Civil Law and Algorithmic Labor Cost Optimization Discrimination Claims in Europe

1. Introduction

Algorithmic labour-cost optimisation refers to the use of software, AI, machine learning, predictive analytics, or automated management systems to reduce or control an employer's labour costs.

Typical applications include:

• automated workforce scheduling;
• shift allocation;
• overtime optimisation;
• automated wage-setting;
• performance scoring;
• productivity targets;
• workforce reduction;
• employee ranking;
• promotion decisions;
• dismissal recommendations;
• recruitment filtering;
• absence management;
• allocation of undesirable shifts;
• platform-worker pricing;
• algorithmic allocation of work;
• monitoring of employee productivity.

The legal problem arises when a system designed to reduce labour costs produces a discriminatory effect.

For example:

Employer introduces an AI scheduling system to reduce overtime expenditure → algorithm systematically gives fewer hours to older workers or workers with disabilities → affected employees lose income and career opportunities.

The legal issue is not simply whether the algorithm saved money. The questions are:

1. Was the optimisation system lawful?
2. Did it use protected characteristics or discriminatory proxies?
3. Did it produce direct or indirect discrimination?
4. Was there meaningful human oversight?
5. Was personal data processed lawfully?
6. Was the worker adequately informed?
7. Can the employer justify the measure?
8. What evidence proves the algorithm's effect?
9. Did the discrimination cause economic or non-economic damage?

EU employment-law case law already provides substantial principles for analysing these issues, although there is still a relatively limited body of judgments directly concerning AI-driven labour-cost optimisation. A 2025 comparative study identified a growing European body of cases involving algorithmic management, labour law and privacy, but many disputes remain about conventional automated-management systems rather than sophisticated generative AI.

2. What Is Algorithmic Labour-Cost Optimisation?

An employer may give an AI system an objective such as:

Minimise total labour cost while maintaining required productivity.

The system might then calculate:

• employee availability;
• hourly wage;
• overtime cost;
• predicted productivity;
• absence probability;
• turnover probability;
• demand forecasts;
• employee performance;
• availability for night/weekend work.

The algorithm then recommends:

“Employee A should receive 20 hours; Employee B should receive 35 hours.”

Although apparently neutral, the variables may correlate with protected characteristics.

3. How Discrimination Can Enter the Algorithm

A. Direct discrimination

The system explicitly uses:

• age;
• sex;
• disability;
• racial or ethnic origin;
• religion;
• sexual orientation.

Example:

“Reduce hours for employees over 55 because their hourly cost is higher.”

This would create an obvious age-discrimination issue.

B. Indirect discrimination

More difficult cases involve apparently neutral criteria.

For example:

“Give the most profitable shifts to employees who can work continuously for eight hours.”

That criterion might disproportionately disadvantage workers with disabilities or workers with certain caring responsibilities.

Indirect discrimination is particularly important because an AI system can discriminate without ever being programmed to consider a protected characteristic.

4. Proxy Discrimination

AI systems can use variables that function as proxies.

Examples:

The employer may therefore argue:

“The model never used age.”

But that does not necessarily resolve an indirect-discrimination claim.

5. Main European Legal Framework

A. Equal Treatment in Employment

Directive 2000/78/EC establishes an EU framework concerning discrimination on grounds including:

• religion or belief;
• disability;
• age;
• sexual orientation.

Directive 2000/43/EC addresses racial and ethnic discrimination.

National equality laws implement and sometimes expand these protections.

6. EU AI Act

The EU AI Act is increasingly important for workplace AI.

Employment and worker-management systems can constitute high-risk AI systems, including systems used for:

• recruitment;
• selection;
• decisions affecting work relationships;
• promotion;
• termination;
• task allocation;
• monitoring;
• evaluation of workers.

The reason is that employment decisions can affect workers' livelihoods, careers and fundamental rights.

The AI Act also recognises the risk that employment AI may reproduce historical discrimination.

