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

Long before any algorithm assigned a shift, works councils in Germany, unions in the Nordic countries and the Netherlands' ondernemingsraden held rights to be informed, consulted, negotiated with and, in some matters, to withhold consent. Those rights now meet systems that direct and measure people's work and change with every release. This chapter sets out the national and European rights that apply to an AI boss and answers the open question Q8: how co-determination and collective agreements apply to a system whose models, rules and thresholds change far more often than the negotiating bodies meet. It concerns four management functions: M1 Direction, because allocation and rostering rules are where most changes happen; M2 Evaluation, because the right to co-determine attaches to systems suited to monitor behaviour and performance; M3 Consequence, because selection guidelines and the use of figures in pay or discipline are the matters representatives guard most closely; and M5 Care, because co-determination is the institutional form of voice at work.

Figure 23.1: An agreement that governs change rather than versions: every release is classified against the delegation profile, and each class carries the representatives' rights in Germany and Sweden.

Co-determination as a European constant

Europe has no single model of employee participation, but it has a common grammar with four degrees. Information means that the employer transmits data so that representatives can examine a subject. Consultation means an exchange of views and a dialogue before the employer decides. Negotiation means that the employer must seek agreement, though it may decide when none is reached. Co-determination or consent means that the employer may not act without the representatives' agreement or a substitute decision by a conciliation body or court. The degree differs by country and subject.

The national rights examined in this chapter can be summarised as follows. The table states what the law grants, not what any agreement adds.

Jurisdiction	Body	Strongest right over an AI boss	Trigger in the text
Germany	Betriebsrat, Gesamtbetriebsrat, Konzernbetriebsrat	Co-determination (BetrVG §87(1) no. 6); consent to selection guidelines (§95)	Introduction and application of technical equipment suited to monitor
Sweden	Unions bound by collective agreement	Duty to negotiate before deciding (MBL §11)	Important change in activity or conditions of work
Norway	Union representatives (tillitsvalgte)	Duty to discuss as early as possible (Working Environment Act §9-2)	Need, design, implementation and significant change of a control measure
Denmark (municipalities)	Employees, via collective agreement	Information 6 weeks before new control measures (KL agreement)	New control measures, including AI-based digital work tools
Netherlands	Ondernemingsraad	Consent (WOR Article 27(1)(k) and (l))	Regulations on personal data; facilities suited to observe or check conduct or performance
Spain	Works council	Information on parameters, rules and instructions (Workers' Statute Article 64.4(d))	Algorithms affecting conditions of work, access to and retention of employment
Italy	Company union representatives	Information, and written notice of changes at least 24 hours ahead (Legislative Decree 104/2022, Article 1-bis)	Fully automated decision or monitoring systems

The triggers are functional, not technical: a rules engine that allocates claims is caught as surely as a large language model. And most attach to change as well as to introduction, which makes Q8 a practical question.

Germany: why almost every AI boss is technical equipment suited to monitor

The Works Constitution Act gives the works council a right of co-determination, so far as no statute or collective agreement regulates the matter, over "the introduction and application of technical equipment intended to monitor the behaviour or performance of employees" (BetrVG §87(1) no. 6) (9, Germany, 2001). If employer and works council cannot agree, the conciliation committee (Einigungsstelle) decides, and its award replaces their agreement (§87(2)).

The word bestimmt, "intended", sounds as if it turned on the employer's purpose. The Federal Labour Court has long read it otherwise, and its 2016 decision on an employer's Facebook page restates the line. Equipment is "intended" to monitor when it is objectively suited to collect and record information about employees' behaviour or performance; the employer's subjective intention to monitor does not matter. It is enough that data entered manually is further processed by the equipment. Monitoring already begins with collecting the data, not only with evaluating it, and the stored data need not allow a complete judgement on its own: it suffices that, combined with other information, it makes one possible (71, Bundesarbeitsgericht, 2016). The case was brought by a Konzernbetriebsrat, a group works council, which is the body that faces group-wide systems.

Read against the AI boss, the test is met by almost every system this book describes: a claims-triage agent records which handler received which claim and when, a rostering system records acceptances of open shifts, and an office assistant records who was assigned which task, so each is objectively suited to show performance even if nobody runs a report. The court did limit the right where data cannot be attributed to individuals: if only a group's total performance is recorded, co-determination applies only where the monitoring pressure on the group passes through to its members (71, Bundesarbeitsgericht, 2016). That limit is the legal reason why team-level reporting, such as Hansa Mutual's oversight-quality metric OQM-HM-CLM-01, is easier to agree than individual figures, and it echoes the conclusion of AI Red Team, Chapters 6 and 15, that approval metrics are agreed with works councils first and reported in aggregate (F52) (43, Nordqvist, 2026).

