Pay and Scheduling Fairness Governance

2. Summary

Pay and Scheduling Fairness Governance requires that any AI agent involved in determining, recommending, adjusting, or materially influencing employee compensation, shift scheduling, workload distribution, or overtime allocation implements enforceable fairness constraints that prevent unlawful bias along protected characteristic lines. Automated pay and scheduling systems present a dual fairness risk: pay algorithms can entrench and amplify historical gender and racial pay gaps by learning from biased compensation histories, while scheduling algorithms can produce discriminatory shift patterns that disproportionately burden employees with caregiving responsibilities, religious observance requirements, or disability-related scheduling needs. This dimension mandates pre-deployment equity audits, continuous pay gap and scheduling disparity monitoring, jurisdictional pay transparency compliance, and documented remediation procedures — ensuring that AI-driven compensation and scheduling decisions comply with equal pay legislation, anti-discrimination law, and emerging algorithmic scheduling regulations across all operating jurisdictions.

3. Example

Scenario A — Pay Algorithm Perpetuates Gender Pay Gap Through Historical Anchoring: A retail financial services company with 6,300 employees deploys an AI agent to recommend annual salary adjustments based on performance scores, market benchmarks, role tenure, and current salary. The model uses current salary as a primary anchor — employees with higher current salaries receive proportionally larger absolute increases. Because historical gender pay gaps mean that women in equivalent roles earn on average 8.2% less than men, the anchoring effect perpetuates and widens the gap: after two annual cycles, the gender pay gap increases from 8.2% to 11.7%. The model recommends £4.1 million in total salary increases, of which women receive £1.6 million and men receive £2.5 million — despite women comprising 49% of the workforce. When the company publishes its mandatory gender pay gap report, the widening gap triggers regulatory scrutiny. The Equality and Human Rights Commission opens a Section 20 investigation. The company spends £890,000 on an independent pay equity audit, £1.2 million on retrospective pay corrections for 1,400 female employees, and £340,000 in legal and advisory fees.

What went wrong: The model used current salary as an input feature without recognising that current salary encodes historical gender pay disparities. No pay equity audit was conducted before deployment. The model's recommendation cycle amplified the gap because percentage-based increases applied to an unequal base produce widening absolute gaps. Continuous monitoring did not track the gender pay gap trajectory.

Scenario B — Scheduling Algorithm Creates Religious Discrimination Through Availability Optimisation: A logistics and warehousing company with 2,800 employees implements an AI scheduling agent to optimise shift coverage across 12 distribution centres. The model maximises coverage efficiency by assigning shifts based on historical availability patterns, skill match, and location proximity. Employees who have repeatedly declined Friday evening and Saturday shifts — predominantly observant Jewish and Muslim employees adjusting for Shabbat and Jumu'ah — are deprioritised for desirable weekday shifts because the model interprets their availability restrictions as lower flexibility. Over six months, 34 employees with religious scheduling needs are assigned 23% more late-night and Sunday shifts than the average — the least desirable shifts with the lowest premium pay. Eleven employees file grievances. Three resign and bring constructive dismissal claims citing religious discrimination. Tribunal awards total £187,000. The company also receives an enforcement notice under the Equality Act 2010 requiring a review of all scheduling practices.

What went wrong: The scheduling model treated religious availability patterns as a flexibility metric without recognising that religious observance is a protected characteristic requiring reasonable accommodation. No protected-characteristic impact analysis was conducted on shift distribution patterns. The model optimised for operational efficiency without fairness constraints, producing a scheduling pattern that systematically disadvantaged employees exercising their right to religious observance.

Scenario C — Workload Algorithm Creates Disability Disparate Impact Through Throughput Normalisation: A customer service operation with 1,900 agents uses an AI system to distribute incoming cases to agents based on historical throughput — agents who resolve cases faster receive more cases. Agents with disabilities that affect processing speed — including dyslexia, chronic pain conditions requiring more frequent breaks, and visual impairments requiring assistive technology — have lower historical throughput. The model assigns them fewer cases, which reduces their performance metrics, which in turn reduces their variable pay (tied to case volume and resolution rate). Over four quarters, agents with disclosed disabilities earn on average 17% less in variable compensation than agents without disabilities in the same role grade. The aggregate underpayment across 83 agents with disabilities totals £312,000 annually. When the company is audited for disability pay gap reporting (now mandatory in several jurisdictions), the algorithmic workload allocation is identified as the primary driver. The remediation costs £540,000 including retrospective compensation adjustments, system redesign, and legal advisory.

