Scheduling, Pay and the Allocation of Work

Most people who are managed by software never see a dispatch app or a rate display. They see a roster. Hospitals, care services, shops, contact centres and claims departments have drafted rosters and allocated work items with optimisation software for years, and the newest generation adds demand forecasting, automatic offers of open shifts and agents that place work in one person's queue rather than another's. This is where algorithmic management is most common outside platform work and least visible, because a roster looks like an administrative product and an allocation looks like routing. Yet a roster decides when a person sleeps, how many hours an hourly worker is paid for and who receives the supplements for evening and weekend work; an allocation decides who gets the complex cases and the ones from which people learn. This chapter describes how these systems direct work, what research measured about unstable schedules, when allocation becomes a decision about pay, which European rules apply, and which degree of delegation fits. It concerns three management functions: M1 Direction, because rosters and allocations assign, sequence and pace work; M3 Consequence, because hours, premiums and the mix of work reach pay and access to work; and M5 Care, because working time and rest are where the duty of care meets the optimiser.

Figure 9.1: 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.

Rosters and routes by forecast

A rostering system solves a constrained optimisation problem. Its inputs are a forecast of demand, the staff available with their contracts and competences, the rules on working time and rest, the rules of the collective agreement, and the preferences staff have entered. Its output is a roster: who works which shift, on which ward or route, over a period of usually two to six weeks. For mobile workers such as home-care staff the same system also sequences visits into routes and allocates travel time between them. The objective function weighs coverage against cost, and preferences against both. Every one of those weights is a management choice, whether or not anyone records it as one.

The forecast is the part that changed most. Older systems repeated a template roster for a manager to adjust. Newer systems forecast demand per hour from history, bookings and, in care, the needs assessments of the people served, and draft the roster that meets the forecast at least cost. Where the forecast is updated daily, the roster can be redrafted daily too, and the cost of staffing to a tighter margin falls on the workers whose hours move. This is the mechanism behind what is called just-in-time scheduling: staffing that follows forecast demand closely, with short notice, cancelled or added shifts and on-call periods taking up the variance that used to sit in the employer's labour budget.

The same tools can do the opposite. A forecast can be used to publish rosters further ahead, to keep a stable core of fixed shifts and to reserve flexibility for a defined set of open shifts that staff may choose to take. Which way the tool is used is set by the objectives and constraints the employer configures and, in most of Europe, negotiates. Nordvik Region's ShiftWise drafts rosters for the Region's two hospitals and routes for its home care; Hansa Mutual's claims-triage agent allocates claims to individual handlers. Both are examined below.

Unstable schedules

The best-known press account of algorithmic scheduling is Jodi Kantor's report for The New York Times on a Starbucks barista in San Diego, a single mother who rarely learned her schedule more than three days before the start of a working week, whose shifting hours put her degree course on indefinite pause and who feared that asking for more stable hours would cost her hours overall (199, Kantor, 2014). A report of one person's experience shows what the pattern can do, not how common it is or what it does to health; the research that followed measured both.

Schneider and Harknett surveyed 27,792 hourly retail and food-service workers employed at 80 large firms in the United States, recruited through targeted advertisements on Facebook, and measured exposure to six kinds of instability: variable hours, short advance notice, cancelled shifts, on-call work, "clopening" shifts in which a worker closes late and opens early the next morning, and having no input into the schedule (125, Schneider & Harknett, 2019). They reported that 63 per cent of the workers had less than two weeks' notice of their schedule and 34 per cent less than one week. Exposure to routine instability was associated with psychological distress, poor sleep quality and unhappiness, and these associations were much stronger than those with low wages. Workers with fewer than three days' notice, and with three to six days' notice, fared significantly worse than those with more than two weeks; the authors found no difference between one to two weeks and longer. The association ran partly through household economic insecurity and to a larger extent through work-life conflict. The authors state that these cross-sectional estimates are not causal.

A later study supplied the causal step. Seattle's Secure Scheduling Ordinance, in force from 2017, required employers in its scope to give two weeks' notice of schedules among other provisions. Harknett, Schneider and Irwin compared workers in Seattle with workers in comparison cities before and after the law, using a difference-in-differences design, and found that the law increased schedule predictability and stability and improved subjective well-being, sleep quality and economic security, with no effect on reported psychological distress (220, Harknett et al., 2021). Using the law as an instrument, they estimated that uncertainty about working time has a substantial effect on well-being, particularly on sleep and economic security.