Section 87(1) no. 6 is the strongest right but not the only one. Four further provisions apply to an AI boss, and the 2021 reform of the Act wrote artificial intelligence into three of them, §§ 80(3), 90 and 95 (207, Germany, 2021). Under §90(1) no. 3 the employer must inform the works council in good time about the planning of work procedures and workflows "including the use of artificial intelligence", and under §90(2) consult it early enough for its proposals to be taken into account. Under §94 personnel questionnaires and general appraisal principles need the works council's consent, which reaches any structured rating of performance that a scoring system applies. Under §95(1) selection guidelines for hiring, transfers, regrading and dismissals need its consent, and §95(2a) makes clear that this applies also where artificial intelligence is used to set up the guidelines. Under §80(3) the works council may call on an expert by agreement with the employer where its tasks require it; where it must assess the introduction or application of artificial intelligence, the expert is deemed necessary, and the same applies to a standing expert agreed for such matters (9, Germany, 2001). The reform ended the argument over whether a works council needs expertise to assess a model.

Three procedural rules matter for systems that change. In co-determined matters a terminated works agreement continues to apply until replaced (§77(6)), so the agreement on version 1 governs version 9 unless someone changes it. A standing conciliation committee can be set up by works agreement (§76(1)), which can shorten disputes. And the group works council is competent for matters that concern several group companies and cannot be regulated by the individual councils (§58(1)), so a group-wide agent such as Hansa's falls to the Konzernbetriebsrat.

Sweden and the Nordic model

The Nordic model rests on a different idea. The law sets a frame of information and negotiation, and collective agreements between unions and employers fill it, often with more detail than any statute. In Sweden the Co-determination in the Workplace Act (MBL) obliges an employer bound by a collective agreement to negotiate on its own initiative with the union before deciding on an important change in its activity, and before deciding on an important change in the working or employment conditions of the union's members (§11). On other questions it must negotiate when the union asks (§12); negotiation runs first locally and then centrally on request (§14), with minutes both parties approve on request (§16). The employer must keep the union continuously informed of how its activity develops and of the guidelines of its personnel policy (§19). Before having work done by people who will not be its employees, it must negotiate (§38) (205, Sveriges riksdag, 1976).

There is no veto over technology: the employer must negotiate before deciding, then may decide. The act invites the parties to agree co-determination in collective agreements on the management and distribution of work, including that certain decisions be taken by employee representatives or joint bodies (§32) (205, Sveriges riksdag, 1976). An AI boss in Sweden is governed less by the statute than by what the parties write. Safety representatives add a second channel under the Work Environment Act, which requires work to be organised so that strictly controlled or tied work is avoided or limited (75, Sveriges riksdag, 1977);

Norway and Denmark show the same model applied to monitoring. Norway's Working Environment Act allows control measures only where they have an objective reason in the undertaking's circumstances and do not place a disproportionate burden on the employee (§9-1). The employer must discuss the need for, design, implementation and significant change of control measures with the employees' union representatives as early as possible, must inform the employees affected of purpose, practical consequences and expected duration before the measure is implemented, and must evaluate the need for the measures regularly with the union representatives (§9-2) (236, Kingdom of Norway, 2005). In January 2026 Spekter, the employers' association for the Norwegian health trusts and other public enterprises, and the Norwegian Confederation of Trade Unions revised their basic agreement: its clause on control measures now opens with a reference to the act's Chapter 9, and a joint committee was asked to show how local cooperation under the agreements serves as a tool for introducing digital solutions, including artificial intelligence (237, Spekter & LO, 2026).

In Danish municipalities, the agreement on control measures between the municipal employers' organisation KL and the unions' negotiating cartel, in force since 1 April 2024, allows control measures under the employer's right to manage but requires them to have an objective operational reason, a reasonable relation between purpose and means, and to cause no offence or appreciable inconvenience. Its explanatory note names monitoring through digital work tools "based on artificial intelligence" among the measures covered and adds that the examples change with technology. The municipality must inform employees at least six weeks before introducing new control measures, stating its operational reasons, what information is collected and how it is handled, and an individual employee cannot consent to a control measure, either on hiring or later (238, KL & Forhandlingsfællesskabet, 2024). That rule stops individual opt-ins from replacing the collective process.

The Nordic arrangements have no German-style veto, but their statutes already speak of significant change and regular evaluation, so a change-based procedure fits them naturally.

The Netherlands, Spain and Italy

The Dutch Works Councils Act gives the ondernemingsraad a consent right over proposed decisions to adopt, amend or withdraw regulations on the processing and protection of personal data of the people working in the undertaking (Article 27(1)(k)) and over facilities aimed at or suited to observing or checking their presence, conduct or performance (Article 27(1)(l)) (206, Kingdom of the Netherlands, 1971). The wording "suited to" (geschikt voor) brings the Dutch test close to the German one, and the reference to amendment means that changes, not only introductions, are caught. A Dutch employer that changes what an AI boss records or who sees it needs a new consent.

