Contestability by Design: Explanation, Appeal and Correction

A worker who believes that a score, a route or a pay input about her is wrong needs three things: to learn what the system did and who decided, to reach a person who can change the outcome, and to see the error corrected in her own case and in the system that produced it. This chapter designs those three things into the AI boss. It concerns M2 Evaluation, where scores and ratings are contested, and M3 Consequence, where the pay, promotion, discipline and access to work that follow from them are contested; it also reaches M1 Direction and M5 Care where a contested instruction, such as a route with no time between two visits, turns out to be a risk to health. The chapter introduces the contestation record (CNT-*), defines the explanation a person receives and the appeal path with its time limits, and shows how contestations feed changes to rules, data and the delegation register. Its argument is that an appeal reaching a person able to change the outcome is the condition on which every degree of delegation above Level 1 Assisted depends.

Figure 27.1: The contestation path: explanation, appeal to a person other than the decider, a written answer within a time limit, correction of the case and of the system, and contest patterns reported to the representatives.

Why contestability

Contestability has three justifications, and a path designed for all three is stronger than one designed for any one of them.

The first is procedural justice. People judge a decision by how it was made as well as by its result. Lee's experiment found algorithmic decisions on tasks requiring human skills perceived as less fair and less trustworthy than the same decisions attributed to a manager (165, Lee, 2018); Binns and colleagues found that people apply the same dimensions of justice to algorithmic decisions as to human ones, and that no single style of explanation was best across scenarios (164, Binns et al., 2018). Chapter 12 noted their limits: online experiments, not workplaces observed over time. Both show that part of what people judge is whether a person will answer.

The second is accountability. Bovens describes accountability as a relationship in which an actor must explain and justify its conduct to a forum that can question it and pass judgement, with consequences for the actor (121, Bovens, 2007). A contest is that relationship at the scale of one person and one decision. A system cannot take the actor's place, because it cannot be questioned in that sense or bear a consequence, which is why the book's rule that every M3 Consequence decision has a named human decider is also the precondition of contest.

The third is information. Hirschman showed that voice, the attempt to change an organisation from within, carries information about decline that exit does not (102, Hirschman, 1970). The person affected is often the only one who can see that a rule misread her work: the manager sees a figure and the owner sees an average. Contestations are the earliest error signal an organisation has, if they reach the people who can act on them.

Lyons, Velloso and Miller analysed 65 submissions on contestability to Australia's proposed AI ethics framework. They found no consensus on its nature, but broad agreement that it protects individuals and that contesting an algorithmic decision should resemble contesting a human one, with an emphasis on human review (175, Lyons et al., 2021). Their account of Australia's automated debt-recovery scheme adds a warning: individuals who sought review had debts overturned, yet the system stayed in place until a class action. They conclude that contestation is the last line of defence and cite Almada's case for contestability by design, built in with those affected from the start. AI Governance, Chapter 10, concluded that oversight must be designed rather than declared, with an interruption path, an override log and the authority to use them (F21) (41, Nordqvist, 2026). Contestability is the same conclusion from the other side: oversight acts before a decision takes effect, contest after.

What a worker can learn

Chapter 12 set out the law on reasons and review; this section takes from it what an appeal path must deliver. European law gives a worker four overlapping routes to information about a management decision that AI made or shaped, and an employer that designs one explanation to satisfy all of them avoids running four processes.

GDPR Article 15(1)(h) gives the right to meaningful information about the logic involved in automated decision-making referred to in Article 22(1) and (4), and its significance and envisaged consequences; Article 12(3) requires an answer within one month, extendable by two further months (5, European Union, 2016). Wachter, Mittelstadt and Floridi argued that this amounts only to a right to be informed about a system's logic in general, not to an explanation of a specific decision (160, Wachter et al., 2017). The Court of Justice has since read the provision more broadly. In Dun & Bradstreet it held that the person is entitled to an explanation of the procedure and principles actually applied to their data, in a concise and intelligible form that neither a complex formula nor a description of every step provides; that for profiling it may be sufficient to say to what extent a variation in the data would have led to a different result; and that information claimed as a trade secret goes to the supervisory authority or court to balance (65, CJEU, 2025). Article 22(3) adds, where a solely automated decision is permitted by contract or explicit consent, the right to human intervention, to express a point of view and to contest. The Gerechtshof Amsterdam found it hard to see how a meaningful examination of Uber's deactivations could take place without hearing the drivers (67, Gerechtshof Amsterdam, 2023): a contest is part of what makes review meaningful, not a step after it.

AI Act Article 86(1) gives a person subject to a deployer's decision based on the output of an Annex III high-risk system, which includes point 4 on employment, the right to clear and meaningful explanations of the role of the system in the procedure and the main elements of the decision, where the person considers its effects adverse to health, safety or fundamental rights; Article 86(3) applies the right only where Union law does not already provide it (1, European Union, 2024). It applies from 2 August 2026, while the deployer duties that make an explanation possible apply to Annex III systems from 2 December 2027 (2, European Union, 2026).

