Who Should Be the Boss? AI May Direct Work, but Not Decide About People
For a general professional audience of managers, HR leaders, workers' representatives and executives in European enterprises and public authorities
Algorithmic management is not a forecast: systems already draft rosters, allocate claims, rank applicants and score performance. This lecture presents the argument of AI Boss. Management divides into five functions that can be delegated one by one, at a degree matched to how easily a mistake can be undone. Direction can be delegated within written bounds; a consequence for a person always needs a named human decider with the means to decide otherwise. The lecture shows how that rule is made real through records, measured oversight, an appeal that corrects the system and a manager whose role is transformed, not abolished.
6 min
Two undramatic scenes
A roster drafted by a procured system waits in a unit manager's approval queue; an agent places a claim in a handler's queue.
Neither scene is dramatic, until a route leaves no time to travel between two visits.
The question of who is the boss is then a question about decision rights, records and accountability.
8 min
Five functions, five degrees, one closed cell
M1 Direction, M2 Evaluation, M3 Consequence, M4 Development and M5 Care differ in reversibility, in the person's stake and in governing law.
Each can be delegated from Level 0 Advisory to Level 4 Autonomous, with a default ceiling per function.
Level 4 Autonomous is never used for M3 Consequence, and care is a duty no system discharges.
8 min
How direction and evaluation become consequence
Data collected to direct work becomes a rating, and the rating reaches pay, hours or access to work without anyone deciding it.
Courts and authorities accepted a human decision only where there was evidence of what the human did.
A score recorded as advisory was decisive in practice where 97 per cent of pay proposals in one unit matched it, at a median of 48 seconds.
7 min
What it does to people, and what the evidence can say
People accept algorithmic assignment far more readily than algorithmic evaluation, and what they miss is recognition.
Bias enters through proxies and can be found only by measuring outcomes by group.
Short-run gains are observable; whether algorithmic management works over years is unmeasured, so the quiet harms must be designed against in advance.
8 min
The accountable design
A delegation profile agreed with the representatives before go-live, and a register that shows the sum and tests whether it can be undone.
A decision record in the decider's own words, and an oversight metric that measures the review, not the reviewers.
An appeal that reaches someone other than the decider and corrects the rule, and a notice the person managed can read.
8 min
The transformed manager and the verdict
Agents take direction and the checking of output; development and care are owed only to people and grow.
Time saved by systems becomes development and care only if a role charter protects it.
The thesis names what would refute it; the author's own verdict for 2036 is a person conducting systems, and it is labelled as his judgement.
If you are managed by an AI boss, ask who your decider is.
What the audience takes away
Management is five functions: direction can be delegated within written bounds, while a consequence for a person always keeps a named human decider.
A human in the loop counts only when records show time, authority, independent information and real departures from the system's proposal.
The AI boss is acceptable where it frees managers for development and care, and it must be evaluated on whether it does.
Who Is the Boss? Five Functions, Five Degrees and One Rule
This lecture carries the argument of Part I. It starts in two ordinary scenes, a roster in an approval queue at Nordvik Region and a claim placed in a handler's queue at Hansa Mutual, and shows why the question of who is the boss is a question about decision rights, records and accountability. It then divides management into five functions, shows that algorithmic management inherits Taylor's ambition and adds three new properties, and ends with the book's instruments: degrees of delegation with a default ceiling per function, the named human decider, and the delegation profile.
7 min
Two undramatic scenes
ShiftWise drafts a four-week roster that a unit manager approves; Hansa Mutual's claims-triage agent places a claim in a handler's queue under rules the group head of claims operations owns.
The question becomes concrete when a roster leaves no travel time between two visits, or nine complex claims go to a handler on a reduced schedule.
Algorithmic management is observable: the chapter cites the Foodinho fine of 2.6 million euros, the Amsterdam Court of Appeal on Uber deactivations and the reported warehouse terminations without input from supervisors.
