Performance Scores, Productivity Metrics and Dismissal by Algorithm

A performance score looks like an instrument of measurement. In practice it is often an instrument of decision: it sets a bonus, triggers a warning, places a name on a list for a restructuring or, in the reported cases, ends a working relationship by email. This chapter examines how individual scores are computed from activity data, what they measure and what they miss, what the reported cases and the courts found about dismissal and deactivation by algorithm, and what the GDPR, the AI Act and national dismissal law require when a score shapes a decision about a person. It concerns two management functions: M2 Evaluation, the measuring, rating and judging of performance and conduct, and M3 Consequence, the decisions on pay, promotion, discipline, dismissal and deactivation that follow from it. It is the chapter in which the book's rule that every M3 Consequence decision has a named human decider, and that Level 4 Autonomous is never used for M3 Consequence, is first tested against practice. The test is harder than it looks, because a score rarely becomes a decision by anyone deciding that it should.

Figure 8.1: How a score travels from M2 Evaluation to M3 Consequence, with the checkpoints that keep it advisory and the automatic path the book excludes.

The score

Every performance score is built in three layers. The first is activity data: scans per hour in a warehouse, deliveries on time, calls handled in a contact centre, tickets closed, documents edited and messages sent in an office suite. The second layer is the indicator, a rate computed over a period and compared with a target or a peer group. The third layer is the composite score, which weights several indicators into one figure or band and often ranks people against each other.

Each layer is a choice. The data are what the system can log, not what the work consists of. The indicators are what someone decided counts. The weights and bands are a judgement about what matters, frozen in configuration and applied to everyone in the same way. In this book's vocabulary the whole construction is M2 Evaluation performed by software, and Ajunwa's account of the quantified worker describes how much of that evaluation has moved from the manager's observation to the system's record (143, Ajunwa, 2023).

Three families of scores matter for this chapter. Rate scores in paced work, such as warehouse units per hour and time off task, were the subject of Chapter 6. Rating scores on platforms, built from customer ratings, acceptance rates and on-time performance, were the subject of Chapter 5. The third family is newer and spreads quietly: composite scores and dashboards in offices, built from the logs of collaboration suites and business systems, and from monitoring software installed on workers' computers to record application use, activity and presence. Critics call such products bossware. Whatever the name, they compute M2 Evaluation from data collected for another purpose and present it to a manager as a figure per person.

The AI Act names this activity directly. Annex III, point 4(b), lists as high-risk AI systems intended to be used to make decisions affecting terms of work-related relationships, the promotion or termination of work-related contractual relationships, or to monitor and evaluate the performance and behaviour of persons in such relationships (1, European Union, 2024). Article 6(3) adds that a system referred to in Annex III is always considered high-risk where it performs profiling of natural persons, which a score of an individual's performance does almost by definition. A simple rate calculation may not be an AI system within the Act's meaning, but composite scores that weight, predict or rank are the kind the annex describes.

What scores miss

The research on feedback and measurement predicts most of what went wrong in the cases below. Kluger and DeNisi's meta-analysis of feedback interventions found that feedback improved performance on average, but that in more than a third of the cases it reduced performance; their explanation was that feedback which draws attention to the self, rather than to the task, can displace the effort it was meant to direct (116, Kluger & DeNisi, 1996). A score attached to a person's name and to their pay is feedback of exactly that self-focused kind.

Campbell stated the second problem as a law: the more any quantitative social indicator is used for social decision-making, the more subject it will be to corruption pressures and the more apt it will be to distort and corrupt the social processes it is intended to monitor (110, Campbell, 1979). An indicator observed is useful. The same indicator used to decide about a person changes the behaviour it was meant to observe. Muller's survey of metric fixation across hospitals, schools, policing and business found the same pattern: what is counted is done, what is not counted is neglected, and the measure is gamed (123, Muller, 2018).

Three kinds of work are systematically missed. The first is work that has no log: helping a colleague, calming an angry customer, training a new starter, noticing that something is wrong. The second is work whose quality is not visible in its volume: the claims handler who takes longer over a complicated claim and gets it right, the care assistant who stays five minutes longer with a confused resident. The third is the work of exceptions, where the case that does not fit is the one that most needs judgement and the one most likely to depress a score. Bernstein's field study in a large factory found that workers who were fully observed hid their improvements and workarounds from the line, and that productivity rose when a team was given a degree of privacy from observation (122, Bernstein, 2012). The figure on the dashboard may therefore show less than the work, and the gap grows with the stakes attached to the figure.