The Platform Work Directive binds only platforms, but its Article 11 is the most complete statement in Union law of a contest path, and Chapter 5 already used it as a benchmark. It requires an explanation in clear and plain language of any decision taken or supported by an automated system; a contact person with the competence, training and authority to discuss it; written reasons for decisions such as suspending an account; a written, substantiated reply to a request for review within two weeks; rectification of an infringing decision within two weeks, or compensation; and the steps necessary, including modifying or discontinuing the system, to avoid such decisions in future (47, European Union, 2024). It links the individual remedy to the correction of the system in a single article.

National law adds the routes that already exist for any management decision. BetrVG §82(2) entitles an employee in Germany to have the calculation of pay explained and the assessment of performance discussed, with a works council member present; §84 gives a right to complain to the competent body, obliges the employer to answer and to remedy a justified complaint, and forbids disadvantage for complaining; §85 lets the works council take up a complaint and, in a dispute over whether it is justified, call the conciliation committee unless it concerns a legal claim (9, Germany, 2001). In Sweden a union may negotiate on any question between the employer and a member (MBL §10), and a claim under the act or a collective agreement must be raised within four months of learning of it (MBL §64) (205, Sveriges riksdag, 1976).

A public employer carries one further duty. The Swedish Administrative Procedure Act requires an authority to act lawfully, objectively and proportionately and to help individuals protect their interests, principles that reach all its administrative activity (§§ 1, 5 and 6). A decision that may affect someone's situation more than insignificantly must state the rules applied and the decisive circumstances (§ 32), although § 2 disapplies that rule for regional matters open to legality review under the Local Government Act, and § 32 allows reasons to be omitted for decisions on appointing someone (239, Sveriges riksdag, 2017). Whether a unit manager's approval of a route is case handling at all need not be settled for a region to act. AI Governance, Chapter 17, concluded that in public authorities legality comes before capability and every case has a reasons log (F31) (41, Nordqvist, 2026); a public employer that gives reasons for contested management decisions as policy has that log, and the record an ombudsman or court will ask for.

Explanations people can use

An explanation that satisfies a lawyer can still be useless to an assistant nurse at the end of a shift. Edwards and Veale warned that a right to explanation is unlikely to be a complete remedy: the law is unclear about when it is triggered, the explanations machine-learning research produces may not be the meaningful information the law means, and the search for such a right risks nurturing what they call a "transparency fallacy" (161, Edwards & Veale, 2017). Selbst and Barocas distinguished models that are inscrutable, whose rules cannot readily be described, from models that are nonintuitive, whose described rules still do not make sense. Describing the rules addresses the first; only an account of why the rules are what they are addresses the second, so they argue for explaining the process behind a model's development as well as the model (246, Selbst & Barocas, 2018).

Most management systems in this book are less opaque than that. A roster optimiser applies constraints and an objective; a sales score weights three indicators. The problem is rarely that the rule cannot be stated, but that it is not stated to the person in terms of her case, and that nobody has asked whether it is right. The explanation therefore has two layers. Kaminski and Malgieri call for multi-layered explanations that join the individual's rights to the systemic governance of the data protection impact assessment, so that what the person is told rests on what the organisation assessed (174, Kaminski & Malgieri, 2021). For an AI boss, the systemic layer is the worker information notice of Chapter 28, the delegation profile and the fundamental rights impact assessment; the individual layer is the explanation of one decision, and it should cite the systemic layer rather than repeat it.

The individual explanation has four elements.

1. The role of the system and the logic it applied. Which system took part, at which degree of delegation, what it proposed and the rule it applied, in a sentence a person can repeat to a colleague: "the roster system drafts each day's route from your assigned visits, their time windows and the travel time between addresses; the unit manager approves it". That meets Article 86 and Dun & Bradstreet; the optimiser's objective function would meet neither.
2. The inputs about the person. The data about her that the system used, and only those: her contracted hours, competences and assigned visits for a route; her conversion, retention and complaint figures and their comparison base for a score. This is what lets her find the wrong input, and says where she can see the data themselves under the right of access.
3. The counterfactual. Wachter, Mittelstadt and Russell proposed counterfactual explanations, statements of the smallest change to the inputs that would have produced a different outcome,, serving three aims without opening the model: to understand a decision, to contest it and to know what would change it in future (162, Wachter et al., 2018); the Court's remark on a variation in the data points the same way. For a management decision the counterfactual should name an input the person or the organisation could actually change ("had your complaints been compared with the Dutch market rate, your band would have been B"), never a protected characteristic, and never a figure the person can reach only by working faster than the risk assessment allows. Where the decision rests on a rule rather than a threshold, the counterfactual states the rule that would have to change.
4. The human decider. Who decided, in what role, what that person considered besides the system's output and where they departed from it, and how to reach them. A decision with no named decider cannot be explained in this form, which is itself a finding.