Two further findings matter for European rostering. The first concerns rest between shifts. Vedaa and colleagues reviewed 22 studies of "quick returns", eleven hours or less between two shifts, which are common in rotating shift work in health care; they found indications of detrimental effects on sleep, sleepiness and fatigue, while the evidence on chronic physical and mental health and work-life balance was inconclusive (222, Vedaa et al., 2016). The second concerns cost. In a randomised field experiment in 28 Gap stores in the San Francisco and Chicago areas from November 2015 to August 2016, a package of scheduling practices that improved consistency, predictability, adequacy of hours and employee control increased store labour productivity by 5.1 per cent, through 3.3 per cent higher sales and 1.8 per cent lower labour hours (221, Kesavan et al., 2022). The study covers one retailer, which co-funded it, but it is evidence that stable schedules are not simply a cost conceded to workers.

The research does not settle how far American retail findings carry to Nordic hospitals and German offices, where notice periods and collective agreements already limit the worst practices. But the mechanisms measured, short notice, variable hours, quick returns and lack of control, are ones a European optimiser can produce inside the law unless it is configured not to.

Allocation by queue and agent

Allocation is the finer-grained sibling of rostering. Where a roster decides when a person works, an allocation decides what the person works on. In offices, cases, tickets and claims arrive in a queue, and a rule, a model or an agent assigns each item to a person. In health care, patients on a ward are assigned to nurses, home-care visits are assigned to staff and sequenced into routes, and theatre lists and on-call duties are distributed among clinicians. The allocator weighs skill, authority, availability, workload and sometimes speed.

Each allocation is small and reversible, which is why allocation is easily delegated. The effects, however, are cumulative. A person who receives the routine items for a year has a different working life from a colleague who receives the complex ones: different learning, different visibility, different figures in the throughput report and, eventually, a different case at the annual review. The European Commission's draft guidelines on the classification of high-risk AI systems put the point directly: an allocation of tasks "is not just an operational arrangement, but may be a determinant of professional development", because complex, visible or lucrative assignments may open pathways to advancement while repetitive or less valued tasks may in practice preclude promotion and affect current or future income (paragraph 267) (3, European Commission, 2026).

Allocation also produces figures. A system that places items in queues knows how many each person received, how long each stayed open and how each was closed. Those figures are a by-product of M1 Direction and become an input to M2 Evaluation unless someone decides otherwise. Hansa Mutual's claims-triage agent, a Mode 5 Autonomous production agents system, produces throughput figures per handler, which under the group works agreement KBV-IT-2024 go to team leaders only. Chapter 8 followed such figures towards pay and dismissal; this chapter concerns the allocation that generates them.

Pay shaped by software

Software shapes pay in three ways, from the most indirect to the most direct.

The first is through hours: for an hourly or part-time worker, the roster is the pay slip, and a system that concentrates extra hours on those who accept fastest or are most available distributes income. The second is through premiums. Collective agreements in health and care pay supplements for evening, night and weekend work, so whoever the optimiser places on those shifts receives more. A roster that is neutral in hours can be unequal in pay. The third is direct: the system calculates the pay for a piece of work itself, as platforms do when they set the fare for a ride or a delivery.

The third form is where the evidence is strongest and the debate sharpest. Dubal's long-term ethnographic study of ride-hail drivers in California names the practice "algorithmic wage discrimination", in which individual workers are paid different hourly wages, calculated with ever-changing formulas from granular data on location, individual behaviour, demand, supply or other factors, for broadly similar work (144, Dubal, 2023). Her definition matters beyond platforms because it encompasses "digitalized decisions to allocate work, which are significant determinants of hourly wages", and because workers under such systems find their hourly wages hard to predict. Dubal argues that the practice runs against long-standing norms of fair wage setting and the spirit of equal pay law, and proposes a non-waivable restriction on it.

The strongest counter-evidence comes from the same industry. Cook and colleagues studied more than a million Uber drivers in the United States and found a gender earnings gap of roughly 7 per cent, which they attribute entirely to experience on the platform, preferences and constraints over where to work, and driving speed; they found no differential effect of taste for specific hours and no customer discrimination (139, Cook et al., 2021). The pay algorithm in their study does not use gender. The finding cuts both ways: a formula applied identically to everyone need not discriminate in its design, yet a gender-blind system operating amid unequal constraints can produce a persistent pay gap, which equal-pay law asks employers to see and explain.

The draft Commission guidelines treat the direct form as high-risk: a ride-hailing system that adjusts individual drivers' pay multipliers from acceptance rates, passenger ratings and completion times falls within Annex III, point 4(b), because remuneration decisions fall within the terms of a work-related relationship (3, European Commission, 2026). Platform work is the subject of Chapter 5, and the Platform Work Directive defines automated decision-making systems to include those that take or support decisions affecting earnings, "including the pricing of individual assignments", and working time (Article 2(1)(i)) (47, European Union, 2024). Neither Nordvik Region nor Hansa Mutual sets pay by algorithm; both shape it through hours, premiums and the mix of work, so the question for them is whether anyone watches what the roster and the queue do to pay.