Important timing point: AI Act obligations have staggered application dates, and the precise application date of the employment-related high-risk provisions must be checked against the relevant deployment date and provision. Current EU developments have also involved changes to the timetable for some high-risk obligations.

7. GDPR Article 22

GDPR Article 22 protects individuals against certain decisions based solely on automated processing, including profiling, where the decision produces legal effects or similarly significant effects.

Employment is particularly important because Recital 71 expressly identifies evaluation of matters such as a person's performance at work as an example of profiling.

The CJEU's interpretation of Article 22 has become increasingly important.

In SCHUFA, the Court explained that automated scoring can itself constitute a decision where another party strongly relies on the score in determining the person's position.

8. Case Law

Case 1 — SCHUFA Holding, C-634/21

Court: CJEU
Judgment: 7 December 2023

This is one of the most important modern authorities for algorithmic employment disputes.

The CJEU held that automated establishment of a probability value can constitute automated individual decision-making where another party strongly relies on that score to establish, implement or terminate a relationship.

The Court also stressed safeguards including:

• appropriate mathematical/statistical procedures;
• minimisation of errors;
• correction of inaccuracies;
• protection against discriminatory effects;
• human intervention;
• opportunity to express one's point of view;
• ability to challenge the decision. 

Application to labour-cost optimisation

Suppose an employer's system produces:

“Employee cost-efficiency score = 42/100.”

The employer then uses the score to determine:

• fewer shifts;
• lower bonus;
• promotion denial;
• termination;
• reduced working hours.

The employer cannot necessarily avoid Article 22 simply by saying:

“The AI only generated a score; management made the final decision.”

The actual degree of reliance on the score matters.

Principle

An apparently preliminary algorithmic score can become legally significant where it effectively determines the employment outcome.

Relevance: Highly important analogy; SCHUFA itself was not an employment dispute.

9. Case 2 — Dun & Bradstreet Austria, C-203/22

Court: CJEU
Judgment: 27 February 2025

The CJEU addressed the meaning of “meaningful information about the logic involved” in automated decision-making.

The Court held that the explanation should describe the procedure and principles actually applied in a way that is:

• relevant;
• concise;
• transparent;
• intelligible;
• accessible.

A complex mathematical formula or disclosure of the entire algorithm is not necessarily required.

Employment application

An employee could potentially ask:

“Why did the algorithm reduce my shifts?”

A meaningful explanation might need to identify relevant factors such as:

• availability;
• productivity score;
• overtime history;
• predicted demand;
• absence data;
• performance indicators.

Simply saying:

“The algorithm determined that your labour cost was inefficient”

may not provide meaningful transparency.

Principle

Trade-secret protection does not automatically eliminate meaningful explanation.

10. Case 3 — Feryn, C-54/07

Court: CJEU
Judgment: 10 July 2008

Feryn concerned public statements by an employer indicating that it did not want to recruit persons of a particular ethnic origin.

The CJEU held that such statements can constitute direct discrimination in recruitment even without an identifiable individual complainant.

Algorithmic application

Suppose an employer's AI recruitment system is deliberately configured to:

reduce recruitment of candidates from a particular ethnic group because management believes customers prefer other workers.

The fact that the discriminatory policy operates through software rather than a manager's verbal statement would not make the underlying employment policy neutral.

Principle

Technology does not sanitise discriminatory intent.

11. Case 4 — Asociația Accept, C-81/12

Court: CJEU
Judgment: 25 April 2013

Accept concerned public statements suggesting that a football club would not recruit a homosexual player.

The CJEU considered the burden of proof under employment-equality law.

Where facts establish a presumption of discrimination, the burden can shift to the defendant to demonstrate that there was no breach of equal-treatment requirements.

Algorithmic application

Imagine an employer's workforce data reveal:

• 60% of workers are women;
• 40% are men;
• AI scheduling assigns undesirable shifts disproportionately to women;
• the employer refuses to disclose how the model makes the allocation.