Every explanation also says how to contest, by when, to whom, and that contesting carries no detriment. The decider explains it in person when asked, because the finding that no one style of explanation is best is an argument for a conversation, not a better template. An AI assistant may draft the explanation from the records; the decider reads, corrects and signs it, and the explanation says so.

The appeal path

An appeal path has five parts: receipt, a first look, independent review, a written answer and escalation. Each part has an owner and a time limit, and the whole path must work without a lawyer.

Receipt. A contest can be lodged with the decider, with HR, through a form or by a representative on the person's behalf, in any form; an oral contest is written down and read back. Receipt is acknowledged within two working days with the reviewer's name and the answer date. The window for lodging should never close before the national route for the same dispute; at Nordvik that means at least the four months of MBL §64.

The first look. The decider may grant the contest in full and correct the decision at once, and usually should: a route corrected the next morning is worth more than one corrected in two weeks. The decider may not reject a contest or grant it in part; those, and every contest not resolved within two working days, go to review. The decider can say yes; only the reviewer can say no.

Independent review. The reviewer has not taken, approved or advised on the decision, and is not the system that produced the output, its supplier or an agent acting under the system owner's mandate. An AI system may retrieve and summarise the evidence for the reviewer, at Level 0 Advisory; it never drafts the finding. The reviewer has the authority to change the decision, to require correction of the data and to refer the rule to the system's owner with a deadline for an answer. Lyons and colleagues distinguish internal review by the decision-making body from external review by an ombudsman, tribunal or court (175, Lyons et al., 2021). An employer's path is internal review, so independence is designed in by role: an HR business partner, a manager from another unit, never the decider's own line where it approved the decision.

The answer. The answer is written, states the rules applied and the decisive circumstances, what will change in the case and in the system, and the escalation route and deadline. Time limits are recorded in the appeal path, not left to each manager. Two weeks, the Platform Work Directive's limit for a reply, is a sound default for decisions that take effect quickly, such as routes, rosters and access to shifts; three weeks can be justified for a pay input in a group with several national arrangements, provided the pay effect is held or applied provisionally until the answer.

Escalation. A person not satisfied can go to a second, more senior review, to the representatives' own routes (negotiation, BetrVG §85, a safety representative's request for investigation), to the data protection officer and supervisory authority, or to a court. The path replaces none of them and says so.

Practical guide: Building an appeal path for a management system

Purpose. Give every worker affected by a management system a route from contest to correction that names who answers, by when and with what authority. Roles: the system owner (accountable), HR (operates the path), line managers, the workers' representatives, the data protection officer and workers who will use the path.

1. Map the decisions that can be contested. From the delegation profile, list every decision type in M1 Direction, M2 Evaluation and M3 Consequence that the system makes or shapes about a person, with its decider, not only consequence decisions.
2. Write the explanation template with the four elements and test it with three workers. Default: one page, linking to the worker information notice.
3. Name the receivers and the reviewers by role for each decision type, with a deputy for each. No reviewer may have approved the decisions they review. Default: HR business partner for the sector; another unit's manager as deputy.
4. Set the time limits in working days and in calendar days and state which governs. Default: acknowledgement within 2 working days; explanation with the acknowledgement or within 5 working days; answer within 10 working days and never later than 14 calendar days for decisions that take effect within a month; up to 15 working days for pay inputs, with the pay effect held or provisional.
5. Write the first-look rule: the decider may grant in full, never reject.
6. Define what correction means for the decision type: the decision, the data, the rule, the model and the register, each with an owner and a test.
7. Connect the record. Create a contestation record for every contest, linked to the decision's record and the delegation profile. Default: retained for the life of the system plus two years; aggregates for representatives.
8. Agree reporting and triggers with the representatives (see "Contest patterns").
9. Test the path. With the representatives' agreement, lodge two test contests per year through each channel and measure time to acknowledgement and answer.

Definition of done. The path is recorded in a decision record referenced from every delegation profile it covers, the worker information notice names it, the first quarterly report has gone to the representatives and a test contest was answered within the limit. Adapt the defaults to the collective agreement or works agreement, which may set shorter limits or give representatives a role in review.

At Nordvik Region the first contest exposed two gaps: it went to the unit manager who had approved the route, and two time limits were in circulation, 14 days in the dashboard record AIDR-0624 (Chapter 3) and 10 working days in FRIA-2026-007 (Chapter 21). Ingrid Halvorsen recorded the Region's rule in AIDR-0631.