What went wrong: The workload distribution model used throughput as the primary allocation criterion without adjusting for disability-related factors. No reasonable adjustment mechanism existed in the algorithm for employees with disabilities. The consequential pay impact was not monitored because the company tracked pay equity by role grade but not by disability status. The model created a feedback loop: lower throughput led to fewer cases, which led to lower performance metrics, which led to lower pay — a compounding disadvantage driven by disability.

4. Requirement Statement

Scope: This dimension applies to any AI agent that determines, recommends, adjusts, distributes, or materially influences: (a) employee compensation including base salary, variable pay, bonuses, commission structures, overtime rates, shift premiums, or any other form of remuneration; (b) shift scheduling including shift assignment, shift swapping recommendations, on-call allocation, or availability-based scheduling; (c) workload distribution including case allocation, task assignment, project staffing, or any mechanism that determines the volume or type of work assigned to an individual and thereby affects their compensation or working conditions. The dimension applies regardless of whether the AI agent makes final decisions or provides recommendations to human decision-makers. If the human decision-maker adopts the AI recommendation without substantive modification in more than 50% of cases, the system is treated as a decision-maker for the purposes of this dimension. The scope extends to gig economy platforms, staffing agencies, and any organisation where algorithmic systems influence pay or scheduling for workers in an employment or quasi-employment relationship.

4.1. A conforming system MUST conduct a pre-deployment pay equity audit analysing compensation outcomes across all protected characteristics recognised in applicable jurisdiction(s), using both the four-fifths rule for favourable compensation outcomes and regression-based pay gap analysis controlling for legitimate pay-determining factors (role, tenure, location, performance, qualifications).

4.2. A conforming system MUST prohibit the use of current salary or salary history as an input feature for compensation recommendations unless a documented pay equity assessment confirms that current salary does not encode historical protected-characteristic pay gaps — and must re-confirm this assessment annually.

4.3. A conforming system MUST implement continuous pay disparity monitoring that analyses compensation outcomes by protected characteristic subgroup at intervals no greater than each compensation cycle or quarterly, whichever is more frequent, generating automated alerts when unexplained pay gaps exceed jurisdiction-specific thresholds or 3% (whichever is more stringent).

4.4. A conforming system MUST analyse shift and workload distribution patterns across protected characteristic subgroups at intervals no greater than monthly, detecting disproportionate assignment of undesirable shifts (defined by the organisation with reference to employee survey data or contractual shift premium structures), excessive workload concentration, or systematic under-allocation to specific subgroups.

4.5. A conforming system MUST implement a reasonable accommodation mechanism in scheduling and workload allocation that allows employees to register protected-characteristic-related scheduling needs (religious observance, disability adjustments, caregiving responsibilities protected under applicable law) and ensures that these needs are treated as hard constraints rather than preference inputs — the system must not penalise employees for exercising these accommodations in any downstream metric or allocation.

4.6. A conforming system MUST provide every affected employee with a plain-language explanation of how their compensation was determined or how their schedule was generated, including the factors considered, the relative influence of each factor, and — for compensation — how their pay compares to the median for their role, level, and location.

4.7. A conforming system MUST implement a contestation mechanism allowing any employee to challenge a pay decision or scheduling pattern, triggering a documented review by an independent reviewer completed within 20 business days, with the ability to order retrospective pay correction or schedule adjustment.

4.8. A conforming system MUST halt compensation or scheduling operations and trigger mandatory remediation when monitoring detects a statistically significant (p < 0.05) unexplained pay gap exceeding 3% for any protected characteristic subgroup, or when scheduling analysis reveals a four-fifths rule violation in desirable shift allocation for any protected characteristic subgroup.

4.9. A conforming system MUST maintain compliance with jurisdiction-specific pay transparency requirements, including but not limited to: EU Pay Transparency Directive (2023/970) requirements for pay reporting, gender pay gap disclosure, and information rights; UK gender pay gap reporting requirements; and US state-level salary range disclosure and pay equity laws applicable in each operating jurisdiction.