Statistical and documentary evidence may become important in establishing facts from which discrimination can be presumed.

The employer may then have to provide a lawful explanation under the applicable equality rules.

Principle

A claimant does not necessarily need to obtain the source code before discrimination can be established.

12. Case 5 — CHEZ Razpredelenie Bulgaria, C-83/14

Court: CJEU Grand Chamber
Judgment: 16 July 2015

CHEZ concerned electricity meters installed much higher in predominantly Roma neighbourhoods.

The CJEU examined:

• direct discrimination;
• indirect discrimination;
• apparently neutral measures;
• statistical/contextual disadvantage;
• justification;
• proportionality;
• stigmatizing effects.

Algorithmic application

This is particularly useful for AI because a labour-cost algorithm may use an apparently neutral variable.

Example:

“Employees who have unpredictable availability receive fewer premium shifts.”

The employer might say:

“Availability has nothing to do with gender.”

But if the criterion disproportionately disadvantages a protected group, the court may need to examine:

• actual impact;
• objective justification;
• necessity;
• less discriminatory alternatives.

Principle

A neutral-looking criterion can still produce indirect discrimination.

13. Case 6 — HK Danmark v Experian, C-476/11

Court: CJEU
Judgment: 26 September 2013

The case concerned age discrimination and an occupational pension arrangement involving different contribution structures according to age.

The CJEU examined the justification of age-related differences under Directive 2000/78/EC.

Algorithmic application

Suppose a labour-cost system calculates:

“Older workers cost more because their wage and pension costs are higher.”

It then automatically:

• allocates fewer hours;
• excludes older employees from training;
• limits promotion;
• prioritises younger workers for recruitment.

The employer cannot simply say:

“The system is economically efficient.”

Age-related differentiation requires examination under the applicable equality framework and any relevant justification requirements.

Principle

Economic cost considerations do not automatically justify age-based employment differences.

14. Case 7 — VL v Szpital Kliniczny, C-16/19

Court: CJEU
Judgment: 26 January 2021

The case involved different treatment within a group of workers with disabilities concerning a salary allowance.

The CJEU held that a measure concerning workers with disabilities could constitute direct or indirect discrimination depending upon the circumstances.

Algorithmic application

Imagine a productivity system determines:

“Employees who have higher expected productivity should receive performance bonuses.”

If disability-related factors systematically reduce the measured productivity of disabled employees, the algorithm may produce discriminatory effects.

The employer would need to examine whether:

• the measurement method is appropriate;
• disability-related limitations are being treated fairly;
• reasonable accommodation is required;
• less discriminatory alternatives exist.

Principle

A system can discriminate even within a protected group; the relevant comparator and actual disadvantage matter.

15. Case 8 — Coleman v Attridge Law, C-303/06

Court: CJEU
Judgment: 17 July 2008

Coleman established that disability discrimination under EU employment law can extend to discrimination by association.

A worker can in certain circumstances receive protection even where the worker does not personally have the disability but is disadvantaged because of association with a disabled person.

Algorithmic relevance

An algorithm could potentially penalise employees because of:

• caring responsibilities;
• dependent-care patterns;
• absence linked to caring;
• family-related scheduling.

The system might therefore create discriminatory consequences even when it does not classify the employee as disabled.

Principle

Algorithmic employment systems must be assessed according to the actual legal discrimination rules, not merely the labels assigned to data fields.

16. Direct and Indirect Algorithmic Discrimination

Direct discrimination

Example:

AI instruction: “Prioritise employees under 35 for overtime because they are cheaper.”

The protected characteristic—age—is directly used.

Indirect discrimination

Example:

AI instruction: “Prioritise workers who can accept last-minute shifts.”

The criterion is apparently neutral.

But it may disproportionately disadvantage:

• workers with disabilities;
• workers with caring responsibilities;
• certain religious groups;
• workers with particular family commitments.