10 min
What a manager does: five functions
Fayol and Drucker described separable functions; Mintzberg and Hill observed work that is informational, interpersonal and fragmented.
The book derives M1 Direction, M2 Evaluation, M3 Consequence, M4 Development and M5 Care from the literature and from the employer's legal duties.
The functions differ in reversibility, in the person's stake and in governing law, and they overlap: evaluation feeds consequence and direction sets the pace that care must answer for.
A function is not a job: line managers, HR, systems, colleagues and the person managed all perform parts of each.
9 min
A century of managing by measurement
From the foreman's simple control through scientific management, the moving line, bureaucratic control and lean production to metrics, platforms and agents.
At every step M1 Direction and M2 Evaluation were handed to a system first, and M3 Consequence followed whenever the measure was tied to pay or a job.
What is new is scale and speed, opacity, and a consequence for a person without a human decision; direction by an agent with its own identity is emerging as a fourth property.
Every regime that measured more eventually needed a contest path.
10 min
Degrees of delegation and default ceilings
The series' autonomy levels, Level 0 Advisory to Level 4 Autonomous, are applied per management function and action class.
Default ceilings: Level 3 Bounded for M1 Direction, Level 1 Assisted for individual ratings in M2 Evaluation and for adverse M3 Consequence decisions, Level 1 Assisted for M4 Development.
M5 Care is asymmetric: a system may flag a risk or block a breach of a protective rule, but never relaxes one.
The function an action belongs to is set by its effect on the person, not by its label.
9 min
The named human decider and the delegation profile
Level 4 Autonomous is never used for M3 Consequence, and every consequence decision has a named human decider with six capabilities.
The rule answers both failure modes of accountability: the problem of many hands needs a name, and the moral crumple zone needs a name that carries the means of deciding.
The delegation profile records degree per function, deciders, the representatives' agreement and the stop path; DLG-NR-ROS-01 and DLG-HM-CTA-01 show the form.
What the audience takes away
Management can be divided into five functions, and the division decides what may be delegated to a system and at what degree.
Algorithmic management is Taylor's ambition in a new medium; what is new is scale and speed, opacity and a consequence without a human decision.
A consequence for a person always needs a named human decider who has the means to decide otherwise, recorded in a delegation profile.
Module 2 · Part II · The AI Boss Today · 50 minutes
The AI Boss Today: How Direction and Evaluation Become Consequence
Part II is a field trip through six workplaces, and this lecture carries the one argument they share. In each, a system that was introduced to direct or measure work accumulates data, the data becomes a rating, and the rating reaches pay, hours or access to work without anyone deciding that it should. Courts and authorities accepted a human decision only where there was evidence of what the human did. The lecture follows that drift from the platform to the office, and shows the instruments the chapters build against it: written input lists, a named decider, a practice test and a feature register.
9 min
The platform loop: allocation, pay, rating, deactivation
Dispatch platforms perform M1 Direction, M2 Evaluation and M3 Consequence in one loop, with blended scores the rated person cannot read.
The Amsterdam Court of Appeal accepted human involvement for the one driver Uber had interviewed, and not for three whose files held only notes.
Three steps carry the pattern into ordinary employers: behaviour is recorded, the record becomes an input, and the hidden rating reaches hours and pay.
8 min
When the system sets the pace
A rate system has a target, a measurement, a comparison and a follow-up; the governance questions concern a parameter and a link, not a model.
The Framework Directive requires the employer to alleviate work at a predetermined work-rate, so a target generated by software is assessed, adapted and consulted on.
M1 Direction becomes M5 Care when the system sets the pace, and the duty of care is exercised over parameters.
8 min
Access to work: hiring by machine
Where recruiters read only the top of a ranking, the ranking has settled the fate of everyone below it: delegation is set by use, not by the configuration screen.
Human recruiters discriminate too, and audits measure groups, not the individual rejected.