Working time, recorded

The Working Time Directive sets the floor on which every European roster is built. Every worker is entitled to a minimum daily rest of 11 consecutive hours per 24-hour period (Article 3), a weekly uninterrupted rest of 24 hours plus the 11 hours of daily rest per seven-day period (Article 5), and average weekly working time, including overtime, of no more than 48 hours (Article 6) (50, European Union, 2003). Article 13 requires an employer who organises work according to a pattern to take account of the principle of adapting work to the worker. Article 17 permits derogations, among them from the rest rules for hospitals and similar establishments (Article 17(3)(c)(i)) and for shift work each time a worker changes shift (Article 17(4)(a)), provided that equivalent compensatory rest is granted; Article 18 allows derogations by collective agreement.

In 2019 the Court of Justice added a duty that matters directly to software. In CCOO v Deutsche Bank, the Grand Chamber held that Articles 3, 5 and 6 of the Directive, read in the light of Article 31(2) of the Charter of Fundamental Rights, and provisions of the Framework Directive 89/391/EEC preclude national law that does not require employers to set up a system enabling the duration of time worked each day by each worker to be measured; the Member States "must require employers to set up an objective, reliable and accessible system" for that purpose (paragraph 60) (66, CJEU, 2019) (48, European Union, 2012) (49, Council of the European Communities, 1989). For a rostering system this means that the planned roster is not the record of time worked, and rest rules must be checked against actual time: a roster that complies on paper can breach the rules once swaps, overruns and sick-leave cover are added.

National law adds detail, and the Region's national rules are illustrated here by Swedish law. The Swedish Working Hours Act requires employers to notify employees of changes in the scheduling of ordinary working time and on-call time at least two weeks in advance, unless the nature of the activity or unforeseeable events justify shorter notice (§ 12); it sets daily rest of at least eleven consecutive hours in every 24-hour period (§ 13) and weekly rest of at least 36 consecutive hours in every seven-day period (§ 14); and it permits departures by collective agreement within the limits of the Directive (§ 3) (219, Sveriges riksdag, 1982). The optimiser's rule set must therefore follow the applicable collective agreement, and change when it changes.

The Directive on transparent and predictable working conditions, Directive (EU) 2019/1152, adds rules on predictability that read like a specification for scheduling software (53, European Union, 2019). Where a work pattern is mostly predictable, the employer must inform the worker of the standard working day or week and any arrangements for shift changes (Article 4(2)(l)). Where it is entirely or mostly unpredictable, the employer must state the guaranteed paid hours, the reference hours and days within which the worker may be required to work, and the minimum notice period (Article 4(2)(m)). Under Article 10, such a worker may refuse, without adverse consequences, an assignment outside the reference hours and days or without reasonable notice, and is entitled to compensation for a late cancellation where national law allows cancellation without it. Article 12 gives a worker with six months' service the right to request more predictable and secure working conditions and to receive a reasoned written reply within one month.

Read together, these instruments turn rest rules, notice periods, reference hours and the right to refuse into constraints the optimiser must respect and evidence the employer must be able to produce. ShiftWise flags rest breaches at Level 1 Assisted and the unit manager resolves each one, because every resolution moves someone else's rest or hours, and the duty of care belongs to the manager.

Task allocation as high-risk

The AI Act, Regulation (EU) 2024/1689, lists among the high-risk use cases in Annex III, point 4(b), "AI systems intended to be used to make decisions affecting terms of work-related relationships, the promotion or termination of work-related contractual relationships, to allocate tasks based on individual behaviour or personal traits or characteristics or to monitor and evaluate the performance and behaviour of persons in such relationships" (1, European Union, 2024). Recital 57 gives the reason: an appreciable impact on career prospects, livelihoods and workers' rights, and the risk of perpetuating historical discrimination. Article 6(3) lets an Annex III system escape the classification where it poses no significant risk of harm, but never where it performs profiling of natural persons.

The draft guidelines draw the line for allocation in four steps (3, European Commission, 2026). Behavioural indicators such as punctuality, responsiveness, reliability metrics or performance ratings count as personal traits or characteristics (paragraph 270). The dividing line is whether an attribute pertains to the worker's personal traits or social identity, rather than to neutral, objective and external factors such as unavailability or location (paragraph 271). Allocation based on availability as workers indicate it in scheduling tools, on a required accreditation or on geographical proximity falls outside point 4(b) (paragraph 272). And not every day-to-day managerial adjustment is a decision affecting terms of the work-related relationship: the decision must reach a threshold of significance (paragraph 262). Among its practical examples of systems inside point 4(b) is an AI-enabled scheduler that ranks workers for each slot from punctuality, no-show history, acceptance and decline rates for offered shifts and ratings, allocates the most important shifts to the highest-ranked and downgrades priority when a worker declines a priority offer; such a system cannot benefit from Article 6(3). The guidelines are a draft, and their final text may change.