The legal analysis then moves toward objective justification and proportionality.

17. Labour-Cost Optimisation and Wage Discrimination

Algorithmic systems may influence:

• starting salary;
• hourly rates;
• bonuses;
• commissions;
• overtime;
• premium shifts;
• tips;
• performance pay.

This can create potential equal-pay and discrimination claims.

The system may appear neutral:

“Pay = predicted productivity.”

But if the underlying model systematically undervalues the work of one protected group, the resulting pay disparity can become legally significant.

18. Algorithmic Scheduling

Scheduling is one of the most realistic forms of labour-cost optimisation.

The employer may instruct AI:

“Minimise labour expenditure while meeting demand.”

The algorithm might then allocate:

• fewer hours to certain workers;
• night shifts to certain employees;
• unpredictable shifts;
• split shifts;
• weekend work.

Potential legal issues

• indirect discrimination;
• disability discrimination;
• pregnancy/maternity discrimination;
• religious discrimination;
• age discrimination;
• working-time law;
• collective bargaining;
• contractual rights;
• privacy/data protection.

The CHEZ proportionality analysis is particularly useful where a neutral scheduling criterion has disparate effects.

19. Algorithmic Performance Scoring

An employer may use:

AI Productivity Score = 72/100

to determine:

• bonus;
• promotion;
• working hours;
• disciplinary action;
• dismissal.

This creates two separate legal problems.

Equality problem

Does the score disadvantage a protected group?

Data-protection problem

Is the score based on lawful, accurate and appropriately processed personal data?

SCHUFA is relevant because the CJEU has rejected an overly narrow conception of what constitutes an automated decision where an automated score substantially determines another person's decision.

20. Automated Dismissal

The most serious scenario is:

AI determines that an employee is economically inefficient → employment terminated.

Potential claims may involve:

• unfair dismissal;
• discrimination;
• GDPR Article 22;
• lack of human intervention;
• inaccurate data;
• contractual breach;
• collective consultation;
• national labour law.

The employer should not assume:

“A manager clicked approve.”

automatically converts an automated process into genuine human decision-making.

The quality and independence of human intervention matter.

21. Automation Bias

Human involvement can sometimes be only formal.

Example:

AI: “Dismiss Employee X.”

Manager:

“Approved.”

If the manager:

• does not inspect the data;
• does not understand the model;
• cannot challenge the result;
• never considers alternatives,

the supposed human review may be legally questionable depending upon the applicable framework.

SCHUFA expressly highlights the importance of meaningful human intervention in qualifying automated decision-making.

22. Historical Bias

AI often learns from historical employment data.

Suppose historically:

• men received more overtime;
• women took more career breaks;
• disabled workers received lower productivity ratings.

The algorithm learns:

historical pattern = predicted productivity.

It may then reproduce the historical inequality.

This produces the classic problem:

Past discrimination → training data → algorithmic prediction → future discrimination.

The EU AI Act's employment-risk framework expressly recognises the possibility that AI systems may perpetuate historical discrimination.

23. Proxy Variables

An employer may claim:

“We removed gender from the dataset.”

But the model may still use:

• job history;
• employment gaps;
• working pattern;
• postcode;
• absence;
• previous salary.

These can function as proxies.

Therefore, removing the protected characteristic itself does not necessarily eliminate discrimination.

24. Employer's Business-Justification Defence

An employer may argue:

“The algorithm is necessary to reduce labour costs.”

This is not automatically sufficient.

For indirect discrimination, the court generally needs to consider questions such as:

1. Is the employer pursuing a legitimate objective?
2. Is the algorithm suitable for achieving it?
3. Is it necessary?
4. Is there a less discriminatory alternative?
5. Is the disadvantage proportionate?

CHEZ is particularly useful for the proportionality analysis of apparently neutral measures.