A screening configuration is in substance a selection guideline; Nordvik Region paused a pilot it could not explain and recorded why in AIDR-0601.
9 min
How a score becomes a decision
In the reported cases the score selected the person and the human role, where there was one, ratified the selection.
A score reaches M3 Consequence through three administrative channels: the pay input, the warning and the list.
Whether a score is advisory is a test of practice: Hansa Mutual found 97 per cent of pay proposals in one unit matching the score band, with a median of 48 seconds.
8 min
Rosters, queues and what they do to pay
Every weight in an optimiser's objective function is a management choice, recorded or not.
Small reversible allocations accumulate into different hours, premiums, learning and working lives.
The input list decides the classification, so it belongs in the delegation profile, with change classes that reopen negotiation.
8 min
The office: features nobody decided on
A meeting assistant's owner field is the point at which a summary becomes an instruction.
Four tests show when help becomes a boss: the addressee, the default, the trace and the decision.
The remedy is not to forbid the feature but to decide it, in a feature register agreed through the representatives' process.
What the audience takes away
Data collected to direct work becomes a rating, and the rating becomes a consequence, unless the drift is closed in advance in a written profile.
A human in the process counts only where there is evidence of what the human weighed and that the outcome could have changed.
Degrees of delegation are measured in practice, by reading rates, time and departures, not read from the configuration screen.
Module 3 · Part III · What It Does to People · 50 minutes
Five Debates and Where They Meet: What the AI Boss Does to People
Part III stages five debates in the same four steps, and this lecture shows what they have in common. In each, the case for concern describes what happens when design is left to default, and the case for calm describes what is possible when it is not; both compare the AI boss with real managers, not ideal ones. The positions meet on conditions that an organisation can design and negotiate: discretion, reasons and a reachable person; outcomes measured by group; purpose and aggregation of data; health limits inside the optimiser; and an evaluation agreed before go-live. The verdicts separate the observable from the expected and from speculation.
6 min
The form of the debate
Four steps before a verdict: the case for concern, the case for calm, where the positions meet and what can be done.
Each position is stated at its strongest, and what a source reports is kept apart from interpretation.
Every verdict is labelled per management function as observable, expected or speculation, and ends with what can be influenced.
9 min
Autonomy and dignity: discretion, reasons, a reachable person
Self-determination theory, the job characteristics model and Karasek's demand-control model show that a system's design choices are job design decisions.
In Lee's experiment people accepted algorithmic assignment about as readily as human assignment, resisted algorithmic evaluation, and missed recognition.
The harm is not inherent in delegating direction; it follows from how much discretion the design leaves.
9 min
Fairness: measure by group, correct the rule
Bias enters through the definition of good performance, proxies, historical data and feedback loops, so a system can pass every individual check and still disadvantage a group.
The appeal fails without notice, without reasons, without anyone to ask, or with a remedy that fixes one case and leaves the rule.
Hansa Mutual's test of claims allocation found a rule on consecutive working days that kept part-time handlers from complex claims.
8 min
Surveillance: what a system requires and what it invites
A visit-routing system requires the start and end of each visit; the by-products invite a figure per person, then a ranking, then a pay input.
Courts and authorities apply a consistent test of notice, necessity, proportionality and safeguards, and treat undifferentiated retention harshly.
The same data can protect or chill; purpose, aggregation, retention and access decide which.
9 min
Health and pace: care designed into direction
An optimiser removes slack because slack looks like waste, yet those are the minutes in which a person recovers.
A route at Strandvik complied with every working-time rule and still could not be worked; travel time became a hard constraint written as a test.
Care cannot be delegated to the system that directs the work, but it must be designed into it.
9 min
Does it work? Quarters and years
Controlled studies show real short-run gains from assistance, largest for novices; a meta-analysis of general monitoring found no average gain in performance.
Gaming and the ironies of automation take time, so first-quarter gains on the optimised metric are the least informative evidence.
Q4 stays open: most organisations do not know whether it works because they have not designed a way to find out.