# decisions/AIDR-0631.yaml (illustrative)
id: AIDR-0631
title: Appeal path for contested management decisions
date: 2027-03-15
status: accepted
commission: AIDR-0600
owner: Ingrid Halvorsen, HR director
reviewers: [AI governance lead, data protection officer, chair of the central work-environment committee]
context:
  - CNT-NR-0001 was lodged with the unit manager who had approved the route
  - AIDR-0624 (dashboards) said 14 days; FRIA-2026-007 said 10 working days
applies_to: [DLG-NR-ROS-01, DLG-NR-HCD-01, every future DLG-NR-* profile]
decision:
  lodge:
    channels: [unit manager, HR business partner for the sector, staff app form, union or safety representative]
    window: four months from the day the worker learnt of the decision   # never shorter than MBL 64
  acknowledge: 2 working days, naming the reviewer and the answer date
  explanation: [role and logic, inputs about the person, counterfactual, human decider, how to contest]
  first_look: the decider may grant in full and correct at once; may not reject or grant in part
  review:
    by: HR business partner for the sector
    deputy: unit manager of another home-care or ward unit
    excluded: [the decider, anyone who approved or advised on the decision, ShiftWise, its supplier]
    authority: [change the decision, order data correction, refer the rule to the system owner]
  answer:
    form: written, with reasons naming the rules applied and the decisive circumstances
    limit: 10 working days from receipt and never later than 14 calendar days; the earlier governs
  escalation: [union negotiation under the co-determination act, safety representative request, data protection officer]
  no_detriment: true
reporting:
  to: [central work-environment committee, unions party to DLG-NR-ROS-01]
  frequency: quarterly
  content: contests by system, decision type and outcome; time to answer; system changes
  minimum_cell: 5
consequences:
  - The 14 days in AIDR-0624 and the 10 working days in FRIA-2026-007 now read as this single rule
  - FRIA-2026-007 and DLG-NR-HCD-01 reference AIDR-0631 at their next revision

The rule reconciles the two figures rather than choosing one. In an ordinary fortnight 10 working days and 14 calendar days are the same period; when a public holiday falls inside it, the calendar limit governs, keeping the Region within the two weeks the Platform Work Directive sets, which it has adopted as its benchmark.

The contestation record

The contestation record sits beside the Management Decision Record, the delegation profile and the oversight record, and follows the discipline that AI Platform and Security, Chapter 5, set for AI Decision Records: short, numbered, immutable once closed, validated against a schema and re-opened only by new evidence (F38) (42, Nordqvist, 2026). Nordvik numbers its records CNT-NR-0001 onwards and Hansa CNT-HM-0001 onwards. A record holds, at minimum:

• Identity and links: the identifier, the contested decision and its record (the MDR where one exists), the system and its version, the delegation profile, the management function and the degree of delegation at the time.
• The contest: who contested, who supported them, when, how, and the grounds in the person's own words.
• Deadlines: acknowledgement, explanation and answer dates, each as due and as met.
• The explanation given, with its four elements.
• The deciders: the original decider, the first-look outcome, the reviewer and the statement that the reviewer is independent of the decision and the system.
• Evidence and findings, including findings about oversight.
• The outcome at five layers: the decision, the data, the rule, the model and the register, each with an owner, a test where one applies and a date verified.
• Closure: when the person was informed, the escalation route and deadline, and the closing date, when the last system correction passed its test.

The record is personal data about the worker and often the decider. The worker sees her record in full, the decider and reviewer the case, the system owner the findings and corrections, and representatives aggregates. The record never feeds an M2 Evaluation of the worker who contested, and the delegation profile says so, which is how BetrVG §84(3)'s prohibition of disadvantage becomes a control rather than an assurance.

Correcting the decision and the system

A successful contest corrects up to five things, and the record names each.

• The decision. The route, the score, the pay input or the shift offer is changed, and any loss is made good. Where the decision cannot be undone, as with an evening already worked, the correction is compensation or an adjustment the person agrees.
• The data. A wrong input is corrected at its source, not only in the case, and every other decision that used it is checked. A misattributed complaint corrected in one employee's score but left in the database will reappear next quarter.
• The rule. A constraint, threshold, weight or comparison base that produced the error is changed by its owner, after the risk assessment and representatives' review the change requires, and written as a test that fails if the error returns.
• The model. Where a learned component produced the error, it is retrained, recalibrated or withdrawn, and the change passes the same release gates as any other, including the disparate-impact test of Chapter 12.
• The register. The delegation profile is updated: a new condition, a changed degree of delegation, a practice classification, a new review trigger.

AI Red Team, Chapter 3, concluded that every finding ends as a verified control, closed only when the re-test passes (F60) (43, Nordqvist, 2026). A contestation is a finding reported by the person best placed to find it, and the same discipline applies: the record stays open until the system correction has passed its test, not when the person has been answered. This is the step Lyons and colleagues found missing in the Australian debt scheme, where individual reviews succeeded and the system stayed in place (175, Lyons et al., 2021), and the step the Platform Work Directive's Article 11(3) makes explicit for platforms (47, European Union, 2024).