25. “Cheaper Worker” Is Not a Complete Legal Category

A labour-cost model might rank employees according to:

Wage cost ÷ predicted productivity.

But this can create discrimination if the model systematically disadvantages:

• older workers;
• disabled workers;
• women;
• ethnic minorities;
• pregnant workers;
• workers with caring responsibilities.

An economically efficient outcome does not automatically become a legally permissible outcome.

26. GDPR Data-Protection Issues

Algorithmic labour optimisation commonly uses personal data.

Possible data include:

• attendance;
• productivity;
• location;
• communications;
• keystrokes;
• work speed;
• health-related absence;
• performance;
• working patterns.

Relevant GDPR principles include:

Lawfulness

There must be a valid legal basis.

Purpose limitation

Data should not simply be reused for unrelated purposes.

Data minimisation

Only appropriate data should be used.

Accuracy

Incorrect data can produce discriminatory results.

Transparency

Workers should receive appropriate information.

Accountability

The controller must be able to demonstrate compliance.

SCHUFA emphasised that automated profiling must satisfy GDPR principles concerning lawfulness, accuracy and safeguards against discriminatory effects.

27. Meaningful Explanation

Under Dun & Bradstreet, meaningful information concerns the actual procedure and principles applied—not merely a mathematical formula.

For labour optimisation, a meaningful explanation could involve:

“The scheduling model reduced your hours because it assigned a lower availability score based on X, Y and Z.”

The employer does not necessarily have to disclose:

complete source code + proprietary model architecture.

The legal objective is meaningful understanding and challenge.

28. Evidence in Algorithmic Discrimination Claims

Algorithmic discrimination cases are heavily evidence-dependent.

Important evidence includes:

Model documentation

• model design;
• variables;
• objectives;
• weighting;
• thresholds.

Employment records

• hours;
• wages;
• bonuses;
• promotions;
• dismissals.

Statistical evidence

• group comparison;
• selection rates;
• pay differences;
• scheduling differences;
• error rates.

Technical records

• model version;
• logs;
• system outputs;
• updates;
• overrides.

Governance records

• DPIA;
• AI impact assessment;
• equality assessment;
• bias testing;
• vendor documentation;
• internal complaints.

29. Statistical Evidence

Suppose an employer has:

The difference alone does not prove discrimination.

The court may need to examine:

• job roles;
• availability;
• seniority;
• qualifications;
• working patterns;
• productivity;
• contractual terms;
• legitimate business requirements.

But substantial statistical disparities can contribute to a prima facie discrimination case, particularly when combined with evidence about the algorithm's design.

30. Burden of Proof

Feryn and Accept are particularly useful here.

Feryn demonstrates that direct discrimination can exist even without a single identified victim in the recruitment context.

Accept demonstrates that once facts establish a presumption of discrimination, the evidentiary burden can shift to the defendant under the relevant equality framework.

Algorithmic significance

The claimant may therefore rely upon:

• statistical evidence;
• algorithm outputs;
• employee comparisons;
• internal documents;
• discriminatory instructions;
• performance patterns.

The claimant does not necessarily need to prove the complete technical architecture of the AI model at the outset.

31. Vendor Liability

Suppose:

Company A purchases workforce AI from Company B.

The AI produces discriminatory results.

There can be several separate questions.

Employer

Potential responsibility for:

• deployment;
• employment decision;
• supervision;
• discrimination;
• data processing.

Vendor

Potential responsibility depending upon:

• contract;
• defective software;
• warranties;
• negligent development;
• statutory obligations;
• contribution/indemnity.

The employer generally cannot simply say:

“The vendor's AI discriminated, so we have no responsibility.”

32. Collective Claims

Algorithmic discrimination can affect hundreds or thousands of employees.

Potential mechanisms may include:

• trade-union claims;
• collective litigation;
• representative actions;
• equality-body proceedings;
• labour-inspectorate intervention;
• data-protection complaints.