Spain and Italy chose information rather than consent, but information of an unusual depth. Spain's 2021 law on delivery riders, the "Rider law", added a point (d) to Article 64.4 of the Workers' Statute: the works council has the right to be informed of "the parameters, rules and instructions on which the algorithms or artificial intelligence systems are based that affect decision-making that may affect working conditions, access to and maintenance of employment, including profiling" (77, Jefatura del Estado, 2021). The right is general, not limited to platforms. Italy's transparency decree, as amended, obliges employers to inform workers of the use of fully automated decision or monitoring systems relevant to hiring, management or termination, the assignment of tasks, and the monitoring and evaluation of performance; to explain their purpose, logic, main parameters and training data categories, control measures and accuracy metrics; to give the same information to the company union representatives; and to inform workers in writing, at least 24 hours in advance, of any change that alters the information given and the conditions of work (78, Repubblica Italiana, 2022). The Italian rule is the only one in this chapter that names a notice period for each change. A later amendment narrowed it to "fully automated" systems and exempted systems protected by trade secrecy.

Poland, where Hansa Mutual's service centre works, takes a fourth route: the purposes, scope and manner of workplace monitoring are set in the collective agreement or the work regulations, and employees are informed at least two weeks before monitoring starts (Labour Code, Articles 22² and 22³) (79, Rzeczpospolita Polska, 1974).

The European level: a floor of information and consultation

Union law sets a floor beneath these systems. Directive 2002/14/EC requires information and consultation in undertakings with at least 50 employees or establishments with at least 20, as each member state chooses (Article 3). Information and consultation cover "decisions likely to lead to substantial changes in work organisation or in contractual relations" (Article 4(2)(c)), and on those decisions consultation must take place "with a view to reaching an agreement" (Article 4(4)(e)). (51, European Union, 2002). An AI boss that changes how work is allocated is the kind of decision Article 4(2)(c) describes.

For groups that span member states, the European Works Council Directive applies to undertakings and groups with at least 1,000 employees in the member states and at least 150 in each of two (Article 2(1)(a) and (c)). Its competence is limited to transnational issues, which are those that concern the group as a whole or at least two undertakings or establishments in two member states (Article 1(3) and (4)). Under the subsidiary requirements, information and consultation cover in particular the "introduction of new working methods or production processes" (Annex I) (52, European Union, 2009). A group-wide AI system that allocates or evaluates work in two countries is a transnational issue; one confined to a single country is not, however large.

Two recent instruments build on that floor. The AI Act obliges deployers who are employers to inform workers' representatives and the affected workers before a high-risk AI system is put into service or used at the workplace, under the rules and procedures of Union and national law (Article 26(7)) (1, European Union, 2024). Recital 92 says that the Regulation is without prejudice to existing information and consultation duties, including those under Directive 2002/14/EC, and that the information right is laid down "without affecting any existing rights of workers"; Article 2(11) allows member states and collective agreements to be more favourable to workers. The Digital Omnibus moved Article 26 for Annex III systems to 2 December 2027 (Chapter 21) (2, European Union, 2026); national co-determination applies now. The Platform Work Directive extends information and consultation under Directive 2002/14/EC, for digital labour platforms, to "decisions likely to lead to the introduction of or to substantial changes in the use of automated monitoring systems or automated decision-making systems", entitles representatives to an expert of their choice, and requires a platform with more than 250 workers in the member state to bear the expert's proportionate costs (Article 13). Where there are no representatives, the platform informs the workers directly (Article 14) (47, European Union, 2024). Its formula, introduction or substantial change, states precisely the trigger this chapter is about.

The social partners reached the same point by agreement. Their 2020 framework agreement on digitalisation, signed by BusinessEurope, SMEunited, CEEP and the ETUC, sets out a "dynamic circular process" of joint mapping, assessment, adoption of measures and joint monitoring, and states that the requirements for trustworthy AI should be met "throughout the system's entire life cycle" under a "human in control" principle (85, BusinessEurope et al., 2020). De Stefano had argued that collective actors should take part in governing technology-enhanced management systems to secure a "human-in-command" approach (136, De Stefano, 2019). A life cycle is exactly what a one-off consent to version 1 lacks.