None of this makes scores useless. A team-level figure, or an individual figure the worker sees first and uses to plan their own day, can serve M4 Development and M5 Care. The problem begins when a score that misreads work is used for M3 Consequence without a person who understands the work standing between the figure and the decision.

Dismissal by algorithm, reported

The cases below are reported, and some are contested; each is described by what was reported, by whom and on what evidence, before any interpretation.

The warehouse

In April 2019 The Verge published documents that Amazon had filed in a dispute before the United States National Labor Relations Board. According to the company, roughly 300 full-time associates at one fulfilment centre in Baltimore were terminated for productivity reasons between August 2017 and September 2018, at a facility the company said employed about 2,500 full-time staff. A letter in the filing stated that the system tracked the rate of each associate's productivity and automatically generated any warnings or terminations regarding quality or productivity without input from supervisors; Amazon said that supervisors could override the process and that an appeal system was part of it (187, Lecher, 2019). Chapter 6 treats the pace and health side. Here the point is that the termination was generated by the system, and the human role was described as the possibility of override, not as a decision.

Amazon Flex

In June 2021 Bloomberg reported on Amazon Flex, the programme under which independent contractors deliver parcels using their own vehicles. The report followed a driver in Phoenix who had delivered for almost four years and was terminated by an automated email, and drew on interviews with drivers and former Amazon managers who said that the largely automated system was insufficiently aware of the obstacles drivers face, such as locked buildings, and that drivers who believed they had been wrongly deactivated had little practical recourse (198, Soper, 2021). Amazon told Bloomberg that the drivers' accounts were anecdotal and did not reflect the experience of most Flex drivers. Because Flex drivers are contractors, the termination was a deactivation of access to work, which Chapter 5 treats as the platform form of M3 Consequence.

Xsolla

In August 2021 Xsolla, a payment-processing company serving the games industry, laid off about 150 employees, roughly a third of its workforce. An email from the chief executive to those affected, reported by Game Developer, said that his big data team had analysed their activities in Jira, Confluence, Gmail, chats, documents and dashboards and had tagged them as unengaged and unproductive employees. In a later interview he linked the layoffs to slowing growth and a plan to cut the salary budget by 10 per cent (200, McAloon, 2021). The case differs from the two above in one important respect: a named executive took the decision and said so. What the analysis did was to choose the people. An activity score built from office tools, with no recorded relation to the quality of anyone's work, became a redundancy list.

Deactivation and the Amsterdam judgment

On 4 April 2023 the Amsterdam Court of Appeal decided the appeal of four drivers, three in the United Kingdom and one in Portugal, whose Uber accounts had been deactivated on grounds of alleged fraud (67, Gerechtshof Amsterdam, 2023). Uber explained that rule-based software raised fraud signals and that its risk team then investigated manually before deciding. The court applied the guidelines on automated decision-making endorsed by the European Data Protection Board, under which human involvement must be meaningful and not a token gesture, must be carried out by someone with the authority and competence to change the decision, and must take all relevant data into account (61, Article 29 Data Protection Working Party, 2018). It held that Uber bore a heightened burden to show such involvement, because the information about it lay in Uber's own domain, and that for three of the drivers it had not done so: the records showed that one to three employees had made a note in response to the fraud signal, but not that their review was meaningful. For the fourth driver, Uber showed that a personal conversation had taken place before the deactivation, and the court accepted that as genuine human involvement. The court ordered Uber to give the three drivers meaningful information about the logic involved under GDPR Article 15(1)(h), on pain of a penalty of EUR 4,000 per day.

The finding transfers directly. A human in the process does not by itself take a decision out of Article 22; what counts is whether the review could change the outcome and whether the employer can show that it did. A note on a file is not evidence of a decision. A conversation with the person concerned is.

What the cases show together

Case	What the system did	Where the human stood	What was found or reported
Baltimore fulfilment centre, 2017–2018	Tracked individual rates; generated warnings and terminations	Supervisors could override	About 300 productivity terminations in a year at one site, reported from the company's own filing
Amazon Flex, 2021	Rated drivers and terminated access by email	Appeal after the event	Drivers and former managers described a largely automated process with little recourse
Xsolla, 2021	Analysed office-tool activity and tagged staff	The chief executive took and announced the decision	A layoff list drawn from activity data
Uber, Amsterdam 2023	Raised fraud signals that led to deactivation	Risk-team notes; one personal conversation	Notes did not prove meaningful involvement; the conversation did

The four cases differ in law and in fact, but they share a structure. In each, the score selected the person and the human role, where there was one, ratified the selection. That is the move from M2 Evaluation to M3 Consequence that this chapter is about.

