent allocating work to people at Level 3 Bounded, Mode 5 pilot agents directed by a person at Level 2 Supervised, and Mode 2 office assistants.
Observable, expected, speculation: The narrative is illustrative and set in 2026 to 2027; the instruments it applies are drawn from the book and the events it draws on are observable elsewhere; its outcomes for the organisations are expected consequences of the decisions recorded, not predictions.
Figure: Two parallel timelines for the year, quarter by quarter, following the four people, with the records each produced and the management functions each concerns

Chapter 31: Case Studies and Anti-Patterns

File: 31_case_studies_and_anti_patterns.md
Status: draft
Purpose: Collect case studies and anti-patterns of the AI boss. Present the real cases of Part II by name, each with what happened, what was decided and what the case teaches, described by what courts, authorities, inquiries and reporters found: the Foodinho and Deliveroo decisions, the Bologna ruling, the Amsterdam Uber and Ola judgments, the Amazon warehouse and Flex reporting, the CNIL and Hamburg monitoring fines, Amazon's recruiting tool, the iTutorGroup settlement, the Workday litigation, the Microsoft productivity-score reversal, and, as the clearest warning about trusting a system over people, the Post Office Horizon litigation and inquiry. Add composite cases, labelled illustrative, that show patterns not yet documented in a single public case, such as a mixed team whose agents' throughput became the benchmark. Distil the anti-patterns into a named, testable list, each with its signal, its consequence and the control in the book that prevents it.
Builds on: F60, F65, F67, F28, F21, Q2
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M4 Development; M5 Care
Sections: How to read the cases; Platforms; Warehouses and pace; Hiring; Scores and monitoring; Trusting the system over people; Composite cases; The anti-patterns; A checklist
Practical guides: Reviewing an organisation's AI boss against the anti-pattern checklist
Worked examples: The anti-pattern checklist as code with signal, consequence and control for each entry, applied to Nordvik's and Hansa's delegation registers
Nordvik Region and Hansa Mutual: Both organisations review their delegation registers against the anti-pattern checklist at the end of their year. Nordvik finds two near-misses: unit managers approving rosters in under two minutes, the approval-in-seconds pattern, and the risk that open-shift acceptance becomes a hidden rating, the hidden-score pattern; both have controls from Chapters 26 and 28. Its paused screening pilot is recorded as a case of the right response to an unexplainable system. Hansa finds that the sales-performance score had drifted towards a silent pay input in two markets, where managers departed from it in fewer than one case in fifty, and that the pilot charter had prevented the benchmark-agent pattern. Thomas Brandt asks for the checklist to be annexed to KBV-IT-2024. The composite cases in the chapter are illustrative and not drawn from either organisation.
Usage modes and degrees of delegation: The anti-patterns are grouped by where they arise: Mode 0 features nobody registered, Mode 4 systems whose advisory output became decisive, and Mode 5 agents whose allocation or output became the measure of people.
Observable, expected, speculation: The named cases are observable and described by what was found; the composite cases are illustrative; the expectation that the anti-patterns will recur as office agents spread is labelled expected.
Figure: A map of the documented cases and composite cases by management function and degree of delegation, with the anti-patterns they illustrate and the controls that prevent them

Chapter 32: Nordvik Region and Hansa Mutual: Conclusions for Two Organisations

File: 32_conclusions_for_nordvik_and_hansa.md
Status: draft
Purpose: Draw conclusions for the two organisations across the whole book, in the manner of the conclusions chapters of the earlier volumes: the decisions each took, what worked, what was hard and what remains unresolved. Answer the HR director's commission at Nordvik Region and the executive board's request at Hansa Mutual explicitly, give each organisation a prioritised 24-month plan for governing AI in management, and compare the two, so that a reader can see how the same five functions and the same rule for consequences lead to different priorities in a public authority that procures its AI boss for shift work in care and a multinational insurer that builds agents that already allocate office work. Show what each organisation's delegation register looks like at the end and which degrees of delegation it raised, lowered or refused.
Builds on: F65, F69, F70, F18, Q1, Q10, Q8
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M4 Development; M5 Care
Sections: One question, two answers; Nordvik Region: the answer to the regional executive committee; Hansa Mutual: the answer to the executive board; The two delegation registers; Two 24-month plans; Where they differ and why; What remains unresolved; What other organisations can take from the comparison
Practical guides: Turning a year's records into a prioritised 24-month plan for governing AI in management
Worked examples: The 24-month plan for governing AI in management as code for each organisation, with measures, owners, dependencies, the representatives' part and verification
Nordvik Region and Hansa Mutual: Nordvik's conclusion is that ShiftWise stays at Level 2 Supervised for rosters, that open-shift offers stay at Level 3 Bounded only while offers are not ranked, that the home-care dashboards are treated as high-risk under Annex III point 4(b) as a precaution, and that recruitment screening stays paused until a supplier can explain its ranking; the first priority is the procurement annex for management functions. Hansa's conclusion is that the claims-triage agent's allocation is treated as high-risk under point 4(b) (AIDR-0651), that the sales-performance score loses its role as a pay input in markets where managers rarely depart from it, and that the agent-boss pilot extends to further teams only with the team charter and MND-HM-CLM-* mandates; the binding constraint is co-determination across seven markets.
Usage modes and degrees of delegation: Nordvik's plan stays with procured Mode 4 systems and Mode 2 assistants, without agents directed by managers; Hansa's plan extends Mode 5 agents directed by people at Level 2 Supervised and keeps allocation at Level 3 Bounded with reassignment rights, and neither raises any M3 Consequence system above Level 1 Assisted.
Observable, expected, speculation: The plans are the organisations' decisions within the illustration; their outcomes are expected, not observed, and the plans name the signals on which they would be revised.
Figure: The two delegation registers at the end of the book, side by side by function and degree, with the changes during the year and the 24-month plans

Chapter 33: Conclusions: Who Should Be the Boss?