Whether this floor will become a general European right over algorithmic management is open. In December 2025 the European Parliament asked the Commission to submit a proposal on algorithmic management in the workplace, and recommended that the deployment of new systems, "or updates to existing systems", which directly affect remuneration, evaluation, working arrangements, task allocation or working time be deemed decisions likely to lead to substantial changes in work organisation, and so subject to consultation under Article 4(2)(c) of Directive 2002/14/EC (62, European Parliament, 2025). The request is observable; a proposal is expected only if the Commission acts on it after an impact assessment; a European framework that gives representatives co-decision over AI at work is speculation.

A system that changes with every release

The rights above were designed for systems installed once, such as a time clock introduced, negotiated and left alone for a decade. An AI boss changes continuously: Hansa Mutual's AIDR-0651 recorded that the claims-triage agent's allocation rules change monthly. A procured system such as Nordvik Region's ShiftWise receives releases on the supplier's schedule, and a vendor can switch on a feature in an office suite by default. Representatives, meanwhile, negotiate a works agreement every few years, and a union's local branch may meet monthly.

The mismatch produces two failure modes. The first is freezing: every change is treated as a new introduction requiring full negotiation, so either nothing changes, including fixes that would protect workers, or representatives approve dozens of changes without reading them, the rubber stamp this book rejects for managers. The second is drift: the parties agree on version 1, the agreement continues to apply under §77(6) BetrVG or the habit of the parties, and version 9 runs with inputs, recipients and objectives nobody negotiated. Hansa found drift in AIDR-0651: a throughput-based capacity factor had entered the allocation rules unnoticed.

The way out is already in the statutes. German co-determination attaches to "introduction and application", so a change in how equipment is applied can be co-determinable. Sweden asks whether a change is "important"; Norway, whether it is "significant"; the Platform Work Directive, whether it is "substantial"; Italy asks for notice of every change that alters the information given. None depends on a release number; all depend on what changes for the people managed: function, degree of delegation, inputs, purpose and recipients of outputs, which are the fields of the delegation profile of Chapter 4. AI Governance, Chapter 8 concluded that classification is triggered by change, not by the calendar (F16) (41, Nordqvist, 2026); the same principle applies to co-determination. An agreement can therefore govern change rather than versions, by classifying each change by its effect on the profile and the notice and attaching a procedure to each class.

Chapter 28 introduced three change classes for the worker information notice. The agreements in this chapter use the same three classes and attach the representatives' procedure to each:

• Class 1, renegotiation. A change that alters the bargain itself: a new management function, a higher degree of delegation, a new objective, a behavioural input, a new use of outputs in M2 Evaluation of individuals or in M3 Consequence, a new scope. In Germany it needs the works council's consent under a new version of the agreement; in Sweden it is an important change requiring negotiation under MBL §11 before the decision.
• Class 2, approval within agreed bounds. A change that the agreement anticipated and bounded: a non-behavioural input, a threshold or cap moved within an agreed range, a model version that passes the agreed tests. A joint committee approves it, or the union is informed with the diff and may call for negotiation.
• Class 3, information. A change that leaves every field of the profile and the notice unchanged: a defect fix, retraining within tested bounds, a forecast refresh, a security patch, the removal of an input. It is entered in a change log the representatives can read at any time and summarised monthly.

Three design rules keep the classes honest. The classification is made against a written test, applied to the diff of the delegation profile and the release manifest, not by the engineer's sense of how large a change is. Any doubt moves a change up a class, never down. And the representatives can reclassify: a works council or union that disagrees with a Class 2 or Class 3 label can require the change to be treated as Class 1, so the employer bears the cost of a wrong label. AI Governance, Chapter 13 concluded that pipelines enforce, observe and remember while forums decide what pipelines cannot (F18) (41, Nordqvist, 2026); change classes are that division of labour applied to co-determination.

Practical guide: Drafting a works agreement or local agreement for an AI system that changes continuously

Purpose. Write an agreement that governs the life of an AI boss, not its first version.

Roles. The system owner and HR (drafting the classes and the release manifest); the works council, its IT committee and its expert, or the local union and its safety representative; the data protection officer; the line managers who use the system; the AI governance lead who keeps the delegation register.

Steps.

1. Start from the delegation profile. Agree the degree per management function, the permitted and prohibited inputs, the recipients of individual outputs and the uses excluded from M2 Evaluation and M3 Consequence. Those fields are what the agreement protects.
2. Define the three change classes by reference to the profile fields, and list examples in each. Default: anything that raises a degree, adds a function, adds a behavioural input, adds a recipient of individual figures or a use in M3 Consequence is Class 1.
3. Set the procedure and lead time per class. Defaults: Class 1, full negotiation or consent before deployment, with the expert's report; Class 2, a decision by a joint committee within 10 working days, or information with the diff at least 5 working days before deployment with the right to call for negotiation; Class 3, change log entry before deployment and a monthly summary. Never set a lead time below a statutory minimum, such as Italy's 24 hours or Denmark's six weeks for new control measures.
4. Require a release manifest for every release: version, profile fields changed, proposed class, test results and the notice diff. For a procured system, put the manifest into the tender.
5. Give the representatives the right to reclassify a change upwards, and agree that a disputed Class 1 change goes to a standing conciliation body or, in Sweden, to central negotiation under MBL §14.
6. Agree data access: the change log, pattern reports at the agreed level, the contestation statistics of Chapter 27 and the expert's access to test environments under confidentiality.
7. Agree an annual review and triggers that reopen the agreement whatever the class of the change that caused them: a contestation pattern, a disparate-impact finding, a stop-path test that fails, a new entity or a change in law.
8. Record the agreement as a decision record, link it from the delegation profile and test the classification on the last ten releases before signing.