Worked example: CNT-NR-0001, a route with no time between two visits

Illustrative. On the second day of ShiftWise routes in the Strandvik home-care unit, an assistant nurse told Mikael Strand that her route for that day planned no time to travel between two visits at addresses about a kilometre apart, and that she had made up the time by shortening the second visit. The safety representative supported the contest. Strand corrected the route for the following day at once, as the first-look rule allows, but could not see why the route had been drafted that way, so the contest went to review.

# contestations/CNT-NR-0001.yaml (illustrative)
id: CNT-NR-0001
opened: 2027-01-26
channel: oral, to the unit manager; written down and read back
contested_record: ShiftWise daily route, Strandvik unit, 2027-01-26, approved by the unit manager
system: ShiftWise (tender NR-2025-089)
delegation_profile: DLG-NR-ROS-01
function_and_level: {M1 Direction: Level 2 Supervised}
contested_by: assistant nurse, Strandvik home-care unit
supported_by: safety representative
grounds: Route planned 0 minutes between two visits about 1 km apart; second visit shortened
explanation_given:
  date: 2027-01-27
  role_and_logic: ShiftWise orders assigned visits by time window and travel time; the unit manager approves
  inputs_about_person: [shift times, competences, assigned visits]   # no performance data
  counterfactual: With travel time computed between the two addresses, 12 minutes would have been planned
  human_decider: Mikael Strand, unit manager
first_look: granted in full; route for 2027-01-27 corrected by the unit manager
reviewer: HR business partner for home care   # independent of the route and of ShiftWise
deadlines: {acknowledged: 2027-01-26, answer_due: 2027-02-09, answered: 2027-02-05}
findings:
  - Both addresses geocoded to the same postcode centroid, so routed travel time was 0 minutes
  - No transition time for parking, entering and leaving was planned on any route
  - The rest-rule flag did not fire because no working-time rule was breached
outcome:
  decision: route corrected 2027-01-27; shortened visit recorded as caused by the route, not the worker
  data: address-level coordinates for all Strandvik residents, supplier ticket closed 2027-02-12
  rule: travel time at least routed time plus 5 minutes, minimum 10 minutes (WRA-NR-HC-01 rev 2, C2)
  model: none; the visit-duration forecast was examined and not implicated
  register: DLG-NR-ROS-01 references constraint C2 and AIDR-0631
system_correction:
  owner: Ingrid Halvorsen, HR director
  reviewed_by: central work-environment committee
  test: tests/test_travel_time.py
  verified: 2027-03-25
person_informed: 2027-02-05, in writing and at a meeting with the unit manager
escalation: union negotiation or safety representative request; open until 2027-05-26
closed: 2027-03-25

The corrected route, and a visit record that no longer counts against her, were the individual remedy. The coordinates were the data. The constraint, written as a test, was the rule, and it came through the risk assessment WRA-NR-HC-01 that Chapter 14 showed, approved by the central work-environment committee, not through a setting changed by the supplier. The register records the constraint and the appeal path. What the case also showed was that the route complied with every working-time rule: a flag designed to catch breaches of rest could not catch a plan that was lawful and unworkable, and only the person doing the work could. Chapter 6 traced how the same contest led to a risk assessment of the planned visit times themselves.

Worked example: CNT-HM-0001, a contested quarterly score

Illustrative. Chapter 8 introduced the contest of a field-sales employee of Hansa Mutual's Dutch entity against her score for the third quarter of 2026. The full record below adds the explanation, the deadlines and the correction at each layer. The contest preceded Hansa's Management Decision Records, which begin with MDR-HM-0001 in January 2027, so the contested pay proposal is referenced by its pay-cycle entry.