4.10. A conforming system SHOULD implement pay equity simulation — the ability to model the projected impact of proposed compensation changes on pay gaps across protected characteristics before the changes are applied, enabling pre-emptive correction.

4.11. A conforming system SHOULD implement scheduling preference learning that adapts to employee preferences over time while maintaining fairness constraints — ensuring that preference learning does not create a feedback loop where employees who historically received unfavourable schedules continue to receive them because their preferences adapted to what was available.

4.12. A conforming system MAY implement real-time fairness-constrained optimisation in scheduling that jointly optimises for operational coverage and demographic fairness, producing Pareto-optimal schedules that maximise coverage subject to equitable distribution constraints.

4.13. A conforming system MAY provide employees with a self-service tool to explore how different legitimate factors (e.g., acquiring a certification, changing location, taking on additional responsibilities) would affect their projected compensation.

5. Rationale

Pay and scheduling decisions are among the most consequential actions an employer takes. Compensation determines employees' economic security, their ability to meet financial obligations, and their long-term wealth accumulation. Scheduling determines employees' ability to manage caregiving responsibilities, maintain health, exercise religious observance, and participate in community life. When AI systems influence these decisions, the scale of impact is unprecedented: a single algorithmic change can affect thousands of employees simultaneously, and a bias encoded in the algorithm can compound across pay cycles, creating widening gaps that would take years to close even after detection.

The legal framework governing pay and scheduling fairness is among the most developed in employment law. The EU Pay Transparency Directive (2023/970), effective June 2026, requires employers to provide pay transparency, conduct joint pay assessments when gender pay gaps exceed 5%, and gives employees the right to information about pay levels for workers in the same category. The UK Equality Act 2010 includes an equal pay provision (Section 66) that implies an equality clause into every contract of employment. Title VII and the US Equal Pay Act prohibit compensation discrimination. Over 20 US states have enacted pay transparency laws restricting the use of salary history. The EU Working Time Directive and national implementations regulate scheduling practices. Emerging algorithmic scheduling laws (e.g., predictive scheduling ordinances in several US cities, proposed EU directive provisions on algorithmic management) impose specific requirements on automated scheduling systems.

The technical risk is well-documented. Compensation algorithms that use current salary as an input feature will perpetuate historical pay gaps — this is a mathematical certainty, not a probability. If women earn 8% less than men in equivalent roles due to historical discrimination, and the algorithm recommends percentage-based increases anchored to current salary, the gap persists indefinitely (and may widen due to compounding effects). Scheduling algorithms that optimise for operational efficiency without fairness constraints will systematically disadvantage employees whose availability patterns differ from the majority — and these differences are strongly correlated with protected characteristics including religion, disability, gender (through caregiving), and age.

The feedback loop risk is particularly acute in workload distribution. When an algorithm allocates fewer tasks to slower workers, and variable pay depends on task volume, the algorithm creates a pay disparity through workload allocation rather than through the pay calculation itself. This indirect path to pay discrimination is harder to detect because it does not appear in a traditional pay equity analysis — the per-task rate is equal; only the volume differs. Governance must therefore encompass not only direct compensation decisions but also workload allocation decisions that have consequential pay effects.

The organisational cost of failure is substantial and multi-dimensional. Direct financial costs include retrospective pay corrections (which can reach millions for large workforces over multiple cycles), legal costs (tribunal awards, settlement payments, regulatory fines), and remediation costs (system rebuild, independent audit, process redesign). Indirect costs include reputational damage affecting talent acquisition, reduced employee engagement and productivity, and increased regulatory scrutiny of all HR technology systems. The EU AI Act's classification of employment AI systems as high-risk means that non-compliance can trigger fines of up to 3% of annual worldwide turnover.

6. Implementation Guidance

Pay and scheduling fairness governance requires organisations to embed equity constraints at every point where an AI system touches compensation or scheduling — from feature selection through operational monitoring to individual remediation.