The draft thus gives a practical test. A system that allocates by availability, competence, authorisation and location is outside point 4(b) on that ground, though it may still fall within it by affecting terms such as working hours. A system that also weighs how a person behaved, how fast they worked or how often they accepted offers is inside. The input list decides the classification, so it belongs in the delegation profile, and adding an input is a change that needs review. Nordvik Region classifies ShiftWise as high-risk under point 4(b), because rosters change the working hours of part-time and hourly staff and the Region did not want the classification to depend on the supplier's choice of inputs. Hansa Mutual reopened the classification of its claims-triage agent, made for its insurance function, for its work-allocation function as AIDR-0651; Chapter 21 follows that reassessment.

Classification brings duties. A deployer who is an employer must, before putting a high-risk system into service or use at the workplace, inform workers' representatives and the affected workers (Article 26(7)), and must assign human oversight to people with the necessary competence, training and authority (Article 26(2)). Under the Digital Omnibus, Regulation (EU) 2026/1744, the obligations of Chapter III, Sections 1 to 3, apply to Annex III systems from 2 December 2027 (2, European Union, 2026). That date is expected to change how suppliers of rostering and workforce-management software document their inputs; an employer that negotiates the input list now does not need to wait for it.

Pay transparency

The Pay Transparency Directive, Directive (EU) 2023/970, reaches the pay that rosters and allocations shape (57, European Union, 2023). Its definition of pay includes "complementary or variable components" as well as basic salary (Article 3(1)(a)), so premiums for unsocial hours are pay. Employers must make easily accessible to workers the criteria used to determine pay, pay levels and pay progression, and those criteria must be objective and gender-neutral (Article 6(1)). Workers may request in writing their individual pay level and the average pay levels, broken down by sex, for workers doing the same work or work of equal value, and receive the answer within two months (Article 7). Employers with 250 or more workers must report the gender pay gap, including the gap in complementary or variable components and the proportion of women and men receiving them, by 7 June 2027 and every year thereafter (Article 9). The Directive had to be transposed by 7 June 2026 (Article 34).

If premiums are pay, the rule that distributes premium-bearing shifts is in effect a pay-setting criterion, even if nobody wrote it as one. An optimiser that gives evening and weekend shifts to those with the widest availability will, in a workforce where women carry more of the care at home, tend to give them to men, or to staff without children, which is the pattern Cook and colleagues found in a different setting. The employer must be able to state the rule, show that it is objective and gender-neutral, and report its effect. That is a design requirement for the objective function and its reports.

The right degree of delegation

The series' autonomy levels apply here per management function. The line between Level 2 Supervised and Level 3 Bounded is the most important one: at Level 2 a person approves each action, while at Level 3 the system acts within a codified playbook enforced outside the model, and the level is decided and recorded, not discovered (AI Red Team, Chapter 6; F50) (43, Nordqvist, 2026). Applied to scheduling and allocation, the reasoning runs as follows.

A roster sets a person's hours, rest and premiums for a month; once published, changing it costs the people whose plans depend on it. That makes Level 2 Supervised the right degree for M1 Direction in rostering: the system drafts, and a named manager approves the roster as a whole, with the power to change any shift. An open-shift offer to staff who opted in is narrower and declinable without consequence, which fits Level 3 Bounded if the bounds are written: who receives offers, in what order, and with what effect of declining. Rest-rule checks belong at Level 1 Assisted for M5 Care, because the system detects and the manager resolves. Allocation of single work items, such as claims, can sit at Level 3 Bounded within allocation rules, because each item is reversible and a person can reassign; but because the effect is cumulative, the profile must also record the right to reassign and a regular review of allocation patterns by person. Level 4 Autonomous has no place in either: a roster or an allocation that reaches pay or access to work must not run without a person who can be named, and Level 4 is never used for M3 Consequence.

Two further conclusions of the earlier volumes apply. Stop decisions are named in advance, and a stop path that runs inside the system it stops fails when it is needed (AI Governance, Chapter 13; F19) (41, Nordqvist, 2026); for a rostering system, that means the unit manager can withdraw an automated offer or revert to a manual roster without asking the supplier. And small delegations add up until the fallback no longer exists, which a delegation register and reversibility tests are meant to catch (AI Threat, Chapter 26; F69) (44, Nordqvist, 2026). A unit that has not drafted a roster by hand for three years has delegated more than its profile says. These are this book's Q1 (which management functions may be delegated, to what degree and who decides) and Q8 (how co-determination applies to a system whose rules change faster than agreements).