# contestations/CNT-HM-0001.yaml (illustrative, full record)
id: CNT-HM-0001
opened: 2026-10-19
channel: written, through the HR portal
contested_record: sales-performance-score Q3 2026 and the pay proposal based on it (pay cycle Q3 2026)
system: sales-performance-score, Mode 4 Built AI applications
delegation_profile: DLG-HM-SPS-01 version 2
function_and_level: {M2 Evaluation: Level 1 Assisted, M3 Consequence: Level 0 Advisory}
contested_by: field-sales employee, Dutch entity
supported_by: ondernemingsraad member
grounds:
  - Complaint indicator compared with the group median, not the Dutch market
  - Market-wide rise in complaints after a motor-policy product change
explanation_given:
  date: 2026-10-23
  role_and_logic: Score weights conversion, retention and complaints against a comparison base; manager proposes pay
  inputs_about_person: [conversion, retention and complaint figures, Q3 2026; portfolio assignments]
  counterfactual: Against the Dutch market rate, complaint band 3 instead of 1 and composite band B instead of C
  human_decider: the employee's sales manager
first_look: not granted by the sales manager; referred to review
reviewer: HR business partner sales, outside the employee's line
deadlines: {acknowledged: 2026-10-20, answer_due: 2026-11-09, answered: 2026-11-02}
evidence_examined:
  - Indicator definitions in DLG-HM-SPS-01 version 2
  - Complaint rates by market, Q2 and Q3 2026
  - Complaint attribution after the portfolio transfer of 2026-07-01
  - The pay proposal and the manager's recorded reasons
findings:
  - Complaint indicator penalised staff in markets with higher complaint rates
  - Two complaints about advice given before the portfolio transfer were attributed to her
  - Pay proposal matched the score band without a written assessment
  - In the Dutch sales unit 97 per cent of proposals matched the band; median 48 seconds to decide
outcome:
  decision: band corrected; pay proposal re-made by the manager with own assessment first
  data: complaints about events before a portfolio transfer excluded from the new holder's indicator
  rule: complaint indicator normalised per market from Q4 2026
  model: none; composite weights unchanged
  register: DLG-HM-SPS-01 version 3; Dutch sales unit classified decisive in practice
person_informed: 2026-11-03, in writing, with reasons
escalation: second review by an HR lead of another entity; open until 2026-12-03
system_correction:
  owner: group head of sales HR
  test: Q4 2026 scores recomputed against market baselines and attribution rule
  verified: 2027-01-15
closed: 2027-01-15

The sales manager declined to grant the contest at the first look, which under Hansa's path sent it to review rather than ending it. The reviewer found an error in the comparison base, an error in the data and an oversight failure. The last of these mattered most: Chapter 8 showed that the profile's Level 0 Advisory for M3 Consequence was, in the Dutch unit, a pay input that followed the score in almost every case within a minute. The contest exposed it; the practice test confirmed it across the unit; and version 3 of the profile introduced the rule that the manager records an assessment before opening the score.

Contest patterns

One contest corrects a case and, if the path works, a rule. A pattern of contests shows whether the system and its oversight work at all, and representatives rarely receive it. Thomas Brandt's request after CNT-HM-0001, for a quarterly report of contests by system and outcome, is the general form.

A useful report shows, per system and decision type: the number of contests per hundred people affected; the share granted at first look, granted on review, partly granted and rejected; the median time to answer and the share answered within the limit; the system corrections made and verified; and contests in which the review found an oversight failure. Cells under five people are suppressed, and the figures become triggers in the delegation profile.

Three patterns call for action. A rising rate of contests granted on review means that the decisions or the system are wrong more often than oversight catches, and the degree of delegation should be reviewed. A cluster of contests about one rule or one group of people, such as part-time staff or one national market, calls for the disparate-impact test of Chapter 12 to run outside its schedule. And no contests at all, from a system that makes thousands of decisions about people, is not evidence that the system is right. It may mean that people do not know the system exists, do not believe a contest will change anything, or fear the consequences; the representatives can say which.

Hansa recorded the report and its triggers in AIDR-0681 and in version 4 of the sales-score profile.

# management/delegation-register.yaml (excerpt, illustrative)
- id: DLG-HM-SPS-01
  version: 4
  decision_record: AIDR-0681
  effective: 2027-01-25
  contestation:
    path: CNT-HM records; acknowledgement 2 working days; answer within 15 working days
    first_look: the sales manager may grant in full, never reject
    reviewer: HR business partner sales outside the employee's line
    pay_effect_while_contested: provisional; corrected in the next salary run
    escalation: second review by an HR lead of another national entity within one month
    no_use_in_evaluation: a contest is never an input to any M2 Evaluation of the person
  contest_report:
    to: [Konzernbetriebsrat, ondernemingsraad, European Works Council summary]
    frequency: quarterly, first report 2027-04-30
    by: [system, national entity, decision type, outcome]
    minimum_cell: 5
  triggers:
    - when: contests granted on review exceed 3 per 100 employees scored in a quarter
      action: practice test and review of the degree of delegation within 30 working days
    - when: two or more granted contests in a quarter concern the same indicator
      action: indicator owner reports cause and fix to the Konzernbetriebsrat
    - when: no contests in two consecutive quarters in an entity with more than 50 people scored
      action: ask the local representatives whether the path is known and trusted
    - when: a review finds a manager followed the score without a recorded assessment
      action: practice test for that unit, as in Chapter 8

Contestability as a condition

The degrees of delegation of Chapter 4 can now be read from the side of the person affected. At Level 0 Advisory and Level 1 Assisted, a person decides each case with the system's output in view, and a contest reaches the decider's reasons. At Level 2 Supervised, the system's proposal takes effect unless a person objects; the default is the system's, so a contest is often the first moment a person examines the case. At Level 3 Bounded, the system acts within limits and people handle exceptions; most individual decisions have no human decider at all, and the contest is the only per-case check that reaches a decision no one examined.