File: 33_conclusion.md
Status: draft
Purpose: Conclude the book by answering its title question: who should be the boss? Restate the thesis and test it against what the chapters found, part by part; state the cross-cutting conclusions and how firm each is; answer Q1 to Q12 or say why they remain open; and say what organisations, managers, representatives, workers and regulators should do first, next and later. Show how this volume extends the series' argument, keep decisions as records that can be corrected, keep the ability to stop outside the system, keep people able to do the work and judge the output, keep dependence reversible and match autonomy to reversibility, to the management of people, where the rule becomes that Level 4 is never used for M3 Consequence. Keep the author's own judgement, including the answer to Q12 about the boss in 2036, in its own labelled section, apart from the findings.
Builds on: F70, F33, F32, F71, F80, Q1, Q2, Q12
Management functions: M1 Direction; M2 Evaluation; M3 Consequence; M4 Development; M5 Care
Sections: The thesis restated; Reflections on each part; Cross-cutting conclusions; How firm the conclusions are; Answered and open questions; What to do first, next and later; The series as one argument; The author's judgement; Who should be the boss?
Practical guides: Ten questions a board, an HR director, a line manager and a works council should be able to answer about their AI boss
Worked examples: A first-90-days plan for an organisation introducing or reviewing an AI boss, as code
Nordvik Region and Hansa Mutual: The chapter brings the two organisations back once, briefly: the same five functions and the same rule for consequences produced a procurement-first answer at Nordvik Region, where the AI boss directs shift work in care and is bought, and a co-determination-first answer at Hansa Mutual, where agents already allocate office work and people have begun to direct agents. In both, the consequence decisions remained with named people, every management decision AI shaped became a record the person could see and contest, and the managers' roles changed rather than disappeared. Ingrid Halvorsen, Mikael Strand, Jana Petersen and Thomas Brandt appear only as witnesses to that conclusion. The fuller comparison is in Chapter 32.
Usage modes and degrees of delegation: The conclusion restates the book's position across the modes and degrees: AI may direct work at Level 2 Supervised or Level 3 Bounded where the action is reversible and contestable; evaluation stays advisory unless its use is overseen; Level 4 is never used for M3 Consequence; and people who direct agents remain accountable for what the agents produce.
Observable, expected, speculation: Each conclusion is labelled by how firm it is; the author's judgement about who will be the boss in 2036 is labelled as a projection and kept to its own section, consistent with Chapter 29.
Figure: The book's argument in one figure: the five management functions, the degrees of delegation with the closed cell for Level 4 in M3 Consequence, the records and the accountable person, set within the argument of the seven volumes

Appendices

34_glossary.md: Defines the vocabulary of algorithmic management, the five management functions (M1 Direction to M5 Care), the degrees of delegation and the usage modes, the book's artefacts (Management Decision Record, delegation register and delegation profile, oversight record, contestation record, worker information notice, transparency register, mandate manifest, decision records), the co-determination and labour-law terms used (works council, Konzernbetriebsrat, European Works Council, collective agreement, safety representative) and the European instruments cited in the book.
35_appendix_a_code_examples.md: Appendix A: complete versions of the management decision, delegation and oversight artefacts in the chapters: the function map management/function-map.yaml, the delegation register management/delegation-register.yaml with delegation profiles DLG-NR-* and DLG-HM-*, the decision-rights file, Management Decision Records MDR-NR-* and MDR-HM-* with their schema, an oversight record and oversight-quality metric, contestation records CNT-*, a worker information notice and transparency register entry, the mandate manifest MND-HM-CLM-01, a team charter for a mixed team, and the anti-pattern checklist.
appendix_templates_and_tools.md: Appendix B: templates and tools for mapping management functions, writing a delegation profile, classifying a system under the AI Act's Annex III point 4, a proportionality review of monitoring, a work-environment risk assessment for algorithmic pace, an Article 22 check, a works-agreement change-class annex, tender requirements for management functions, a worker information notice, an appeal path, a multi-year evaluation plan, a line manager's role charter and an anti-pattern review.
36_references.md: The only bibliography of the book: European and national instruments on AI, data protection, platform work, co-determination, safety and health and equality, court judgments and supervisory decisions, the research literature on algorithmic management, management, work design, automation and oversight, reporting on documented cases, and the six earlier volumes of the series. Chapters cite only entries listed here.
37_about_the_author.md: About the author of The Architecture as Code Series, placed at the very end of the book.