Recommended patterns:

• Salary history exclusion with documented exception process. Default to excluding current salary and salary history from all compensation recommendation models. If an organisation determines that salary history is necessary for a specific, documented purpose (e.g., ensuring offers do not represent a pay decrease for lateral hires), implement a gated exception process requiring legal review, pay equity impact analysis, and annual re-certification that the use of salary history does not encode protected-characteristic pay gaps. This approach aligns with the growing legal trend toward salary history bans and eliminates the most common vector for algorithmic pay gap perpetuation.
• Regression-based pay gap analysis pipeline. Implement an automated pipeline that runs regression analysis on compensation outcomes, controlling for legitimate pay-determining factors (role, level, tenure, location, performance, qualifications, relevant experience) and measuring the residual gap attributable to each protected characteristic. The pipeline should run on every compensation cycle and produce a standardised report showing: controlled pay gap by protected characteristic, statistical significance, comparison to prior cycles, and flagged subgroups requiring investigation. Use robust regression methods that handle outliers and non-linear relationships.
• Scheduling fairness constraint engine. Implement scheduling constraints as hard boundaries in the optimisation function, not as soft preferences that can be traded off against efficiency. Define fairness constraints including: maximum deviation in desirable shift allocation across protected characteristic subgroups (recommended: no subgroup receives more than 10% fewer desirable shifts than the population average), mandatory accommodation of registered protected-characteristic scheduling needs, and prohibition of penalty signals for accommodation use. The constraint engine should produce a fairness certificate for each generated schedule documenting constraint compliance.
• Workload-to-pay impact tracing. Implement tracing that connects workload allocation decisions to their compensation consequences. For any system where variable pay depends on work volume, monitor the pay impact of workload distribution by protected characteristic subgroup. If workload allocation produces a pay disparity (even if per-unit pay rates are equal), the disparity falls within scope of this dimension and triggers the same monitoring, alerting, and remediation requirements as direct compensation disparities.
• Multi-jurisdictional pay transparency compliance engine. For organisations operating across jurisdictions, implement a compliance engine that tracks pay transparency requirements per jurisdiction (pay range disclosure, pay gap reporting thresholds, employee information rights, joint pay assessment triggers) and generates compliance artefacts appropriate to each jurisdiction. As the EU Pay Transparency Directive is transposed into national law across 27 member states with potential local variations, maintaining jurisdiction-specific compliance manually becomes impractical.

Anti-patterns to avoid:

• Efficiency-only scheduling optimisation. Optimising schedules purely for coverage, cost, or throughput without fairness constraints. This approach treats employees as fungible resources and consistently produces schedules that disadvantage employees with protected-characteristic-related availability patterns. Fairness must be a constraint in the optimisation, not an afterthought.
• Aggregate pay gap analysis without subgroup decomposition. Monitoring the overall gender pay gap without analysing by role, level, and intersectional subgroups. An organisation-wide pay gap of 2% can conceal a 15% gap in a specific role/level combination. Analysis must be granular enough to detect localised disparities.
• Accommodation-as-preference. Treating protected-characteristic scheduling needs (religious observance, disability adjustments) as soft preferences that the scheduling algorithm can override when operational needs are high. Protected-characteristic accommodations must be hard constraints — the system must never schedule an employee during their registered religious observance period, regardless of coverage pressure.
• One-time pay equity audit without continuous monitoring. Conducting a pay equity audit at deployment and treating it as a permanent assurance of fairness. Pay algorithms interact with changing workforce composition, market conditions, and performance distributions. A model that is fair at deployment can develop disparate impact within months.
• Variable pay exclusion from pay equity analysis. Analysing base salary equity but excluding variable pay components (bonuses, commissions, shift premiums, overtime pay). Variable pay is often a larger proportion of total compensation for certain roles and is frequently the component most influenced by algorithmic decisions (through workload allocation, performance scoring, or shift assignment).

Industry Considerations

Retail and hospitality. Scheduling fairness is a dominant concern due to variable shift patterns, weekend and holiday scheduling, and workforces with high diversity across religion, age, gender, and disability. Predictive scheduling laws (enacted in several US cities and proposed in other jurisdictions) impose specific requirements: advance schedule notice, premium pay for schedule changes, and good-faith estimates of expected hours. AI scheduling systems must comply with these requirements while maintaining fairness across protected characteristics.

Financial services. Pay equity is subject to heightened scrutiny under FCA remuneration governance rules (SYSC 19A, 19D, 19G) and the Senior Managers and Certification Regime. Variable compensation often constitutes 30-60% of total remuneration for regulated roles. An AI system that produces biased variable compensation recommendations undermines the regulatory remuneration framework. Firms must demonstrate that AI-assisted compensation decisions do not create or amplify pay gaps across protected characteristics.