The productivity-score backlash

Knowledge work had its own warning in 2020. Microsoft's Productivity Score, a feature of Microsoft 365 that reported on the use of the suite's tools, could show activity figures for named users to administrators. In response to criticism, Microsoft announced on 1 December 2020 that it was removing user names from the product so that the data would be aggregated only at the organisation level, that the score could not be used to monitor individual employees, and that it measured organisational adoption of technology and not individual user behaviour; the announcement said the change was made in response to feedback over the preceding week (189, Spataro, 2020).

The episode shows that a vendor can move a feature from individual to aggregate reporting quickly once the individual figure is seen for what it is: an evaluation of people built from data they generate in the ordinary course of work. It also shows where the pressure comes from. Whether an office feature reports on named individuals is a configuration choice, often made by an administrator rather than a manager or a works council. Chapter 10 treats the office analytics in detail. For this chapter the lesson is narrower: whether a figure is reported per person is the first decision about whether it can become M3 Consequence, and it should be taken deliberately, with workers' representatives, and recorded.

Trusting the system over the person

The most serious modern warning about trusting system data over people does not involve AI at all. From 2000 to 2013 the Post Office in the United Kingdom prosecuted sub-postmasters and branch staff in reliance on accounting data from its Horizon system, asserting each time that the losses Horizon showed were real and that the data were wholly reliable. In the Horizon Issues judgment of December 2019 the High Court found that Legacy Horizon was not remotely robust, citing the number, extent and impact of the bugs, errors and defects that the evidence revealed (72, High Court of Justice & Queen's Bench Division, 2019). The first volume of the public inquiry's final report, published in July 2025, concluded that many hundreds of people were wrongly convicted and many thousands wrongly held responsible for losses that were illusory, and found it likely that about 1,000 people were prosecuted and convicted on Horizon evidence (203, Post Office Horizon IT Inquiry, 2025).

Sub-postmasters were contractors, and Horizon was an accounting system, not a performance score. The lesson still transfers in three ways. An organisation that has invested in a system tends to treat its output as fact and the person who disputes it as the problem. The person affected cannot see inside the system, so the burden falls on the party with the least information, the asymmetry the Amsterdam court corrected by placing a heightened burden on Uber. And everyone who relied on the data could say that the data, not they, had decided. Elish called this position the moral crumple zone: the human nearest to an automated system absorbs the blame for its failures while having had little control over them (170, Elish, 2019). In algorithmic management the crumple zone can run the other way too: the worker absorbs the consequence while the manager points to the score.

Scores and Article 22

GDPR Article 22(1) gives the data subject the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning them or similarly significantly affects them (5, European Union, 2016). Article 22(2) allows such decisions where they are necessary for a contract, authorised by Union or member-state law with suitable safeguards, or based on explicit consent, and Article 22(3) requires in the contract and consent cases at least the right to obtain human intervention, to express a point of view and to contest the decision. A dismissal, a disciplinary sanction, a deactivation and a pay decision all produce legal effects or similarly significant ones. The question for a performance score is when a decision that a person formally takes is nevertheless based solely on automated processing.

The Court of Justice answered a closely related question in December 2023. In SCHUFA it held that the automated establishment of a probability value by a credit information agency is itself automated individual decision-making within the meaning of Article 22(1) where a third party to which the value is transmitted draws strongly on it to establish, implement or terminate a contractual relationship with the person (63, CJEU, 2023). Treating the score as a mere preparatory act, the Court reasoned, would risk circumventing Article 22 and leave a lacuna in legal protection.

The case concerned credit, and its application to scores inside one employer is a matter of interpretation, but its architectural implication is clear: labelling a score advisory does not settle whether Article 22 applies; what settles it is whether the decider draws strongly on the score or genuinely weighs it with other information. The guidelines on automated decision-making, applied by the Amsterdam court, point the same way: a controller cannot avoid Article 22 by fabricating human involvement, and routine application of automatically generated profiles without actual influence on the result remains a solely automated decision (61, Article 29 Data Protection Working Party, 2018). Article 88 allows member states to lay down more specific rules for processing employees' data, by law or by collective agreement, with particular regard to transparency and to monitoring systems at the workplace (5, European Union, 2016).