Co-determination answers part of Q8 already. In Germany, the works council co-determines the beginning and end of daily working time, breaks and the distribution of working time over the days of the week (BetrVG §87(1) no. 2), temporary reductions or extensions of normal working time (no. 3), technical equipment intended to monitor behaviour or performance (no. 6), questions of pay structure including new pay methods (no. 10), and performance-related pay rates (no. 11); where no agreement is reached, the conciliation committee decides (§87(2)) (9, Germany, 2001). A rostering system touches nos. 2 and 3 by its very purpose. In Sweden, the co-determination act requires an employer bound by a collective agreement to negotiate on its own initiative before deciding on an important change in its activity or in the working or employment conditions of union members (MBL §11), and to keep the union informed of the guidelines of its personnel policy (MBL §19) (205, Sveriges riksdag, 1976). Neither statute was written for software that changes weekly, so the profile must say which changes count as important.

Practical guide: Negotiating a delegation profile for a rostering or allocation system with unions or a works council

Purpose. Agree, before go-live, what the system may do in each management function, at which degree, under whose approval, with which inputs and with which reviews, and which changes reopen the agreement.

Roles. The employer's system owner (in a public authority, the head of the service; in a company, the operations owner); the line managers who will approve rosters or reassign work; HR and labour relations; the union representatives or the works council; the data protection officer; the safety representative or the chair of the work-environment committee; and a group of affected workers who have used a pilot.

Steps.

1. List every decision the system takes or supports, and classify each by management function: drafting the roster, sequencing routes, offering open shifts, flagging rest breaches, allocating items, producing figures per person.
2. For each decision, propose a degree of delegation and a named decider by role. Default to Level 2 Supervised for rosters, Level 3 Bounded only for narrow, declinable actions such as open-shift offers to staff who opted in, and Level 1 Assisted for working-time and rest checks.
3. Publish the input list. State which inputs the optimiser or allocator uses and which it must not use: no ratings, acceptance rates, sickness absence or other behavioural indicators unless the parties agree otherwise and the AI Act classification is revisited.
4. Write the objective function in plain language: what the system optimises, in which order, and how preferences are weighted against cost. Agree the rules for distributing premium-bearing shifts and extra hours.
5. Encode the working-time rules from the applicable collective agreement as constraints, with the notice period for publication. A default of publishing rosters at least two weeks before they start follows the Swedish statute and the Seattle evidence; adapt it to the agreement.
6. Agree the rights of the person managed: to see the roster's basis, to state preferences, to decline an offer without consequence, and to contest an allocation or a roster (Chapter 27).
7. Agree a review of patterns by person, monthly at first and quarterly once stable: hours offered and accepted, premium-bearing shifts, quick returns, items allocated by complexity tier, and the share of decisions changed by managers. Report to the representatives at group level, and to individuals only about themselves.
8. Agree change classes. Changes to the objective function, the input list or the rules for offers are important changes that reopen negotiation; forecast refreshes and bug fixes are notified.
9. Record the agreement as a delegation profile in the delegation register, and the decision to adopt it as an AI Decision Record.

Definition of done. The profile names a decider for every decision, lists inputs and prohibited inputs, carries the change classes and review dates, and has been agreed with the representatives; the first roster or allocation approved under it has produced a Management Decision Record.

Worked example: DLG-NR-ROS-01 and the first approved roster MDR-NR-0001

Illustrative. Nordvik Region procured ShiftWise under tender NR-2025-089 and deploys it as a high-risk system under Annex III, point 4(b); its fundamental rights impact assessment is FRIA-2026-007. Before the system went live in home care, the Region negotiated the delegation profile with the public-sector unions under the co-determination act and recorded the decision as AIDR-0602. The Strandvik home-care unit, led by Mikael Strand, is the first home-care unit to use it.

# management/delegation-register.yaml (excerpt, illustrative)
id: DLG-NR-ROS-01
system: ShiftWise
organisation: nordvik-region
usage_mode: Mode 4 Built AI applications
deployer_role: deployer
classification: high-risk, AI Act Annex III point 4(b)
fria: FRIA-2026-007
decision_record: AIDR-0602
scope: [hospital wards, home care]
functions:
  M1_Direction:
    roster_and_routes: {level: Level 2 Supervised, decider: unit manager, approves: whole roster}
    open_shift_offers: {level: Level 3 Bounded, decider: unit manager sets the bounds}
  M2_Evaluation: not_delegated
  M3_Consequence: not_delegated
  M5_Care:
    rest_rule_checks: {level: Level 1 Assisted, resolves: unit manager}
open_shift_bounds:
  recipients: staff who opted in
  order: published rotating order, moving one place with each offer; nothing about the person enters it
  declining: no effect on future offers or hours
  condition: any ordering based on a personal aspect requires renegotiation and a new Article 22 assessment (AIDR-0626)
inputs_allowed: [forecast demand, contract hours, competences, stated preferences, availability, working-time rules]
inputs_prohibited: [acceptance rates, sickness absence, dashboard figures, ratings]
constraints:
  source: sector collective agreement
  publication_notice_days: 14
premium_shifts: distributed by the unit's rotation rule, reported monthly by person
stop: unit manager may withdraw offers and revert to a manual roster at any time
change_classes:
  renegotiate: [objective function, inputs, open-shift rules]
  notify: [forecast refresh, bug fix]
review:
  pattern_review: monthly, unit level to the union, individual to the person only
  profile_review: "2027-08-01"
agreed_with: [union for municipal and care workers, union for nurses and midwives]