The exact mechanism depends upon the Member State.

33. Worker Privacy and Labour-Cost Optimisation

A cost-optimisation system may monitor workers continuously.

For example:

AI records keystrokes → calculates productivity → adjusts future shifts.

This raises separate privacy concerns.

Relevant European cases include:

Bărbulescu v Romania

ECtHR Grand Chamber, 5 September 2017.

The Court examined workplace electronic communications monitoring and required a proper balancing of employee privacy and employer interests.

López Ribalda and Others v Spain

ECtHR Grand Chamber, 17 October 2019.

The Court examined covert workplace video surveillance and proportionality.

These cases are not discrimination cases, but they are important when the labour-cost system relies upon intensive worker surveillance.

34. Discrimination + Privacy Can Overlap

One system may create multiple legal claims.

Example:

Employer monitors employees' health-related absences → AI predicts productivity → employees with disabilities receive fewer shifts.

Potential claims:

1. disability discrimination;
2. unlawful processing of health data;
3. lack of transparency;
4. automated decision-making;
5. workplace privacy;
6. contractual/labour-law breach;
7. compensation.

Thus, an algorithmic labour dispute should not automatically be analysed as only an equality-law case.

35. Remedies

Potential remedies vary by Member State and applicable EU law.

They may include:

1. Compensation

For financial loss and, where permitted, non-material harm.

2. Back pay

Where discriminatory scheduling or wage-setting caused underpayment.

3. Reinstatement

Where discriminatory termination is established and national law provides for it.

4. Reconsideration

The employer may need to repeat the decision through a lawful process.

5. Correction of data

Incorrect personal information can be corrected.

6. Injunction

Future discriminatory algorithmic processing may be restricted.

7. Non-discrimination order

The employer may have to change the relevant practice.

8. Regulatory sanctions

Data-protection or AI/equality regulators may impose separate consequences.

36. Damages and Causation

A worker claiming damages must generally establish the relevant causal connection.

Example:

AI scheduling system → fewer premium shifts → reduced monthly income.

The worker may calculate:

Normal expected earnings − actual earnings = economic loss

But the employer may argue that:

• the employee was unavailable;
• demand declined;
• another worker was more qualified;
• the algorithm's recommendation was not followed;
• other factors caused the reduced hours.

Causation therefore remains a factual and evidentiary issue.

37. AI-Created Wage Inequality

Consider an algorithm that recommends salaries based upon predicted employee retention.

Historical data show that women accepted lower salaries more frequently.

The AI therefore predicts:

“Female candidate likely to accept €35,000.”

while predicting:

“Male candidate likely to require €40,000.”

If the employer uses those predictions to set actual salaries, historical inequality may be transformed into future pay inequality.

This raises potential questions concerning:

• equal pay;
• sex discrimination;
• indirect discrimination;
• data protection;
• AI governance.

38. Algorithmic Workforce Reduction

A company may instruct AI:

“Reduce labour expenditure by 15%.”

The system selects employees for redundancy based on:

• salary;
• performance;
• absence;
• predicted productivity;
• replacement cost.

This can create a serious discrimination risk.

For example:

Older workers have higher salaries → algorithm selects them disproportionately.

The employer may argue:

“The system only selected the most expensive positions.”

But if age and cost are strongly correlated, the legal analysis may require examination of whether the selection method produces unjustified age discrimination.

39. Reasonable Accommodation and Disability

An algorithm may classify a disabled worker as:

“lower productivity.”

But the worker's productivity may be measured without considering reasonable accommodation.

For example:

AI measures keyboard activity rather than completed work.

A worker using assistive technology might appear “slow” under the model even though actual work output is satisfactory.

VL demonstrates the importance of examining how employment practices affect disabled workers and whether a difference in treatment falls within the Directive's discrimination concepts.

40. Religious Discrimination

Suppose an algorithm is instructed:

“Allocate high-demand shifts to workers with the greatest availability.”