Gig economy and platform work. Algorithmic pay and scheduling decisions in gig economy platforms are subject to increasing regulation under the EU Platform Work Directive and national equivalents. The line between scheduling recommendations and scheduling mandates is often blurred — a platform that offers fewer assignments to certain workers is effectively scheduling them out of earnings. This dimension applies in full to platform work contexts where the algorithmic system materially influences pay or scheduling outcomes.

Public sector. Public sector pay is often governed by collective bargaining agreements and published pay scales, which reduces but does not eliminate the risk of algorithmic pay bias (discretionary elements such as starting point on scale, accelerated progression, and allowances remain vulnerable). Scheduling in public services (healthcare, emergency services, education) has direct service delivery implications — but fairness constraints must still apply. The public sector equality duty requires proactive consideration of how scheduling decisions affect equality of opportunity.

Maturity Model

Basic Implementation — The organisation has conducted a pre-deployment pay equity audit using regression analysis controlling for legitimate factors. Salary history is excluded from compensation models (or its use is justified and annually re-certified). Continuous monitoring tracks pay gaps and scheduling distributions by protected characteristic per compensation/scheduling cycle. Reasonable accommodation is implemented as hard constraints in scheduling. Employees receive explanations of pay and scheduling decisions. A contestation mechanism exists. Halt-and-remediate procedures are documented and functional.

Intermediate Implementation — All basic capabilities plus: workload-to-pay impact tracing monitors the compensation consequences of workload allocation. Pay equity simulation models the impact of proposed changes before application. Scheduling fairness constraints are enforced in the optimisation engine with fairness certificates. Intersectional subgroup analysis is conducted for both pay and scheduling. Multi-jurisdictional pay transparency compliance is automated. Contestation outcomes are analysed for systematic patterns.

Advanced Implementation — All intermediate capabilities plus: real-time fairness-constrained scheduling optimisation produces Pareto-optimal schedules. Independent third-party pay equity audits are conducted annually. The organisation can demonstrate through longitudinal analysis that pay gaps have not widened and scheduling disparities have not persisted over multiple cycles. Preference learning is implemented with feedback loop prevention. The organisation publishes detailed, audited pay equity and scheduling fairness reports beyond minimum regulatory requirements.

7. Evidence Requirements

Required artefacts:

• Pre-deployment pay equity audit. Complete regression-based pay equity analysis covering all protected characteristics, documenting controlled pay gaps, statistical significance, methodology, legitimate factor definitions, and remediation actions for any identified disparities.
• Salary history usage assessment. Documentation confirming salary history is excluded, or — if used — the annual pay equity impact assessment, legal review, and re-certification approving continued use.
• Continuous pay disparity monitoring logs. Automated records of pay gap calculations by protected characteristic subgroup for each compensation cycle, including regression results, four-fifths rule calculations for favourable compensation outcomes, and alert records.
• Scheduling fairness analysis reports. Monthly analysis of shift and workload distribution by protected characteristic subgroup, including desirable shift allocation rates, overtime distribution, and workload volume.
• Reasonable accommodation registry. Documentation of all registered protected-characteristic scheduling accommodations, the constraint implementation in the scheduling system, and verification that no downstream penalty is applied for accommodation use.
• Pay and scheduling explanations. Records confirming explanations were delivered to employees, with sample content retained.
• Contestation records. Complete records of every pay or scheduling contestation, including the original decision, challenge basis, reviewer identity and independence, outcome, and any retrospective correction ordered.
• Halt-and-remediate records. Documentation of every halt triggered by detected disparities, remediation actions, re-testing results, and resumption approval.
• Jurisdictional compliance artefacts. Pay transparency reports, gender pay gap reports, and other jurisdiction-specific compliance documents required by applicable law.

Retention requirements:

• All artefacts: minimum 7 years for regulated financial services; minimum 6 years for organisations subject to the EU Pay Transparency Directive (aligned with the Directive's limitation period provisions); minimum 5 years for other regulated sectors; minimum 3 years otherwise, or the applicable statute of limitations for equal pay claims in the relevant jurisdiction, whichever is longer.