The first four-week roster for the Strandvik unit under this profile shows the profile at work. ShiftWise drafted the roster; the rest-rule check flagged one breach; Mikael Strand changed three shifts and approved the rest.

# management/decisions/MDR-NR-0001.yaml (illustrative)
id: MDR-NR-0001
organisation: nordvik-region
unit: Strandvik home care
decision: approve four-week roster
period: {from: "2027-02-08", to: "2027-03-07"}
function: M1 Direction
delegation_profile: DLG-NR-ROS-01
level: Level 2 Supervised
system_output: ShiftWise draft roster, version 3
named_decider: Mikael Strand, unit manager
flags_from_system:
  - {type: daily rest under 11 hours, shift: evening route to morning route, resolved: true}
changes:
  - shift: assistant nurse A, 2027-02-09 morning route
    change: swapped with a colleague on the same route
    reason: resolve flagged rest breach after an evening route ending 22:00
  - shift: care assistant B, Wednesday afternoons
    change: restored a fixed afternoon off in all four weeks
    reason: stated preference for childcare outweighed in the draft by travel cost
  - shift: care assistant C, 2027-02-20 weekend evening
    change: moved to an hourly colleague who asked for more hours
    reason: draft gave C four premium-bearing shifts against a unit rotation of three
drafted: "2027-01-21"
published: "2027-01-22"
notice_days: 17
open_shifts_offered: 6
affected_staff_informed: roster published in the staff app with the reasons for changes
contest_via: unit manager, then HR business partner

The record does what the profile promised. It shows what the system proposed, what the manager changed and why, and the notice given. It names the person who answers for the roster; staff appear here by letter, while the Region's own record carries employee identifiers under its access rules. The third change is the one that matters for pay: without a rule for premium-bearing shifts, the optimiser had concentrated them on the person with the widest availability.

Worked example: DLG-HM-CTA-01 and the first override MDR-HM-0001

Illustrative. At Hansa Mutual the claims-triage agent allocates incoming claims to handlers within allocation rules at Level 3 Bounded. Team leaders can reassign. Jana Petersen leads about 20 claims handlers in Hamburg.

# management/delegation-register.yaml (excerpt, illustrative)
id: DLG-HM-CTA-01
system: claims-triage agent, allocation to handlers
organisation: hansa-mutual
usage_mode: Mode 5 Autonomous production agents
decision_records: [AIDR-0024, AIDR-0034, AIDR-0063, AIDR-0651]
classification: work-allocation function under reassessment, AIDR-0651
functions:
  M1_Direction:
    allocation: {level: Level 3 Bounded, rules: allocation-rules v4, reassign: team leader}
  M2_Evaluation:
    throughput_figures: {level: Level 1 Assisted, individual_figures: team leader only, basis: KBV-IT-2024}
  M3_Consequence: prohibited; allocation data not used for pay, discipline or dismissal
  M5_Care:
    capacity: team leader enters available capacity per handler, without reasons
allocation_inputs: [line of business, complexity tier, authority limit, language, open workload, capacity]
allocation_inputs_prohibited: [throughput per handler, quality scores, absence history]
pattern_review: monthly by handler, complexity tier and volume, team-level report to the works council
co_determination: Konzernbetriebsrat under BetrVG 87(1) no. 6, KBV-IT-2024
review: "2027-06-30"

In the record below, the agent had allocated nine new motor claims in one morning to a handler returning from long-term absence on a reduced schedule agreed with HR. The agent read her open workload, which was low after her absence, and had no means of knowing about the reduced schedule, because health information is deliberately kept out of the claims system.

# management/decisions/MDR-HM-0001.yaml (illustrative)
id: MDR-HM-0001
organisation: hansa-mutual
team: motor claims, Hamburg
decision: reassign claims allocated by the claims-triage agent
function: M1 Direction
delegation_profile: DLG-HM-CTA-01
system_level: Level 3 Bounded
named_decider: Jana Petersen, claims team leader
items_reassigned: 5 of 9
reason: handler on an agreed reduced schedule; agent read open workload only
health_data_recorded: none
correction:
  immediate: claims reassigned to three colleagues within authority limits
  lasting: capacity field added to allocation inputs, entered by team leader without reasons
change_class: allocation-rule change, notified to the works council under KBV-IT-2024
affected_handler_informed: true
date: "2027-01-12"

The correction was the valuable part: a new input, capacity, which lets the team leader protect a person without telling the system why. The record shows that Level 3 Bounded works only because a person watches the pattern and can reassign. It also shows the one function the agent cannot perform, M5 Care, because care depends on information the agent should never hold.