If the system repeatedly assigns certain religious workers to shifts conflicting with religious observance, the employer may need to consider the applicable religious-discrimination and accommodation rules.

The fact that:

“The algorithm simply maximised availability”

does not automatically resolve the equality issue.

41. Age Discrimination

Labour-cost optimisation is especially likely to produce age-related problems where salary increases with seniority.

Example:

Worker A: 25 years old — €25/hour
Worker B: 58 years old — €40/hour

An AI tasked with reducing payroll may conclude:

“Replace Worker B.”

If the system systematically uses age-correlated variables to reduce the workforce, the employer may face age-discrimination questions.

HK Danmark provides an important framework for assessing age-related differences and justification.

42. The Importance of Proportionality

A useful analytical structure is:

Legitimate objective

Reducing unnecessary labour costs may generally constitute a legitimate business objective.

Suitability

Does the algorithm actually help achieve that objective?

Necessity

Is the discriminatory impact avoidable through a less discriminatory method?

Balancing

Does the business benefit justify the disadvantage imposed on the affected workers?

The CHEZ judgment provides a useful illustration of this proportionality reasoning for apparently neutral measures with discriminatory effects.

43. Key Defences for Employers

An employer may argue:

1. No protected characteristic was used

The model did not explicitly use age, sex, ethnicity, etc.

2. Legitimate business objective

The purpose was workforce efficiency.

3. Objective criteria

The algorithm used measurable productivity and availability.

4. No significant automated decision

A human made the final decision.

5. No disparate effect

Statistical evidence does not establish meaningful disadvantage.

6. Objective justification

Any indirect disadvantage was necessary and proportionate.

7. Data accuracy

The underlying information was accurate.

8. Alternative cause

The employee's loss resulted from business conditions rather than the algorithm.

44. Strongest Risk Indicators

An algorithmic labour-cost system becomes particularly legally sensitive where it:

• automatically determines pay;
• automatically determines working hours;
• automatically selects employees for dismissal;
• uses health information;
• uses disability-related information;
• predicts employee behaviour;
• relies heavily on historical discriminatory data;
• uses protected characteristics;
• uses obvious proxies;
• produces unexplained rankings;
• cannot be meaningfully challenged;
• has no independent human review;
• systematically disadvantages one protected group.

45. Practical Litigation Framework

A claimant can analyse a case through the following sequence.

Step 1 — Identify the employment decision

Was it:

• recruitment;
• pay;
• scheduling;
• promotion;
• performance assessment;
• dismissal?

Step 2 — Identify the algorithm

What did it actually do?

Step 3 — Identify the data

What information was supplied?

Step 4 — Identify the protected characteristic

Age? Disability? Sex? Race? Religion? Sexual orientation?

Step 5 — Establish the disparity

Who suffered the disadvantage?

Step 6 — Establish causation

Did the algorithm cause or materially contribute to the outcome?

Step 7 — Examine justification

Was the criterion legitimate, necessary and proportionate?

Step 8 — Examine GDPR

Was personal data lawfully processed?

Step 9 — Examine Article 22

Was the decision solely automated and significantly affecting the worker?

Step 10 — Examine AI Act

Was the system a regulated employment/workforce-management AI system, and which provisions applied at the relevant date?

Step 11 — Determine remedy

Compensation? Back pay? Reconsideration? Injunction? Regulatory complaint?

46. Case-Law Summary

47. Important Distinction: Cost Optimisation Is Not Automatically Discrimination

An employer is generally not prohibited merely because it uses AI to reduce labour costs.

The legally important question is how the cost objective is implemented.

Lawful possibility

AI predicts customer demand → recommends number of workers required → manager allocates shifts using objective and non-discriminatory criteria.

Higher-risk possibility

AI predicts which workers are “too expensive” → systematically reduces hours for older workers → no meaningful review.

The second scenario raises substantially different equality and data-protection questions.