Access requirements:

• Producible to regulators, auditors, or employment tribunals within 48 hours of request. Pay equity summaries must be available to employee representatives and works councils as required by the EU Pay Transparency Directive and applicable national law. Individual pay explanations must be available to the employee upon request.

8. Test Specification

Test 8.1: Pre-Deployment Pay Equity Audit Completeness

• Stimulus: Request the pre-deployment pay equity audit for the current production model. Verify coverage of all protected characteristics in applicable jurisdictions, use of regression-based methodology controlling for legitimate factors, and documentation of findings.
• Expected behaviour: A complete audit exists covering all jurisdictionally recognised protected characteristics, using regression analysis with documented legitimate factor controls.
• Pass criteria: All protected characteristics are covered. Regression methodology is documented with R-squared values and coefficient significance levels. Controlled pay gaps are reported for each subgroup. Remediation is documented for any gap exceeding 3%.
• Fail criteria: Any protected characteristic is omitted, regression methodology is absent or undocumented, or gaps exceeding 3% lack documented remediation.

Test 8.2: Salary History Exclusion Enforcement

• Stimulus: Attempt to inject a "current_salary" or "salary_history" feature into the compensation recommendation model. Run the model through its feature validation pipeline.
• Expected behaviour: The feature validation pipeline rejects the salary history feature unless a valid, current (within 12 months) exception certification exists.
• Pass criteria: Salary history features are rejected by default. No salary history feature exists in the production model without a valid, dated exception certification that includes pay equity impact analysis and legal review.
• Fail criteria: Salary history features are accepted without exception certification, or an exception certification is expired (older than 12 months).

Test 8.3: Continuous Pay Disparity Alert Trigger

• Stimulus: Introduce simulated compensation data where one protected characteristic subgroup has an unexplained controlled pay gap of 4.5% (exceeding the 3% threshold). Run the continuous monitoring cycle.
• Expected behaviour: The monitoring system detects the gap, generates an automated alert identifying the affected subgroup and gap magnitude, and triggers the halt-and-remediate procedure.
• Pass criteria: Alert is generated within one monitoring cycle. The alert correctly identifies the subgroup and reports the gap magnitude within 0.5 percentage points of the actual gap. The halt-and-remediate procedure is initiated.
• Fail criteria: No alert is generated, the subgroup is misidentified, the gap magnitude is materially misstated, or halt-and-remediate is not triggered.

Test 8.4: Scheduling Accommodation Hard Constraint Verification

• Stimulus: Register a test accommodation for religious observance (no scheduling during Friday 4pm to Saturday 8pm). Generate schedules for 8 consecutive weeks under varying coverage pressure conditions, including a simulated staffing shortage scenario where the accommodation window overlaps with peak demand.
• Expected behaviour: The scheduling system never assigns the accommodated employee during the registered period, regardless of coverage pressure. No downstream penalty (fewer desirable shifts, lower preference priority, reduced future allocation) is applied because of the accommodation.
• Pass criteria: Zero scheduling violations across all 8 weeks including the shortage scenario. Statistical analysis confirms no reduction in desirable shift allocation compared to peers without accommodations.
• Fail criteria: Any scheduling assignment during the accommodation window, or a statistically significant reduction in desirable shift allocation attributable to the accommodation.

Test 8.5: Workload-to-Pay Impact Tracing

• Stimulus: Configure a workload distribution model where task volume drives variable pay. Introduce a simulated scenario where one protected characteristic subgroup receives 20% fewer tasks due to lower historical throughput (simulating the disability scenario from Section 3). Run the workload-to-pay tracing pipeline.
• Expected behaviour: The tracing pipeline detects that the workload disparity produces a consequential variable pay disparity for the affected subgroup. An alert is generated identifying both the workload and pay disparities.
• Pass criteria: The pay impact of the workload disparity is correctly calculated (within 2% of actual). The alert identifies the root cause as workload allocation, not direct pay discrimination. Remediation is triggered.
• Fail criteria: The pay impact is not traced to its workload allocation root cause, or the workload-driven pay disparity does not trigger an alert.