What the audience takes away
In every debate the outcome depends on design and negotiation, not on the technology alone: what is left to default produces the case for concern.
The honest comparison is with real managers, and the honest evidence is outcomes measured by group, at team level and over years.
A verdict is only as good as its labels: say what is observable, what is expected and what remains speculation.
Module 4 · Part IV · Managing Machines · 50 minutes
Managing Machines: What Transfers, What Does Not, and Who Answers
Part IV asks whether managing agents is the same activity as managing people, and answers that it is not. Instruction, scope and the checking of output transfer to agents as engineering tasks; development, care and accountability are owed only to people and grow when agents arrive. The lecture follows that asymmetry through the agent team, the orchestrated chain and the mixed team, shows why the manager's own judgement is at risk when the work that taught it is automated, and ends with the decision-rights file in which the book states its answers to open questions Q1 and Q2 as rules.
9 min
The agent boss: claim and evidence
The vendor material reports expectations and intentions, not outcomes; people directing agents is observable in pilots, and every employee becoming an agent boss is speculation.
M1 Direction and M2 Evaluation transfer to agents as instruction, scope, budgets and the checking of a product.
M4 Development and M5 Care are owed only to people, and no agent a person directs holds any part of M3 Consequence over a person.
9 min
Boundaries in the environment, not in instructions
The cited cases show instructions lowering the rate of unwanted agent behaviour without setting a ceiling.
A mandate manifest is read by the systems that enforce it: own identity per agent, a never list, budgets and a stop path outside the agents.
A weekly routine measures the review, not the reviewers: seeded cases, unassisted samples and team-level figures.
9 min
Chains: where accountability ends
Many hands, the gap between answerability and liability, and the moral crumple zone explain why oversight of the last step is not oversight of the chain.
The rule for Q7: a named chain owner who owns the top mandate, holds one stop path outside the chain and can reconstruct every output that reached a person.
No link, at any depth, holds M3 Consequence over a person, not even at Level 0 Advisory.
8 min
The mixed team
Work is divided in three parts: work that stays with people, work agents do under supervision, and reserved learning work.
Two sets of figures are never compared; an agent's throughput is never a norm for people.
A handover is a recorded moment of transfer with time to check and a right to return without cost.
7 min
Who becomes a manager
Rostering, allocating and first assessments were where managerial intuition was calibrated; systems now do that work first.
The people who would have become managers remain; what narrows is the route by which they showed and developed the ability.
Keeping the craft is a design decision: unassisted planning, drafting first, seeded errors, reserved conversations and a competence record.
8 min
Decision rights written down
One file assigns each class of management decision a decider, a ceiling, stop rights and the rights fixed by law or agreement.
For a consequence accepted without reading, the decider answers to the person, the function owner for the conditions and the management body for the allocation.
Lowering a degree or pausing a system needs no prior approval; restart is deliberately more expensive than stop.
What the audience takes away
Direction and checking transfer to agents; development, care and accountability remain owed to people, and agents add to that duty.
Accountability for agents ends at a named person who can stop them and reconstruct what they did, with boundaries enforced outside the agents.
The consequences of a failed decision fall on whoever held the right that failed, not on whoever stood nearest to the system.
Europe's Answer: Reading the Law as Design Requirements
Part V argues that European law already contains most of what the book proposes, spread across instruments written at different times for different purposes. The lecture reads four bodies of law as one specification: the AI Act names the functions and the deployer's duties; data protection law requires a human decision that is real and an explanation of the logic actually applied; co-determination attaches to change, not to version numbers; and safety, equality and public law keep the duty with the employer, who needs evidence it can produce itself. Everything is presented as organisational and architectural implication, not as legal advice.
10 min
The AI Act names the functions
Annex III, point 4, covers recruitment and selection, decisions on terms, promotion and termination, task allocation on behaviour or traits, and monitoring and evaluation.