48. Important Distinction: Algorithmic Error vs Discrimination

These are also different.

Algorithmic error

The system incorrectly predicts productivity.

Discrimination

The system disproportionately disadvantages a protected group.

An algorithm can be:

• accurate but discriminatory;
• inaccurate but non-discriminatory;
• both inaccurate and discriminatory;
• neither.

Therefore, technical accuracy does not by itself establish equality-law compliance.

49. Simple Example

Suppose a supermarket introduces an AI scheduling system.

Its objective is:

Reduce payroll by 10%.

The system uses:

• hourly wage;
• predicted sales productivity;
• availability;
• absence history;
• customer ratings.

After six months:

• older workers receive fewer hours;
• disabled workers receive fewer premium shifts;
• women receive fewer late-evening shifts.

The employer says:

“The algorithm never used age, disability or gender.”

A court could nevertheless need to examine:

AI variables

↓

disparate effects

↓

protected group disadvantage

↓

indirect discrimination

↓

objective justification

↓

proportionality

↓

possible compensation/remedial measures

At the same time, if the system makes significant automated decisions based on personal data, GDPR safeguards may independently become relevant.

50. Key Legal Principles

1. An AI system does not become legally neutral merely because it is mathematical.
2. Labour-cost optimisation is not itself unlawful.
3. Direct discrimination can arise where protected characteristics are deliberately used.
4. Indirect discrimination can arise from apparently neutral optimisation criteria.
5. Proxy variables can create discrimination even when protected characteristics are removed.
6. SCHUFA is important where an algorithmic score effectively determines an employment outcome. 
7. Dun & Bradstreet strengthens the importance of meaningful explanations for automated decisions. 
8. Feryn demonstrates that discriminatory recruitment policies can be legally significant even without an identifiable individual complainant. 
9. Accept demonstrates the importance of burden-of-proof rules in employment discrimination. 
10. CHEZ demonstrates that neutral criteria can produce indirect discrimination and must be tested for proportionality. 
11. HK Danmark is relevant when labour-cost systems create age-based differences. 
12. VL is important for disability-related employment discrimination. 
13. GDPR requires attention to accuracy, lawfulness, transparency and accountability.
14. Article 22 does not prohibit every algorithmic workplace tool; its applicability depends on the conditions concerning solely automated decisions and significant effects, together with applicable exceptions and safeguards. 
15. Human review should be meaningful rather than merely formal.
16. Economic efficiency does not automatically justify discriminatory treatment.
17. Statistical evidence can be important in proving disparate effects.
18. Employer and AI-vendor liability are separate legal questions.
19. Algorithmic discrimination can coexist with privacy, contract and labour-law claims.
20. The AI Act adds an important employment-AI governance layer, but the exact obligation and application date must be checked for the relevant system and date.

51. Final Conclusion

Algorithmic labour-cost optimisation liability in Europe sits at the intersection of employment discrimination law, GDPR, the EU AI Act, contract law, labour law and civil remedies.

The central legal formula is:

Cost-Optimisation Objective + Algorithmic Employment Decision + Protected-Group Disadvantage + Causal Connection + No Adequate Justification = Potential Discrimination Liability

A second, independent formula applies to automated decision-making:

Personal Data + Significant Automated Employment Decision + Insufficient Safeguards/Explanation/Human Review = Potential GDPR Liability

The most important authorities for building a European legal analysis are SCHUFA, Dun & Bradstreet, Feryn, Asociația Accept, CHEZ, HK Danmark, VL and Coleman. The first two provide particularly important modern principles concerning automated scoring and explanation, while the equality cases supply the established legal framework for analysing direct discrimination, indirect discrimination, burden of proof, age and disability. The available European case law therefore supports a detailed accountability framework, but there is still no single CJEU judgment establishing a general civil cause of action specifically for AI systems that optimise labour costs and thereby discriminate against workers.