The AI Act adds obligations that do not depend on whether a decision is solely automated. A deployer of a high-risk system under Annex III point 4(b) must assign human oversight to natural persons who have the necessary competence, training and authority, as well as the necessary support (Article 26(2)); must inform workers' representatives and the affected workers before putting the system into service at the workplace (Article 26(7)); and must, under Article 86, give a person affected by a decision taken on the basis of the system's output clear and meaningful explanations of the role of the system and the main elements of the decision (1, European Union, 2024). Article 14(4)(b) requires providers to design high-risk systems so that the people overseeing them can remain aware of automation bias. The Digital Omnibus set the application of the high-risk obligations for Annex III systems at 2 December 2027 (2, European Union, 2026); the GDPR applies now. Article 5(1)(f) already prohibits inferring emotions at the workplace, except for medical or safety reasons, which rules out engagement scores built on inferred emotion.

The Platform Work Directive is more direct. Article 10(5) requires that any decision to restrict, suspend or terminate the contractual relationship or the account of a person performing platform work, or any other decision of equivalent detriment, be taken by a human being, and Article 11 gives the right to an explanation, a contact person, a written statement of reasons and a review within two weeks (47, European Union, 2024). Member states must transpose it by 2 December 2026. It does not bind ordinary employers, but it states as law the rule this book proposes for all of them.

National dismissal law has always required a reasoned human decision. In Germany a dismissal of an employee with more than six months' service is invalid if it is not socially justified by reasons in the person, in their conduct or in urgent operational requirements (KSchG §1) (217, Germany, 1969); the works council must be heard before every dismissal and told the reasons, and a dismissal without that hearing is invalid (BetrVG §102(1)); and the works council co-determines on technical equipment designed to monitor employees' behaviour or performance (§87(1) no. 6) and must consent to general assessment principles (§94(2)) (9, Germany, 2001). In Sweden a dismissal by the employer must rest on objective grounds, which may be redundancy or circumstances relating to the employee personally (LAS 7 §), and an employer who wishes to dismiss for personal reasons must notify the employee in advance and, if the employee is a union member, the local union at the same time (LAS 30 §) (218, Sveriges riksdag, 1982). A score can supply facts to such a procedure. It cannot supply the reasons, because the reasons must be an employer's judgement that a works council, a union and a court can examine.

From evaluation to consequence

Scores still become decisions because the move from M2 Evaluation to M3 Consequence rarely happens through a decision to delegate. It happens through three channels, each of which looks administrative.

The first is the pay input. A score is defined as one input to variable pay, alongside the manager's assessment. Over a few cycles the manager's assessment converges on the score, because departing from it requires an explanation and following it does not. The second is the warning. A threshold is set below which the system generates a notice, a task for the manager or an entry in the personnel file. The notice is formally a prompt, but it is dated, filed and later cited. The third is the list. When costs must be cut or a bonus pool shared, a ranked list is already available, and the people at its bottom become the shortlist. Xsolla is the extreme case of the list; the Baltimore warnings are the extreme case of the warning.

In each channel, the degree of delegation on paper stays at Level 0 Advisory or Level 1 Assisted, while the degree in practice moves towards Level 3 Bounded or Level 4 Autonomous: the system effectively selects, and the human signs within bounds the system has already set. The degree of delegation is set by use, not by the software. A score read by a manager who knows the work and departs from it when the work says so is Level 0 Advisory. The same score read by a manager who follows it in every case, in seconds, and cannot say why is Level 4 in fact, and for M3 Consequence that is excluded.

The drift is the default, not the exception. Green reviewed policies that require human oversight of government algorithms and concluded that people are often unable to perform the oversight functions such policies assign to them, so that the requirement can legitimise flawed systems rather than correct them (178, Green, 2022). The series drew the practical conclusion earlier: AI Governance, Chapters 10 and 21, concluded that human oversight must be designed, not declared, and that approvals given in seconds are a signal of rubber-stamping (41, Nordqvist, 2026); AI Red Team, Chapters 6 and 15, measured approval quality by time to approval, rejection rate and reversals, and required that where works councils have rights those metrics be agreed first and reported in aggregate (43, Nordqvist, 2026); and The Transformed Security Architect, Chapter 14, made review that does not start from the AI's answer the core of verification (46, Nordqvist, 2026). This chapter applies those conclusions (F21, F52 and F76) to managers reading scores about people.