The dividing line for allocation runs through the optimiser's feature list: availability and competence are left alone, behavioural inputs are not.
Article 26(2) is the legal anchor of the named human decider: competence, training and authority are separate requirements, and employers most often omit the third.
The Digital Omnibus deferred the Annex III obligations to 2 December 2027; the prohibition on emotion inference at work has applied since 2 February 2025.
10 min
The right to a human decision
GDPR Article 22 binds every employer now, and the guidelines say a controller cannot avoid it by fabricating human involvement.
Whether a score determines the decision is a measurable fact about use: departure rate, reading below the line, information besides the score, and time per decision.
The Platform Work Directive specifies for platforms what the book recommends for every employer: a human being takes any decision to restrict, suspend or terminate.
10 min
Co-determination meets a system that changes monthly
Europe has four degrees of participation: information, consultation, negotiation, and co-determination or consent; the triggers are functional, not technical.
Two failure modes: freezing, where every change is a new introduction, and drift, where version 9 runs under an agreement for version 1.
The answer to Q8 is to govern change by class: renegotiation, approval within agreed bounds, or information, classified from the delegation profile's fields.
10 min
Duties that stay with the employer
Safety law: a protective measure is a feasibility condition, and only a test that tries to break it shows which kind the system implements.
Equality law: the burden of proof makes a disparate-impact test on every release the employer's defence file.
For a public authority, procurement is the decisive moment, because it fixes what the authority will be able to show.
5 min
From articles to controls
Each article becomes useful when it is a control with an owner, a trigger and evidence.
Classification, agreements and assessments are reopened by change, not by the calendar.
The law is in force in part, expected in part, and its enforcement in employment is still speculation.
What the audience takes away
The law asks the employer to point to a person who decided, with competence, authority and time, and to show it from records.
Agreements and classifications should follow what changes for the people managed, not release numbers or the supplier's label.
No duty of safety, equality or reasons can be met with a supplier's assurance; the employer must hold the evidence itself.
Module 6 · Part VI · Designing the Accountable AI Boss · 55 minutes
Designing the Accountable AI Boss: Record, Oversight, Appeal, Notice, Role
Part VI is the design part of the book, and its five chapters form one loop. A management decision that AI made or shaped leaves a record; the record shows whether the human in the chair really decided; the person affected is told that the system exists, can learn what it did and can contest it; a successful contest corrects the rule and not only the case; and the line manager, whose role is transformed and not abolished, holds the loop together. The lecture walks the loop with the book's two first contests, a route with no travel time at Nordvik Region and a quarterly score at Hansa Mutual.
11 min
The decision as a record
A Management Decision Record is about a person, so it is personal data, has a reader with rights, is evidence in disputes, and must be produced by the work and not written beside it.
The reasons are in the decider's words, and the system's version and proposal make the decider's part visible.
A schema rejects Level 4 Autonomous for M3 Consequence and a decider named system or agent, which makes the record a condition of the decision.
The delegation register shows the sum of delegation, runs reversibility tests and freezes further delegation where a test fails.
11 min
Oversight that is real
A person who signs in a second has not read; data protection law treats a human without influence as absent.
Six conditions make oversight real: time, authority to deviate, independent information, a measured deviation rate, no penalty for deviation and a stop path.
Oversight of the system can be sampled; the decision in an M3 Consequence case cannot.
The oversight-quality metric tests the design, not the people: aggregate by default, separate the uses, own figures only.
11 min
Contest as a condition of delegation
An explanation has four elements: the role of the system, the inputs about the person, an actionable counterfactual and the human decider.
In the appeal path the decider can say yes; only an independent reviewer can say no.
A successful contest corrects up to five things: the decision, the data, the rule, the model and the register, and the record closes only when the system correction passes its test.
Every degree above Level 1 Assisted rests on an appeal that reaches a person able to change the outcome.
10 min
Telling the managed
Five things are disclosed: existence, logic, inputs, the degree of delegation and the named decider.