Definition of done. Every release since the agreement has a manifest and a class; no Class 1 change has been deployed without its procedure; the representatives have the change log and have exercised at least one review; and a test against past releases showed that the classes would have caught the changes that mattered.

Worked example: The change-class annex to Hansa Mutual's KBV-IT-2024

Illustrative. Hansa Mutual's group works agreement on IT systems KBV-IT-2024, negotiated by Thomas Brandt for the Konzernbetriebsrat, already carries an annex on the claims-triage agent's figures, shown in Chapter 13: individual throughput visible only to the handler and the handler's own team leader, ranking lists prohibited, access logs reviewed quarterly, and new metrics or new recipients only with the works council's consent. That annex settled what the figures may be; it did not say what happens when the allocation rules change monthly. The revision recorded in AIDR-0677 adds a second annex on change classes, applied first to the claims-triage agent and then to TalentLens and the sales-performance score. It adopts the three classes of Chapter 28 with one deliberate difference: new individual metrics and new recipients of individual figures, which Chapter 28 treats as Class 2 by default, are Class 1 here, because the first annex already made them subject to consent.

# agreements/KBV-IT-2024/annex-change-classes.yaml (illustrative)
agreement: KBV-IT-2024
annex: change classes
version: 1.0.0
parties: [Hansa Mutual group management, Konzernbetriebsrat]
legal_basis_de: [BetrVG 87(1) no. 6, BetrVG 94, BetrVG 95(2a), BetrVG 90(1) no. 3, BetrVG 58(1)]
decision_record: AIDR-0677
signed: "2026-12-14"
applies_to:
  - {system: claims-triage agent, profile: DLG-HM-CTA-01, first: true}
  - {system: TalentLens, profile: DLG-HM-REC-01, from: "2027-04-01"}
  - {system: sales-performance score, profile: DLG-HM-SPS-01, from: "2027-04-01"}
release_manifest_required: [version, fields_changed, proposed_class, tests, notice_diff]
classes:
  class_1_renegotiation:
    procedure: consent of the Konzernbetriebsrat; new annex version
    lead_time: before deployment, with the expert report under BetrVG 80(3)
    includes:
      - new management function or higher degree of delegation
      - change of allocation objective
      - behavioural input such as throughput, quality score or absence history
      - new individual metric or new recipient of individual figures
      - any use of figures in M2 individual rating or in M3 Consequence
      - new country or entity in scope
      - change to selection criteria in TalentLens
  class_2_committee_approval:
    procedure: joint IT committee decides within 10 working days within this annex
    bounds:
      daily_cap_per_handler: {min: 8, max: 20}
      preference_rules: within PREF-01 limits, at most 30 per cent of weekly allocation
      model_release: passes allocation regression and DIT-HM tests on agreed data
    includes: [new non-behavioural input, cap or threshold within bounds, model release]
    escalation: any committee member may refer the change to the full Konzernbetriebsrat
  class_3_information:
    procedure: change log before deployment; monthly summary to the IT committee
    includes: [defect fix, retraining within tested bounds, security patch, input removed]
reclassification: works council may require a higher class; the employer may not lower one
disputes: standing conciliation committee under BetrVG 76(1)
review: {annual: "2027-12-01", triggers: [contestation pattern, failed stop test, new law]}

The first Class 2 change came within a month. MDR-HM-0001 on 2027-01-12 recorded that the agent had allocated nine motor claims to a handler on an agreed reduced schedule, because it read only her open workload. The fix, a capacity field the team leader enters without reasons, is a new input, but not a behavioural one, so the joint IT committee approved it and the full Konzernbetriebsrat received the diff with the reissued worker information notice of Chapter 28. The daily cap per handler that AIDR-0675 introduced and the preference rule PREF-01 of Chapter 11 sit within the Class 2 bounds; anything beyond them would be Class 1. The capacity factor based on four weeks of throughput, removed in release 7.0 under AIDR-0651, is the example the annex was written to catch: as a behavioural input it is Class 1, and its silent arrival would now fail the manifest check.