It also answers part of Q2, raised in AI Threat, Chapter 30: who is accountable for a consequence decision an AI recommended and a manager accepted without reading it (44, Nordqvist, 2026). The manager is, and so is the employer that let the conditions arise; but accountability on paper is worth little unless the record shows what the manager actually weighed. That is why the test of whether a score is advisory must be a test of practice. It also bears on Q3, how a decision made with AI is shown to the worker, a works council, a court and a supervisor: what each of them needs is the same record of what the decider weighed, not the score alone.

Keeping a score advisory

A score stays advisory when five conditions hold together, and each can be checked.

1. The manager forms a view before seeing the score. For any M3 Consequence use, the manager records a short assessment of the person's work before the score is shown: review that does not start from the AI's answer.
2. Thresholds trigger a conversation, not a sanction. A score below a threshold opens a conversation about the work, which is M4 Development and M5 Care as much as M2 Evaluation, and never generates a warning, a file entry or a pay effect by itself.
3. The score is never the sole ground of a consequence. Every consequence has reasons in the manager's own words; a score may appear among the facts but not as the reason.
4. The worker sees what the manager sees. The worker can see their score, its indicators and the data behind them before any consequence, and can correct errors and add context.
5. Departures are expected and recorded. A manager who departs from the score records why, and a manager who never departs is asked why. Departures are the evidence that the decision was human.

The record that shows all five is the Management Decision Record of Chapter 25, with the score as an input, the manager's assessment and reasons, the named decider and the path to contest; the contestation record of Chapter 27 carries any challenge. F38's discipline applies: short, numbered, immutable records validated against a schema, which an auditor, a works council or a court can read (42, Nordqvist, 2026). In a public authority, F31 adds that the line between decision support and automated decision is drawn in the data model and the audit log, and every case has a reasons log (41, Nordqvist, 2026).

Practical guide: Testing whether a performance score is decisive in practice

Purpose. Establish, from records rather than declarations, whether a score that is classified as advisory for M3 Consequence is advisory in practice, and act when it is not.

Roles. The HR lead who owns the delegation profile; the line managers who use the score; the works council or union representatives, who agree the design and receive the results; the data protection officer, who checks that the test does not become monitoring of individual managers; and the workers affected.

Steps.

1. Agree the design first. Before any data are drawn, agree with the workers' representatives what will be measured, at what level of aggregation and for what purpose. Default: results reported per unit of at least ten managers, never per named manager, as F52 requires.
2. Draw the decisions. Take every consequence decision in which the score was an input over the last full cycle: pay inputs, warnings, promotion shortlists. Default: one annual pay cycle or two quarterly cycles.
3. Measure agreement. Compute the share of decisions that match the score's band or rank exactly. Default alert: agreement above 90 per cent across a unit.
4. Measure departures and their reasons. Compute the share of decisions that depart from the score, and the share of departures with written reasons in the manager's own words. Default alert: departures below 5 per cent, or departures without reasons above 20 per cent.
5. Measure time. Where the system logs it, compute the median time between the manager opening the score and recording the decision. Default alert: a median under two minutes for a pay or disciplinary input.
6. Test with seeded cases. With the representatives' agreement, include in a review cycle a small number of cases in which the score is known to misread the work, for example a period of approved leave counted as low activity, and check whether managers catch them. Default: at least three seeded cases per unit per year, removed before any real decision is taken.
7. Read the contestations. Check whether each contestation in the period changed the score, the decision or neither.
8. Classify and act. If two or more alerts fire in a unit, record the score as decisive in practice for that unit, treat the use as Level 3 Bounded or above, which is excluded for M3 Consequence, suspend the score's use as a pay or disciplinary input there, and agree corrective measures with the representatives: training, the view-first rule, a changed threshold or withdrawal of the score from that use.
9. Report. Give the aggregate results to the representatives and the register owner, and a summary to the workers affected.

Architecture as Code (Chapter 9C) treated risk as a versioned record with an owner and expiring exceptions, and made a failing control metric fail the pipeline (40, Nordqvist, 2025). The practice test is that control metric for a score: a unit classified as decisive in practice fails the check, and the score's consequence use there stops until the measures are in place, without waiting for the next policy review (F2).