Across the degrees of delegation

Decision	Typical mode	Degree that fits	Named decider	What must be recorded
Drafting a roster or route	Mode 4 Built AI applications	Level 2 Supervised	Unit manager approves the whole roster	MDR with changes, reasons and notice
Offering open shifts to staff who opted in	Mode 4 Built AI applications	Level 3 Bounded, rotating order	Unit manager sets bounds	Offers sent, accepted, declined; no effect of declining
Rest-rule and working-time checks	Mode 4 Built AI applications	Level 1 Assisted	Unit manager resolves	Flag and resolution, against actual time
Allocating cases or claims	Mode 5 Autonomous production agents	Level 3 Bounded within rules	Team leader may reassign	Override MDRs; monthly pattern review
Distributing premiums and extra hours	Inside the roster	Level 2 Supervised, by agreed rule	Unit manager	Monthly report by person; pay-gap reporting
Setting individual pay per task	Platform pay engines	Not delegated in either organisation	Employer, under the collective agreement	Not applicable

Forecast-driven rostering, open-shift offers and agent allocation are observable in both organisations and across Europe. Agents that also negotiate swaps between workers and propose premiums are expected, because workforce-management products are adding agentic features; each such feature reaches M3 Consequence and belongs in a renegotiated profile. Fully autonomous scheduling without approval in a regulated care setting is speculation, and under this book's rule it would remain outside the profile even if it became technically routine.

Nordvik Region and Hansa Mutual

At Nordvik Region the AI boss directs shift work in care through a procured, high-risk system. ShiftWise drafts four-week rosters for the hospital wards and visit schedules and routes for home care at Level 2 Supervised, offers open shifts to staff who opted in, in a published rotating order, at Level 3 Bounded, and flags rest-rule breaches at Level 1 Assisted for M5 Care. The profile DLG-NR-ROS-01 was negotiated with the unions before go-live in home care (AIDR-0602), and MDR-NR-0001 records what Mikael Strand changed. Allocation reaches earnings directly at Nordvik: through hours for the third of the Strandvik unit on part-time or hourly contracts, and through the supplements for evening, night and weekend work set in the collective agreements. Its monthly pattern review is a pay control as much as a care control; Chapter 27 follows an assistant nurse's contestation of a Strandvik route.

At Hansa Mutual the AI boss directs office work through a built agent. The claims-triage agent allocates claims at Level 3 Bounded, team leaders can reassign, and MDR-HM-0001 records Jana Petersen's first override. Allocation reaches earnings only indirectly, through the annual objectives cycle in which a handler's results depend in part on the claims the handler was given. Hansa's challenge is legal plurality: the same allocation rule is co-determined by the Konzernbetriebsrat in Germany, needs consent from the ondernemingsraad in the Netherlands and negotiation with the unions in Sweden, and AIDR-0651 must settle whether the work-allocation function is high-risk before 2 December 2027. The difference is instructive: Nordvik's system is plainly a roster that affects pay and was negotiated as one, while Hansa's agent was introduced as claims routing and is only now being treated as a manager of people.

Summary

Rosters and allocations are the most common and least visible form of algorithmic management outside platform work: optimisation problems whose objectives are management choices, and small reversible decisions that accumulate into different working lives. Both direct work, which places them in M1 Direction. Both reach pay and access to work through hours, premiums and the mix of tasks, which places them in M3 Consequence. And both decide when people rest and how much they carry, which places them in M5 Care.

The research is clear on the mechanisms. Schedule instability and lack of control were associated with distress and poor sleep among 27,792 American service workers; a two-week notice rule improved well-being, sleep and economic security; quick returns are associated with acute sleep and fatigue problems; and stable scheduling raised productivity at one retailer. Pay set by formula can be unpredictable, and a gender-blind formula can still produce a pay gap.

European law already supplies most of the constraints: rest and working-time limits, the duty to record actual working time after CCOO, notice and the right to refuse in Directive (EU) 2019/1152, objective and gender-neutral pay criteria in the Pay Transparency Directive, and the high-risk listing of task allocation based on behaviour or traits in Annex III, point 4(b), of the AI Act, with deployer duties from 2 December 2027. Co-determination makes the delegation profile a negotiated document.