The annex takes a legal form that German procedure requires. The full Konzernbetriebsrat concluded the agreement and its annex; the joint committee acts only within the bounds the annex sets. The Act allows a works council to transfer tasks for independent handling to committees, including joint committees with members named by both sides, and applies those rules to the group works council, but excludes the conclusion of works agreements from such a transfer (BetrVG §§ 27(2), 28 and 59(1)) (9, Germany, 2001). The architectural implication is the class boundary itself: whatever the annex has not already bounded cannot be approved by the committee and is Class 1.

The delegation profile as a negotiated document

The change classes work only because the delegation profile states, field by field, what the parties agreed. That makes the profile a negotiated document, not an engineering note. In Germany its fields correspond to the matters of co-determination: the outputs suited to monitor (§87(1) no. 6), appraisal principles (§94) and selection guidelines (§95). In Sweden its fields are the content of the negotiation under MBL §11 and of the information under §19. A profile version is the state of the bargain on a given date, and the version history is the record of how it changed and who agreed.

For M1 Direction, where most changes happen, the profile sets bounds inside which the employer may adjust without renegotiation. For M3 Consequence, it states which outputs are not used, and any change to that statement is Class 1 by definition. The profile also gives representatives the review triggers they need: AI Platform and Security, Chapter 5 concluded that decision records are short, immutable and re-opened by evidence (F38) (42, Nordqvist, 2026), and a negotiated profile is re-opened the same way.

Worked example: Nordvik Region's MBL negotiation protocol for ShiftWise

Illustrative. Nordvik Region negotiated the delegation profile DLG-NR-ROS-01 for ShiftWise with the union for nurses and midwives and the union for municipal and care workers under its co-determination act, which follows MBL §11, before the system's routes went live in home care, and recorded the profile as AIDR-0602. The negotiation also produced a standing protocol for what happens after go-live, recorded as AIDR-0627 and signed in the minutes both parties approved. It agrees that changes to the optimisation objectives or to the rules for open-shift offers are important changes that require renewed negotiation, while bug fixes and forecast refreshes are notified. It ties the classes to the worker information notice standard of AIDR-0634 (Chapter 28).

# decisions/AIDR-0627.yaml (illustrative)
id: AIDR-0627
title: ShiftWise, negotiation protocol for changes after go-live
date: "2027-01-20"
status: accepted
organisation: Nordvik Region
system: ShiftWise
delegation_profile: DLG-NR-ROS-01
related: [AIDR-0602, AIDR-0634, FRIA-2026-007]
basis: co-determination act on the pattern of MBL 11, 12, 14, 16, 19 and 38
parties: [union for nurses and midwives, union for municipal and care workers]
safety_representatives: consulted on every class 1 change and on rest-rule bounds
change_classes:
  class_1_negotiate_before_decision:
    - change to the optimisation objectives
    - change to open-shift offer rules or their published rotating order
    - any input on acceptance history, sickness absence or dashboard figures
    - any use of ShiftWise output in M2 Evaluation or M3 Consequence
    - agency staff offered shifts before own part-time staff, also under MBL 38
    - new unit or wave of the home-care rollout
  class_2_inform_with_diff:
    lead_time_working_days: 10
    union_may_call_for_negotiation: true   # MBL 12
    includes: [new working-time rule from the collective agreement, travel-time constraint]
  class_3_notify:
    includes: [forecast refresh, bug fix, supplier security patch]
    channel: change log and monthly summary to the unions
supplier_input: release notes with profile fields changed, 30 days before each release
escalation: central negotiation on request, MBL 14
review: {annual: "2027-12-31", triggers: [contestation pattern, fairness finding, new collective agreement]}
deciders: [HR director, Ingrid Halvorsen]

The protocol was tested within weeks. CNT-NR-0001, an assistant nurse's contestation of a route that left no time between two visits, led to a minimum travel-time constraint of routed time plus five minutes and never under ten. That changed the time a worker has between visits but no objective, input or offer rule, so the Region informed the unions with the diff as Class 2 and reissued the notice; the union for municipal and care workers did not call for negotiation. The next wave of the home-care rollout, in the third quarter of 2027, is Class 1 by definition. The protocol does not stop the Region deciding after negotiation, but it cannot decide an important change without it.