The higher the degree, the more of the work of catching individual errors falls to the person affected. Oversight of the system, which Chapter 26 designs, samples and measures; contest sees the cases the person knows to be wrong. Every degree above Level 1 Assisted therefore rests on an appeal that reaches a person able to change the outcome, within a time that still matters, with a correction that reaches the rule. Where that path is absent, slow or answered by the system that produced the decision, the degree is not justified, whatever the profile says, and should return to Level 1 until the path works. Nordvik's open-shift offers and Hansa's claims allocation, both at Level 3 Bounded, depend on it.

AI Governance, Chapter 13, concluded that a stop path running inside the system it stops fails when it is needed (F19) (41, Nordqvist, 2026). A contest path that runs only inside the system it contests fails in the same way: the form, the record and the reviewer's authority must work when the system is down, wrong or withdrawn.

Across the degrees of delegation

Degree of delegation	What a person contests	Who answers	Condition for the degree
Level 0 Advisory	A manager's decision informed by a score or proposal	The reviewer, with the manager's recorded reasons	Explanation names what the manager considered besides the output
Level 1 Assisted	A decision the manager took on the system's proposal	The reviewer, never the proposing system	Departures and reasons recorded; practice test where it feeds M3 Consequence
Level 2 Supervised	A proposal that took effect because no one objected	The decider at first look, then an independent reviewer	Answer within two weeks; contests granted on review reported
Level 3 Bounded	An action the system took within limits	A reviewer with authority over the rule as well as the case	Contest patterns trigger review of the bounds; first correction at the rule
Level 4 Autonomous	Not used for M3 Consequence	Not applicable	Never for a consequence decision about a person

The usage mode changes how hard contest is to design. In Mode 0 Ambient AI the person may not know a feature produced a figure about her, so the transparency register of Chapter 28 must list it before anyone can contest it. Mode 2 Embedded workplace assistants that summarise a person's work for a manager need an explanation that names the summary as an input. Mode 4 Built AI applications give the organisation the data to explain and correct; procured ones require the contract to provide them, as Nordvik learnt with the geocoding. Mode 5 Autonomous production agents that allocate work need records that show which agent and which version produced the allocation, or the explanation cannot name the role of the system.

Outlook

Rights to explanation and review are observable in force: GDPR Articles 15 and 22 as the Court has read them, AI Act Article 86 since 2 August 2026, and the national complaint and negotiation routes. Effective appeal paths for algorithmic management, with independent review, written answers within a limit and correction of the system, are observable in few organisations; the cases of Chapters 5, 8 and 12 record their absence. National transposition of the Platform Work Directive, due by 2 December 2026, and the application of the deployer duties for Annex III systems from 2 December 2027 make more complete contest paths expected, at least for platforms and high-risk systems. That contest data will become a routine governance signal, reported to representatives each quarter and wired to triggers in delegation profiles, is expected where works councils and unions ask for it and speculation elsewhere.

Nordvik Region and Hansa Mutual

At Nordvik Region, CNT-NR-0001 was the first contest under DLG-NR-ROS-01, lodged on the second day of routes in the Strandvik unit. Mikael Strand corrected the route the next morning; the review found a data error and a missing rule, which became constraint C2 in WRA-NR-HC-01, approved by the central work-environment committee; and the contest showed that the Region's path sent contests to the person who had approved the decision and carried two time limits. AIDR-0631 settled both. As a public employer, the Region gives written reasons for contested management decisions as policy, without waiting to settle whether the Administrative Procedure Act binds it in each case.

At Hansa Mutual, CNT-HM-0001 came from a Dutch employee supported by the ondernemingsraad, so the group path had to work beside Dutch works-council rights and the German routes of BetrVG §§84 and 85. The review found a wrong comparison base, misattributed complaints and a unit where the advisory score was decisive in practice. Thomas Brandt's request for a quarterly contest report became, through AIDR-0681, a set of triggers in version 4 of DLG-HM-SPS-01, with the first report due on 2027-04-30, the date of the group's first report under AIDR-0650. The organisations differ in where contest leads: at Nordvik into the work-environment system and the unions' negotiation rights under one regime; at Hansa into pay and several national arrangements, with a 15-working-day limit that holds only because the pay effect stays provisional until the answer.

Summary

This chapter designed contestability into the AI boss for M2 Evaluation and M3 Consequence, and for the M1 Direction and M5 Care decisions, such as routes, whose errors reach a person's health. Contest serves procedural justice, gives accountability a person who must explain and justify, and is the earliest error signal an organisation has; the research agrees it should resemble contest of human decisions, with human review, and warns that it is the last line of defence, not a substitute for designing the system well.