The trade secret objection limits the form of disclosure, not its fact.
A notice nobody reads informs nobody: layer it, time it at the moment of relevance, and have the line manager explain it.
Change classes close the route by which Mode 0 Ambient AI features become managers without anyone being told.
12 min
The transformed manager
Work that concerns the work moves towards systems under written delegation; work that concerns the person stays with the manager and gains weight.
The AI boss is acceptable only where it frees a manager to do what the evidence on good managers describes, and must be evaluated on whether it does.
The role charter adds a time budget with a protected floor for M4 Development and M5 Care, and a recorded decision on what happens to time saved.
The thesis is falsifiable, and the author's projections are labelled as his own judgement.
What the audience takes away
Records, oversight, appeal and notice are one loop: each depends on the others, and a gap in any one undoes the degree of delegation it supports.
Measure the review and not the reviewers, and correct the system and not only the case.
The manager's role is transformed, not abolished: time saved by systems becomes development and care only if a charter protects it.
Module 7 · Part VII · Experience and Verdict · 50 minutes
Experience and Verdict: A Year, Twelve Anti-Patterns and a Named Person in the Chair
The closing lecture puts the instruments of the book to work together and then judges them. A year at two organisations shows that the order of the instruments matters, and that every failure was found by a designed mechanism and none by a system's own account. Documented cases show one recurring movement: a system built to direct or evaluate comes to carry a consequence without anyone deciding it. The two organisations reach different plans because they hold different levers. The thesis holds with qualifications, seven of twelve open questions are answered, and the book adds a sixth commitment to the series: a named person in the chair for every consequence.
11 min
A year in the order it needed
Commission, inventory, profiles, records, and only then measured oversight and reversibility tests; where the order was reversed, the cost appeared in the next record.
A contest about one route became a hard constraint, an appeal path and a notice change within two months.
Co-determination shaped every outcome that lasted, from the rotating order of open shifts to the workload indicator.
Nordvik's profiles moved sideways and Hansa's inward; none moved upward.
10 min
The one movement in the cases
In the eight documented cases a system built for M1 Direction or M2 Evaluation came to carry M3 Consequence, and the record, the decision right or the contest path that would have shown it was missing.
Twelve anti-patterns fall into four families: delegation unrecorded, oversight in name only, people measured as machines, voice and craft lost.
The anti-patterns feed each other, and a checklist of signals turns them into a quarterly review that cannot be argued away.
10 min
Two organisations, two levers
Nordvik mostly procures, so its lever is the contract and the acceptance test; Hansa mostly builds, so its lever is the allocation rule and the release.
Every degree raised during the year was protective or confined to agents' own actions; every degree touching evaluation or consequences was held, lowered or refused.
Both 24-month plans put the measures that make delegation safe before the measures that extend it, and send at most three decisions to the management body.
11 min
The thesis judged
Direction may be delegated up to Level 3 Bounded only with health constraints inside the optimiser and its figures fenced off from consequence decisions.
The named human decider is the firmest claim and the one most easily emptied; it holds where the six capabilities hold.
Records are supported as design while their effect is unmeasured, and reversibility decays within months of go-live.
Q4, Q5, Q10 and Q11 are answered in part; Q12 is open by nature and left to the author's labelled verdict.
8 min
What to do first
First map and write authority, next record and measure, later test reversibility and evaluate over years; raise no degree before the records exist.
Ten questions test each year whether the organisation governs its AI boss or is governed by it.
If you are managed by an AI boss, ask who your decider is: that question is the beginning of every correction the book describes.
What the audience takes away
What makes delegation safe is the human capacity to check it, and the mechanisms that find failures must be designed before the pressure arrives.
The lever differs, contract or code, consent or negotiation, but the rule is shared: no system holds a consequence decision about a person beyond Level 1 Assisted.
The thesis holds with qualifications and names what would refute it; a verdict that cannot be revised is a forecast in disguise.