Definition of done. For every score used as an input to M3 Consequence, the delegation register records the latest test date, the aggregate results, the classification (advisory in practice or decisive in practice) and the measures taken, and the next test is scheduled within twelve months. Adapt the thresholds to the collective agreement or works agreement in force.

Worked example: Hansa Mutual's sales-performance score

Illustrative. Hansa Mutual's sales-performance score rates employed field-sales and contact-centre sales staff on conversion, retention and complaint indicators. It is built on the group platform (Mode 4). Sales managers use it as one input to variable pay. Its delegation profile records Level 1 Assisted for M2 Evaluation and Level 0 Advisory for M3 Consequence, and excludes every other consequence use.

# management/delegation-register.yaml (excerpt, illustrative)
- id: DLG-HM-SPS-01
  version: 3
  system: sales-performance-score
  usage_mode: "Mode 4 Built AI applications"
  owner: group-head-of-sales-hr
  approved_by: chief-hr-officer
  representatives:
    germany: Konzernbetriebsrat, annex to KBV-IT-2024
    netherlands: ondernemingsraad, consent recorded 2026-05-12
    european_works_council: informed 2026-05-20
  ai_act:
    annex_iii: "4(b) monitor and evaluate performance"
    article_26_7_notice: issued 2026-06-01
  functions:
    M2_Evaluation:
      level: "Level 1 Assisted"
      note: Score computed quarterly; employee sees score and indicators first
    M3_Consequence:
      level: "Level 0 Advisory"
      permitted_use: one input to the variable-pay proposal
      named_decider: sales manager of the employee
      conditions:
        - Manager records own assessment before opening the score
        - Score is never the sole or stated reason for the pay proposal
        - Departure from the score band needs no approval, only a reason
        - Pay proposal is an MDR with the score as input and the reasons in the manager's words
  excluded_uses:
    - Warnings, disciplinary measures or dismissal
    - Selection for redundancy or restructuring
    - Promotion shortlists without a separate manager assessment
  practice_test:
    last_run: 2026-11-02
    result: decisive in practice in one unit; corrective measures agreed
    next_due: 2027-05-31
  contestation_path: CNT-HM records, decision within 15 working days

In October 2026 a field-sales employee of the Dutch entity contested her score for the third quarter. Her complaint indicator had fallen into the lowest band, and with it her composite score. She pointed out that complaints in the Dutch market had risen across the board after a change to a motor-policy product, and that the indicator compared her with the group median rather than with her own market. The contestation record shows what was examined and decided.

# contestations/CNT-HM-0001.yaml (illustrative)
id: CNT-HM-0001
opened: 2026-10-19
contested_record: sales-performance-score, Q3 2026, one employee, Dutch entity
contested_by: employee, supported by an ondernemingsraad member
grounds:
  - Complaint indicator compared with group median, not market baseline
  - Market-wide rise in complaints after a product change in the quarter
reviewer: HR business partner sales, not the employee's line
named_decider: sales manager of the employee
evidence_examined:
  - Indicator definitions in DLG-HM-SPS-01 version 2
  - Complaint rates by market, Q2 and Q3 2026
  - Pay proposal MDR of the employee and the manager's recorded reasons
findings:
  - Complaint indicator penalised staff in markets with higher complaint rates
  - Manager's pay proposal matched the score band without a written assessment
  - In the Dutch sales unit 97 per cent of proposals matched the score band
  - Median time from opening the score to recording the proposal was 48 seconds
outcome:
  score: recalculated against the market baseline; band corrected
  pay_proposal: re-made by the manager with own assessment first
  system_change: complaint indicator normalised per market from Q4 2026
  delegation_profile: DLG-HM-SPS-01 raised to version 3
  practice_classification: decisive in practice in the Dutch sales unit
decided: 2026-11-02
employee_informed: 2026-11-03, in writing, with reasons
escalation_open_until: 2026-12-03

The second record matters more than the corrected score. The profile said Level 0 Advisory for M3 Consequence. In one unit the practice was Level 4 in all but name: the pay proposal followed the score in almost every case, within a minute, without written reasons. The contestation exposed it; the practice test of the guide above confirmed it across the unit; and the corrective measures, the view-first rule and a recalculated indicator, are recorded in version 3 of the profile.