The law gives a worker routes to an explanation through GDPR Articles 15(1)(h) and 22(3) as Dun & Bradstreet and the Amsterdam court read them, AI Act Article 86, the Platform Work Directive's Article 11 as a benchmark, and national routes such as BetrVG §§82, 84 and 85, MBL §§10 and 64 and a public employer's duty of reasons. The explanation has four elements: the role of the system and its logic, the inputs about the person, an actionable counterfactual and the human decider. The appeal path has five parts: receipt, a first look in which the decider may only grant, independent review by someone who is neither the decider nor the system, a written answer within a stated limit, and escalation.

The contestation record CNT-* holds the contest, explanation, deadlines, the reviewer's independence, findings and the correction of the decision, the data, the rule, the model and the register, and closes only when the system correction passes its test. CNT-NR-0001 led to a minimum travel-time constraint and to AIDR-0631, which reconciled Nordvik's time limits; CNT-HM-0001 led to per-market normalisation, a rule on complaint attribution and AIDR-0681, which made contest patterns a quarterly report and a trigger in the delegation profile. Every degree of delegation above Level 1 Assisted depends on such a path, because the higher the degree, the more of the work of catching errors falls to the person affected.

A contest presupposes that the person knows the system exists, what it does and whom to ask. Chapter 28 turns to that precondition: what workers are told about the AI boss, when and by whom.

Sources

1. Lee, M. K. (2018). Understanding perception of algorithmic decisions: Fairness, trust, and emotion in response to algorithmic management. Big Data & Society, 5(1). (165, Lee, 2018)
2. Binns, R., Van Kleek, M., Veale, M., Lyngs, U., Zhao, J. & Shadbolt, N. (2018). 'It's Reducing a Human Being to a Percentage': Perceptions of Justice in Algorithmic Decisions. Proceedings of CHI 2018. (164, Binns et al., 2018)
3. Bovens, M. (2007). Analysing and Assessing Accountability: A Conceptual Framework. European Law Journal, 13(4). (121, Bovens, 2007)
4. Hirschman, A. O. (1970). Exit, Voice, and Loyalty: Responses to Decline in Firms, Organizations, and States. Harvard University Press. (102, Hirschman, 1970)
5. Lyons, H., Velloso, E. & Miller, T. (2021). Conceptualising Contestability: Perspectives on Contesting Algorithmic Decisions. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1). (175, Lyons et al., 2021)
6. Nordqvist, G. (2026). AI Governance: Governing Artificial Intelligence in European Enterprises and Public Authorities. The Architecture as Code Series, Book 2. (41, Nordqvist, 2026)
7. European Parliament and Council of the European Union (2016). Regulation (EU) 2016/679 (General Data Protection Regulation). Official Journal of the European Union. (5, European Union, 2016)
8. Wachter, S., Mittelstadt, B. & Floridi, L. (2017). Why a Right to Explanation of Automated Decision-Making Does Not Exist in the General Data Protection Regulation. International Data Privacy Law, 7(2). (160, Wachter et al., 2017)
9. Court of Justice of the European Union (2025). Judgment of 27 February 2025, CK v Dun & Bradstreet Austria GmbH and Magistrat der Stadt Wien, Case C-203/22, ECLI:EU:C:2025:117. EUR-Lex. (65, CJEU, 2025)
10. Gerechtshof Amsterdam (2023). Judgment of 4 April 2023, ECLI:NL:GHAMS:2023:793 (drivers and App Drivers & Couriers Union v Uber). Rechtspraak.nl. (67, Gerechtshof Amsterdam, 2023)
11. 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)
12. 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)
13. 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)
14. Federal Republic of Germany (2001). Betriebsverfassungsgesetz (BetrVG). Gesetze im Internet. (9, Germany, 2001)
15. Sveriges riksdag (1976). Lag (1976:580) om medbestämmande i arbetslivet (MBL). (205, Sveriges riksdag, 1976)
16. Sveriges riksdag (2017). Förvaltningslag (2017:900). (239, Sveriges riksdag, 2017)
17. Edwards, L. & Veale, M. (2017). Slave to the Algorithm? Why a 'Right to an Explanation' Is Probably Not the Remedy You Are Looking For. Duke Law & Technology Review, 16. (161, Edwards & Veale, 2017)
18. Selbst, A. D. & Barocas, S. (2018). The Intuitive Appeal of Explainable Machines. Fordham Law Review, 87(3). (246, Selbst & Barocas, 2018)
19. Kaminski, M. E. & Malgieri, G. (2021). Algorithmic impact assessments under the GDPR: producing multi-layered explanations. International Data Privacy Law, 11(2). (174, Kaminski & Malgieri, 2021)
20. Wachter, S., Mittelstadt, B. & Russell, C. (2018). Counterfactual Explanations Without Opening the Black Box: Automated Decisions and the GDPR. Harvard Journal of Law & Technology, 31(2). (162, Wachter et al., 2018)
21. Nordqvist, G. (2026). AI Platform and Security: Building and Securing AI Platforms in European Enterprises and Public Authorities. The Architecture as Code Series, Book 3. (42, Nordqvist, 2026)
22. 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)