Negotiating the algorithm: what representatives ask for

The statutes and agreements in this chapter converge on what representatives ask for, and it maps onto the artefacts of this book: the logic (Spain's parameters, rules and instructions; the delegation profile and the worker information notice), limits on what individual data may feed (Annex 1 of KBV-IT-2024; the prohibited inputs of DLG-NR-ROS-01), expertise (BetrVG §80(3); Article 13(3) of the Platform Work Directive), standing review (Norway's statutory evaluation; the social partners' circular process) and, increasingly, a say in change. Eurofound's review of research on algorithmic management concluded that the opacity of algorithms complicates negotiation and makes adapted bargaining frameworks essential, that implementation of information and consultation rights on new technologies varies across member states, and that national implementation of the social partners' agreement has concentrated on digital skills and work-life balance more than on surveillance and dignity (150, Riso et al., 2025). That unevenness is observable; agreements that govern change are expected to spread as the Platform Work Directive is transposed and the AI Act's deployer duties apply. Where consent is required, representatives obtain terms; where only information is required, they obtain knowledge that becomes leverage in bargaining. Either way, the AI boss is better governed when they negotiate over a structured artefact rather than a supplier's brochure, which is the case for writing the delegation profile before the first meeting.

Across the usage modes

Co-determination attaches to function and effect, not to architecture, so every mode can trigger it. The modes differ in where representatives learn of the system late and where change classes matter most.

Usage mode	Typical AI boss	Where co-determination bites	Change-class risk
Mode 0 Ambient AI	Analytics features in a licensed suite	Feature switched on by a vendor release; often learned of late	Highest: register features and treat enabling one as Class 1
Mode 1 Conversational assistants	Chat assistant used by a manager to draft feedback	§94 appraisal principles where it structures ratings	Low; policy on use, not on releases
Mode 2 Embedded workplace assistants	Task lists with named owners in an office suite	§87(1) no. 6 because assignments are recorded	Medium; vendor releases change what is recorded
Mode 3 Agentic work tools	Agent that triages a team's inbox and assigns items	Recorded assignments and outcomes per person	Medium; new tools are Class 1
Mode 4 Built AI applications	ShiftWise, TalentLens, sales score	Every right in this chapter; §95 for selection	Monthly rule and model changes; manifest essential
Mode 5 Autonomous production agents	Claims-triage agent allocating to people	§87(1) no. 6, §58(1) group scope, EWC if transnational	Highest for M1 Direction: behaviour changes with every update

Nordvik Region and Hansa Mutual

Nordvik Region answers to unions under one regime. Its co-determination follows the Swedish model, its staff are covered by the sector's collective agreements, and its safety representatives add a second channel. DLG-NR-ROS-01 was negotiated before go-live in home care, AIDR-0627 governs change after it, and the classes have already handled a contestation and are ready for the next rollout wave. As a deployer of a procured system, it can classify only what the supplier discloses, hence the release notes 30 days ahead. Ingrid Halvorsen reports the protocol to the regional executive committee with the commission AIDR-0600.

Hansa Mutual answers to a Konzernbetriebsrat with co-determination rights and to six other national arrangements. The change-class annex to KBV-IT-2024 governs the claims-triage agent in Germany; each other entity has its own gate before the agent's allocation function or any revised profile reaches it. In the Netherlands the ondernemingsraad's consent under WOR Article 27(1)(k) and (l) is required, and the Dutch entity adopts the classes with Class 1 changes as consent matters. In Sweden the entity negotiates with its unions under MBL §11. In Norway the entity discusses the measure with union representatives as early as possible and evaluates it with them regularly under Working Environment Act §9-2. At the Polish service centre the purposes and scope of monitoring are set in the work regulations, consulted with the company trade union organisation, and staff are informed two weeks before use. The UK staff forum and the Swiss employee committee receive the manifest as group policy. The European Works Council is informed of the agent-boss pilot HM-AB-PILOT-01 because its next phase adds a claims team at the Polish service centre, which makes it a matter concerning establishments in two member states, and Thomas Brandt, who sits on both bodies, asked for the change-class annex to be the template for that information.

The difference is one of form: Nordvik's unions cannot veto a change but sit at the table for every important one; Hansa's Konzernbetriebsrat can withhold consent, but the group must keep one agent consistent across seven arrangements.

Summary

Co-determination is a European constant with national dialects. Germany grants co-determination over technical equipment suited to monitor under BetrVG §87(1) no. 6, and the Federal Labour Court's reading in 1 ABR 7/15, objective suitability without any intention to monitor, brings almost every AI boss within it; §§ 90, 94, 95(2a) and 80(3), as reformed in 2021, add information on AI in work procedures, consent to appraisal principles and selection guidelines, and an expert deemed necessary. Sweden obliges the employer to negotiate important changes under MBL §11 and leaves the detail to collective agreements; Norway requires discussion and regular evaluation of significant changes to control measures; Danish municipalities give six weeks' notice of new control measures, AI-based tools included. The Netherlands requires consent, Spain and Italy deep information, and Union law sets a floor extended by the AI Act's Article 26(7) and the Platform Work Directive's Article 13.