The degree of delegation follows from reversibility and accumulation: Level 2 Supervised for rosters, Level 3 Bounded for narrow, declinable offers and for allocation of single items with a right to reassign, Level 1 Assisted for rest checks, and never Level 4 Autonomous where pay or access to work is at stake. What makes those levels real is a written input list, change classes, Management Decision Records for approvals and overrides, and a regular review of patterns by person. Chapter 10 follows the AI boss into knowledge work, where assistants and workflow agents assign, summarise and report without anyone having decided that they should.

Sources

1. Kantor, J. (2014). Working Anything but 9 to 5: Scheduling Technology Leaves Low-Income Parents With Hours of Chaos. The New York Times; reprinted by Portside. (199, Kantor, 2014)
2. Schneider, D. & Harknett, K. (2019). Consequences of Routine Work-Schedule Instability for Worker Health and Well-Being. American Sociological Review, 84(1). (125, Schneider & Harknett, 2019)
3. Harknett, K., Schneider, D. & Irwin, V. (2021). Improving health and economic security by reducing work schedule uncertainty. Proceedings of the National Academy of Sciences, 118(42). (220, Harknett et al., 2021)
4. Vedaa, Ø., Harris, A., Bjorvatn, B., Waage, S., Sivertsen, B., Tucker, P. & Pallesen, S. (2016). Systematic review of the relationship between quick returns in rotating shift work and health-related outcomes. Ergonomics, 59(1). (222, Vedaa et al., 2016)
5. Kesavan, S., Lambert, S. J., Williams, J. C. & Pendem, P. K. (2022). Doing Well by Doing Good: Improving Retail Store Performance with Responsible Scheduling Practices at the Gap, Inc. Management Science, 68(11). (221, Kesavan et al., 2022)
6. European Commission (2026). Draft Commission guidelines on the classification of high-risk AI systems. Shaping Europe's digital future. (3, European Commission, 2026)
7. Dubal, V. (2023). On Algorithmic Wage Discrimination. Columbia Law Review, 123(7). (144, Dubal, 2023)
8. Cook, C., Diamond, R., Hall, J. V., List, J. A. & Oyer, P. (2021). The Gender Earnings Gap in the Gig Economy: Evidence from over a Million Rideshare Drivers. The Review of Economic Studies, 88(5). (139, Cook et al., 2021)
9. European Parliament and Council of the European Union (2024). Directive (EU) 2024/2831 on improving working conditions in platform work. Official Journal of the European Union. (47, European Union, 2024)
10. European Parliament and Council of the European Union (2003). Directive 2003/88/EC concerning certain aspects of the organisation of working time. Official Journal of the European Union. (50, European Union, 2003)
11. Court of Justice of the European Union (2019). Judgment of 14 May 2019, CCOO v Deutsche Bank SAE, Case C-55/18, ECLI:EU:C:2019:402. EUR-Lex. (66, CJEU, 2019)
12. European Union (2012). Charter of Fundamental Rights of the European Union. Official Journal of the European Union. (48, European Union, 2012)
13. Council of the European Communities (1989). Council Directive 89/391/EEC on the introduction of measures to encourage improvements in the safety and health of workers at work. Official Journal of the European Communities. (49, Council of the European Communities, 1989)
14. Sveriges riksdag (1982). Arbetstidslag (1982:673) (Working Hours Act). (219, Sveriges riksdag, 1982)
15. European Parliament and Council of the European Union (2019). Directive (EU) 2019/1152 on transparent and predictable working conditions in the European Union. Official Journal of the European Union. (53, European Union, 2019)
16. European Parliament and Council of the European Union (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union. (1, European Union, 2024)
17. European Parliament and Council of the European Union (2026). Regulation (EU) 2026/1744 (Digital Omnibus on AI). Official Journal of the European Union. (2, European Union, 2026)
18. European Parliament and Council of the European Union (2023). Directive (EU) 2023/970 on pay transparency and enforcement mechanisms. Official Journal of the European Union. (57, European Union, 2023)
19. Nordqvist, G. (2026). AI Red Team: AI as an Active Part of Cyber Defence in European Enterprises and Public Authorities. The Architecture as Code Series, Book 4. (43, Nordqvist, 2026)
20. Nordqvist, G. (2026). AI Governance: Governing Artificial Intelligence in European Enterprises and Public Authorities. The Architecture as Code Series, Book 2. (41, Nordqvist, 2026)
21. Nordqvist, G. (2026). AI Threat: Are the Dystopias Coming True? Evidence, Risk and Response for European Enterprises and Public Authorities. The Architecture as Code Series, Book 5. (44, Nordqvist, 2026)
22. Federal Republic of Germany (2001). Betriebsverfassungsgesetz (BetrVG). Gesetze im Internet. (9, Germany, 2001)
23. Sveriges riksdag (1976). Lag (1976:580) om medbestämmande i arbetslivet (MBL). (205, Sveriges riksdag, 1976)