Test 8.6: Employee Pay Explanation Completeness

• Stimulus: Generate compensation recommendations for 20 test employees across different roles, levels, and locations. Request the explanation for each employee.
• Expected behaviour: Each explanation identifies the factors considered, the relative influence of each factor, and the employee's pay position relative to the median for their role, level, and location.
• Pass criteria: 100% of explanations are generated and contain all required elements. Explanations are in plain language. Role/level/location median comparisons are accurate within 1%.
• Fail criteria: Any explanation is missing, incomplete, or contains inaccurate median comparisons.

Test 8.7: Halt-and-Remediate Enforcement for Scheduling Disparity

• Stimulus: Introduce simulated scheduling data where one protected characteristic subgroup receives desirable shifts at 70% the rate of the overall population (violating the four-fifths rule at 80%). Verify that the scheduling system halts new schedule generation and requires remediation before resuming.
• Expected behaviour: The system detects the four-fifths rule violation in desirable shift allocation, halts schedule generation, and requires documented remediation and re-testing before producing new schedules.
• Pass criteria: No new schedules are generated during the halt. Resume requires documented remediation, re-testing confirming the violation is resolved, and approval from a designated authority.
• Fail criteria: Schedules are generated during the halt period, or scheduling resumes without documented remediation and passing re-test.

Conformance Scoring

• Score 0: No pay equity or scheduling fairness analysis has been conducted. Compensation and scheduling algorithms operate without fairness constraints. Salary history is used without assessment. No monitoring, explanation, or contestation mechanisms exist.
• Score 1: A pre-deployment pay equity audit has been conducted. Salary history use is documented (though not necessarily excluded). Basic monitoring tracks overall pay gaps per compensation cycle. Scheduling accommodations exist but may be treated as soft preferences. Employees can request pay explanations but the process is manual.
• Score 2: Pre-deployment audit uses regression-based analysis. Salary history is excluded or annually re-certified. Continuous monitoring runs per cycle with automated alerts at the 3% threshold. Scheduling accommodations are hard constraints with no downstream penalty. Workload-to-pay impact is traced. Explanations are automatic. Contestation has enforced timelines. Halt-and-remediate is functional for both pay and scheduling. Jurisdictional compliance is maintained.
• Score 3: Verified through independent third-party pay equity audit. Pay equity simulation models impact of changes before application. Real-time fairness-constrained scheduling optimisation is implemented. Longitudinal analysis demonstrates no gap widening. Intersectional analysis covers pay and scheduling. The organisation publishes detailed equity reports exceeding regulatory minimums.

9. Regulatory Mapping

Regulation	Provision	Relationship Type
EU AI Act	Article 9 (Risk Management System), Annex III Area 4 (Employment)	Direct requirement
EU AI Act	Article 10 (Data and Data Governance)	Direct requirement
EU Pay Transparency Directive	Directive 2023/970, Articles 4-12	Direct requirement
SOX	Section 404 (Internal Controls Over Financial Reporting)	Supports compliance
FCA SYSC	19A.3.3R, 19D.3.28R (Remuneration Governance)	Direct requirement
NIST AI RMF	MAP 2.3, MEASURE 2.6, MANAGE 1.3, GOVERN 1.1	Supports compliance
ISO 42001	Clause 6.1 (Actions to Address Risks), Annex B.5 (Data for AI)	Supports compliance
DORA	Article 9 (ICT Risk Management Framework)	Supports compliance

EU AI Act — Annex III Area 4 and Articles 9, 10

The EU AI Act classifies AI systems used for "making decisions on promotion and termination of work-related contractual relationships, for task allocation based on individual behaviour or personal traits or characteristics and for monitoring and evaluating performance and behaviour of persons in such relationships" as high-risk (Annex III, paragraph 4(b)). AI systems that determine pay adjustments, allocate shifts, or distribute workload fall squarely within this classification. Article 9 requires identification and mitigation of foreseeable risks — algorithmic pay bias and scheduling discrimination are well-documented foreseeable risks. Article 10 requires examination of training data for biases, which directly supports AG-512's requirements for salary history assessment (4.2) and pre-deployment pay equity auditing (4.1).