The chapter's answer to Q8 is that agreements should govern change rather than versions. Each release is compared with the delegation profile and the worker information notice and sorted into renegotiation, approval within agreed bounds, or information, by what changes for the people managed rather than by the size of the change; doubts move a change up and representatives can reclassify. Hansa Mutual's change-class annex to KBV-IT-2024 (AIDR-0677) and Nordvik Region's negotiation protocol for ShiftWise (AIDR-0627) show the approach under co-determination and under the duty to negotiate.

The chapter concerned M1 Direction, where rules change most often and Class 2 bounds give room to adjust; M2 Evaluation, where new metrics or recipients are Class 1; M3 Consequence, where any new use is Class 1 by definition and selection guidelines need consent; and M5 Care, where co-determination is the collective form of voice. The national rights are observable law; change-class agreements are observable in some organisations and expected to spread; a European co-determination right over AI is speculation. What the reader can do now is write the delegation profile first, agree the classes before go-live and test them against the last ten releases.

Chapter 24 turns to the remaining law that constrains the AI boss: occupational safety and health, equality law and the public authority as employer.

Sources

1. Federal Republic of Germany (2001). Betriebsverfassungsgesetz (BetrVG). Gesetze im Internet. (9, Germany, 2001)
2. Bundesarbeitsgericht (2016). Beschluss vom 13. Dezember 2016, 1 ABR 7/15. (71, Bundesarbeitsgericht, 2016)
3. Federal Republic of Germany (2021). Gesetz zur Förderung der Betriebsratswahlen und der Betriebsratsarbeit in einer digitalen Arbeitswelt (Betriebsrätemodernisierungsgesetz). Bundesgesetzblatt. (207, Germany, 2021)
4. Sveriges riksdag (1976). Lag (1976:580) om medbestämmande i arbetslivet (MBL). (205, Sveriges riksdag, 1976)
5. Sveriges riksdag (1977). Arbetsmiljölag (1977:1160). (75, Sveriges riksdag, 1977)
6. Kingdom of Norway (2005). Lov om arbeidsmiljø, arbeidstid og stillingsvern mv. (arbeidsmiljøloven). Arbeidstilsynet. (236, Kingdom of Norway, 2005)
7. Spekter & LO (2026). Protokoll fra forhandlingsmøte 27. og 28. januar 2026 om revisjon av Hovedavtalen. (237, Spekter & LO, 2026)
8. KL & Forhandlingsfællesskabet (2024). Aftale om kontrolforanstaltninger. (238, KL & Forhandlingsfællesskabet, 2024)
9. Kingdom of the Netherlands (1971). Wet op de ondernemingsraden (WOR). wetten.overheid.nl. (206, Kingdom of the Netherlands, 1971)
10. Jefatura del Estado (2021). Ley 12/2021, de 28 de septiembre (Rider law), amending Article 64.4 of the Workers' Statute. BOE. (77, Jefatura del Estado, 2021)
11. Repubblica Italiana (2022). Decreto legislativo 27 giugno 2022, n. 104 (Decreto Trasparenza). Normattiva. (78, Repubblica Italiana, 2022)
12. Rzeczpospolita Polska (1974). Ustawa z dnia 26 czerwca 1974 r. Kodeks pracy. (79, Rzeczpospolita Polska, 1974)
13. European Parliament and Council of the European Union (2002). Directive 2002/14/EC establishing a general framework for informing and consulting employees in the European Community. Official Journal of the European Communities. (51, European Union, 2002)
14. European Parliament and Council of the European Union (2009). Directive 2009/38/EC on the establishment of a European Works Council (recast). Official Journal of the European Union. (52, European Union, 2009)
15. European Parliament and Council of the European Union (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act). Official Journal of the European Union. (1, European Union, 2024)
16. 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)
17. 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)
18. BusinessEurope, SMEunited, CEEP and ETUC (2020). European Social Partners Framework Agreement on Digitalisation. (85, BusinessEurope et al., 2020)
19. De Stefano, V. (2019). 'Negotiating the Algorithm': Automation, Artificial Intelligence and Labour Protection. Comparative Labor Law & Policy Journal. (136, De Stefano, 2019)
20. European Parliament (2025). Digitalisation, artificial intelligence and algorithmic management in the workplace: shaping the future of work. Resolution of 17 December 2025. (62, European Parliament, 2025)
21. Riso, S., Adăscăliţei, D. & Oosterwijk, G. R. (2025). Implications of algorithmic management for work, employment and social dialogue: Literature review. Eurofound. (150, Riso et al., 2025)
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)
23. Nordqvist, G. (2026). AI Governance: Governing Artificial Intelligence in European Enterprises and Public Authorities. The Architecture as Code Series, Book 2. (41, Nordqvist, 2026)
24. 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)