Across the degrees of delegation

Degree of delegation	What it looks like for a performance score	Acceptable for M3 Consequence?
Level 0 Advisory	Score shown to a manager who has already formed a view and who departs from it when the work says so	Yes, with reasons recorded
Level 1 Assisted	Score proposes a band or rank; the manager reviews each case and writes reasons	Yes, with the practice test
Level 2 Supervised	Score sets the pay input unless the manager objects within a window	No: the default is the system's
Level 3 Bounded	System issues warnings or pay effects within thresholds; managers handle exceptions	No
Level 4 Autonomous	System terminates, deactivates or dismisses, as reported at Baltimore and on Flex	Never

The usage mode changes where the risk enters. Scores are usually computed by Mode 4 Built AI applications or by Mode 0 Ambient AI analytics in licensed suites, and Mode 2 Embedded workplace assistants increasingly summarise a person's output for their manager. A Mode 0 feature switched on by an administrator can create an individual score that nobody chose, which is why the delegation register lists office features as well as built systems.

Outlook

Individual productivity scores, automated warnings and deactivations are observable in the cases and court records above. The extension of scoring into knowledge work through office suites and monitoring software is observable and growing, and the Microsoft episode shows that individual reporting is a configuration choice that can be reversed. Regulatory limits on scores used in pay and discipline are expected: the high-risk obligations of the AI Act apply to Annex III point 4 systems from 2 December 2027, the Platform Work Directive's human-decision rule for deactivation must be transposed by 2 December 2026, and national courts are applying the reasoning of SCHUFA and the Amsterdam judgment. That ordinary employers will adopt the Directive's rule voluntarily for all consequence decisions is speculation; this book recommends it.

Nordvik Region and Hansa Mutual

Nordvik Region's home-care dashboards show planned against actual time per visit and per route at Level 0 Advisory for M2 Evaluation under DLG-NR-HCD-01. Individual figures are visible to the unit manager. Ingrid Halvorsen, the HR director, has decided that they may not be used in the annual pay review or in any disciplinary matter, recorded as a condition in the profile, not left as an understanding. For Mikael Strand in the Strandvik unit, a visit that runs long therefore opens a conversation about the route and the residents, not a file note, which is how the unit handled the route contested in CNT-NR-0001. Because the dashboards are built in house and their status under Annex III point 4(b) is an open question in this book, the Region treats them as if they were high-risk for the purpose of informing the unions and safety representatives, under its co-determination act, before any change in what is shown per person.

Hansa Mutual uses AI as an input to pay, which Nordvik does not. The sales-performance score is advisory on paper under DLG-HM-SPS-01, and CNT-HM-0001 showed that in one unit it was decisive in practice; the practice test is now part of the profile, and the Konzernbetriebsrat, through Thomas Brandt, receives its aggregate results. Neither organisation uses AI in discipline or dismissal, and both record that as a decision rather than an absence: Nordvik in the conditions of DLG-NR-HCD-01, Hansa in the excluded uses of DLG-HM-SPS-01. A recorded exclusion can be audited and enforced. An absence can quietly be filled the next time a cost reduction needs a list.

Summary

Performance scores turn activity data into ranked figures about people, which is M2 Evaluation performed by software. The measurement literature predicts their weaknesses: feedback focused on the person can lower performance, an indicator used for decisions distorts what it measures, and unlogged work, quality and exceptions are what a score misreads.

The reported cases, from the Baltimore fulfilment centre and Amazon Flex to Xsolla and the Uber drivers in Amsterdam, share one structure: the score selected the person and the human role ratified the selection. The Amsterdam court held that notes on a file do not show meaningful human involvement and that a conversation with the person does. The Post Office Horizon judgments and inquiry show what happens when an organisation trusts its system's output over the people who dispute it.

The law converges on the rule this book proposes for M3 Consequence. GDPR Article 22, as the Court of Justice read it in SCHUFA, reaches a score on which a decision draws strongly; the AI Act classifies performance evaluation as high-risk and requires competent, authorised oversight, information to workers' representatives and explanations; the Platform Work Directive requires a human being to take every deactivation decision; and German and Swedish dismissal law require reasons that a works council, a union and a court can examine. Because a score becomes a decision through a pay input, a warning or a list, whether it is advisory is a test of practice: agreement rates, departures with reasons, time to decide, seeded cases and contestations, agreed with workers' representatives and reported in aggregate.