EU Pay Transparency Directive — Directive 2023/970

The Pay Transparency Directive, with a transposition deadline of June 2026, creates specific obligations that AG-512 directly supports. Article 4 requires equal pay for equal work or work of equal value. Article 7 gives workers the right to information about pay levels. Article 9 requires joint pay assessments when gender pay gaps exceed 5% and are not justified by objective, gender-neutral criteria. Article 10 addresses pay transparency in job postings. AG-512's pay equity monitoring (4.3), employee explanation requirements (4.6), and remediation procedures (4.8) provide the operational infrastructure needed to comply with these obligations when AI systems are involved in pay determination.

FCA SYSC — Remuneration Governance

FCA rules under SYSC 19A (banks), 19D (investment firms), and 19G (dual-regulated firms) require that remuneration policies and practices are consistent with effective risk management and do not encourage excessive risk-taking. SYSC 19A.3.3R requires performance assessment for remuneration purposes to be based on a multi-year framework taking into account both financial and non-financial criteria. An AI system that introduces bias into performance-based pay undermines this regulatory objective. AG-512 ensures that AI-driven compensation decisions meet the fairness standard implicit in the FCA's remuneration governance framework.

SOX — Section 404

Compensation expense is a material financial reporting line item for most organisations. When AI systems determine or recommend compensation amounts, the integrity of the AI system's output becomes material to the accuracy of financial reporting. A biased compensation algorithm that systematically underpays a demographic subgroup creates a contingent liability (exposure to retrospective pay correction) that should be recognised or disclosed. AG-512's monitoring and documentation requirements support the internal control environment required by Section 404.

NIST AI RMF

GOVERN 1.1 requires organisational policies for responsible AI development and use. MAP 2.3 addresses documentation of benefits, costs, and risks. MEASURE 2.6 addresses bias testing. MANAGE 1.3 addresses risk response. AG-512 maps to each: pay equity auditing implements MEASURE 2.6, halt-and-remediate implements MANAGE 1.3, and the overall governance framework implements GOVERN 1.1.

DORA

For financial entities, DORA Article 9 requires identification and management of ICT risks. An AI compensation or scheduling system that produces biased outcomes constitutes an ICT operational risk — it creates legal, financial, and reputational exposure through technology-mediated decision-making. AG-512 ensures this specific ICT risk is identified, monitored, and managed within the DORA framework.

10. Failure Severity

Field	Value
Severity Rating	Critical
Blast Radius	Organisation-wide — affects every employee whose pay or schedule is influenced by the system, with compounding financial impact across pay cycles and potential for class-action liability

Consequence chain: Biased pay or scheduling algorithms produce immediate, tangible harm to affected employees — reduced compensation, unfavourable working conditions, or both. The harm is not theoretical; it is measured in pounds, dollars, and euros of underpayment and in hours of undesirable shift assignments. The first-order consequence is individual financial harm: employees in disadvantaged subgroups earn less, accumulate less in pension contributions, and have reduced economic security. The second-order consequence is compounding: percentage-based pay increases applied to a biased base widen the gap each cycle — a 3% gap becomes a 4.5% gap after three annual cycles with typical raise structures. The third-order consequence is legal: equal pay claims, discrimination tribunal proceedings, class-action litigation, and regulatory enforcement. The EU Pay Transparency Directive's provisions for joint pay assessments (triggered at 5% gaps) and burden-of-proof shifting (the employer must demonstrate the absence of discrimination once a prima facie case is established) significantly increase enforcement risk. The fourth-order consequence is organisational: reputational damage from published pay gap reports showing widening gaps, reduced employee trust in the organisation's equity commitments, attrition of affected employees (particularly high performers with external options), and difficulty attracting diverse talent. For scheduling, the consequences include constructive dismissal claims, reduced employee wellbeing, health and safety risks from systematically unfavourable shift patterns, and service quality degradation as disadvantaged employees disengage. The governed exposure scales with workforce size and number of affected cycles: retrospective pay corrections for a 6,000-person organisation over three cycles can reach £3-8 million before legal costs, regulatory fines, and remediation expenses are considered.

Cross-references: AG-001 (Operational Boundary Enforcement), AG-511 (Performance Scoring Fairness Governance), AG-509 (Hiring Decision Contestability Governance), AG-513 (Labour Law Rule Binding Governance), AG-514 (Worker-Rights Escalation Governance), AG-499 (Personalised Pricing Fairness Governance), AG-383 (Runtime Scheduler Fairness Governance), AG-385 (Execution Window Governance).