Chapter 9 turns from evaluation to direction: scheduling, pay and the allocation of work, where M1 Direction carries M3 Consequence and M5 Care inside it, and where the rosters and claims allocations of Nordvik Region and Hansa Mutual produce the first Management Decision Records of the two organisations.

Sources

1. Ajunwa, I. (2023). The Quantified Worker: Law and Technology in the Modern Workplace. Cambridge University Press. (143, Ajunwa, 2023)
2. 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)
3. Kluger, A. N. & DeNisi, A. (1996). The effects of feedback interventions on performance: A historical review, a meta-analysis, and a preliminary feedback intervention theory. Psychological Bulletin. (116, Kluger & DeNisi, 1996)
4. Campbell, D. T. (1979). Assessing the impact of planned social change. Evaluation and Program Planning. (110, Campbell, 1979)
5. Muller, J. Z. (2018). The Tyranny of Metrics. Princeton University Press. (123, Muller, 2018)
6. Bernstein, E. S. (2012). The Transparency Paradox: A Role for Privacy in Organizational Learning and Operational Control. Administrative Science Quarterly. (122, Bernstein, 2012)
7. Lecher, C. (2019). How Amazon automatically tracks and fires warehouse workers for 'productivity'. The Verge. (187, Lecher, 2019)
8. Soper, S. (2021). Fired by Bot at Amazon: 'It's You Against the Machine'. Bloomberg. (198, Soper, 2021)
9. McAloon, A. (2021). Xsolla lays off 150 after an algorithm ruled staff 'unengaged and unproductive'. Game Developer. (200, McAloon, 2021)
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. Article 29 Data Protection Working Party. (2018). Guidelines on Automated individual decision-making and Profiling for the purposes of Regulation 2016/679 (WP251rev.01). (61, Article 29 Data Protection Working Party, 2018)
12. Spataro, J. (2020). Our commitment to privacy in Microsoft Productivity Score. Microsoft 365 Blog. (189, Spataro, 2020)
13. High Court of Justice, Queen's Bench Division. (2019). Bates and others v Post Office Ltd (No 6) "Horizon Issues", [2019] EWHC 3408 (QB). (72, High Court of Justice & Queen's Bench Division, 2019)
14. Post Office Horizon IT Inquiry. (2025). Volume 1 of the Post Office Horizon IT Inquiry's final report. (203, Post Office Horizon IT Inquiry, 2025)
15. Elish, M. C. (2019). Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction. Engaging Science, Technology, and Society. (170, Elish, 2019)
16. 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)
17. Court of Justice of the European Union. (2023). Judgment of 7 December 2023, OQ v Land Hessen (SCHUFA Holding (Scoring)), Case C-634/21. (63, CJEU, 2023)
18. 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)
19. 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)
20. Federal Republic of Germany. (1969). Kündigungsschutzgesetz (KSchG). Gesetze im Internet. (217, Germany, 1969)
21. Federal Republic of Germany. (2001). Betriebsverfassungsgesetz (BetrVG). Gesetze im Internet. (9, Germany, 2001)
22. Sveriges riksdag. (1982). Lag (1982:80) om anställningsskydd. (218, Sveriges riksdag, 1982)
23. Green, B. (2022). The flaws of policies requiring human oversight of government algorithms. Computer Law & Security Review. (178, Green, 2022)
24. Nordqvist, G. (2026). AI Governance: Governing Artificial Intelligence in European Enterprises and Public Authorities. (41, Nordqvist, 2026)
25. Nordqvist, G. (2026). AI Red Team: AI as an Active Part of Cyber Defence in European Enterprises and Public Authorities. (43, Nordqvist, 2026)
26. Nordqvist, G. (2026). The Transformed Security Architect: Role, Craft and Career in the Age of AI in European Enterprises and Public Authorities. (46, Nordqvist, 2026)
27. Nordqvist, G. (2026). AI Threat: Are the Dystopias Coming True? Evidence, Risk and Response for European Enterprises and Public Authorities. (44, Nordqvist, 2026)
28. Nordqvist, G. (2026). AI Platform and Security: Building and Securing AI Platforms in European Enterprises and Public Authorities. (42, Nordqvist, 2026)
29. Nordqvist, G. (2025). Architecture as Code. (40, Nordqvist, 2025)