AI in Performance Reviews, Pay and Redundancy Selection
AI-driven decisions about how much employees earn, how they are evaluated, and whether they keep their jobs during restructuring are among the highest-stakes applications of AI in employment. These decisions affect not just livelihoods but also dignity and equal opportunity. HR teams must understand the legal frameworks, the risks, and the controls required.
Performance Reviews and Rating Systems
AI is increasingly used to generate performance review scores, often by analysing email communications, productivity metrics, project completion rates, and manager notes. The system might score each employee and recommend a ranking.
Performance review systems are subject to multiple legal constraints. First, under data protection law, automated decision-making that produces legal or significantly significant effects (like a performance rating that affects salary, promotion, or redundancy eligibility) triggers the right to human review. A performance rating generated entirely by algorithm, without manager judgment, is likely unlawful.
Second, under employment law, performance ratings must not discriminate. If AI-generated ratings systematically rate female employees lower than male employees with identical performance, that is sex discrimination. If ratings systematically disadvantage older workers, that is age discrimination.
If ratings account for disability as a factor (such as penalising employees for disability-related absences or accommodations), that is disability discrimination.
Third, performance ratings must be based on objective, job-relevant criteria. A rating system that down-scores employees because the algorithm detected “low engagement” in emails is measuring sentiment, not performance. Sentiment analysis in performance rating carries high risk of discrimination; it tends to penalise communication styles that are more common among women, people from diverse backgrounds, and neurodiverse employees.
Pay Decisions and Algorithmic Discrimination
AI is also used to determine pay increases, bonuses, and benefits. An algorithm might suggest that an employee receives a 2 percent raise because their productivity score is 2 percent above baseline. It might recommend no bonus for an employee whose productivity dipped because they took parental leave.
Pay decisions are among the most serious applications of AI, because they compound over time. A small algorithmic bias in one year’s pay decision becomes a larger gap in subsequent years. If an algorithm consistently gives slightly lower raises to women, the gap compounds and becomes a significant differential over decades.
Under the Equality Pay Act 1970 and the Equality Act 2010, employees have the right to equal pay for equal work. Pay discrimination is unlawful. If an AI system recommends or determines pay in a way that discriminates, the organisation is in breach of law regardless of whether the discrimination was intentional.
Redundancy Selection and Algorithmic Fairness
AI systems are increasingly used to score employees for redundancy: the system produces a ranked list of “highest risk” employees who are most likely to be selected for redundancy. Managers then make final decisions from that list.
This is problematic on several levels. First, it creates a huge potential for discrimination. If the system was trained on historical redundancy data, and historical redundancy decisions were biased (perhaps older workers were more frequently selected in the past), the system will learn and reproduce that bias.
Second, it violates the principle that redundancy decisions should be based on job-related criteria, not on protected characteristics. A system that scores employees on productivity, which may proxy for disability status or parental leave history, is likely discriminatory.
Third, it violates Article 22 rights: employees have the right to human judgment in decisions with significant effects on them, particularly decisions affecting their employment.
An organisation that uses AI to select employees for redundancy should expect legal challenge unless it has documented that: the system is not biased, the criteria are job-relevant and not proxies for protected characteristics, alternative selection criteria were considered, and each employee has the right to human review of the decision before their employment is terminated.
Compliant Frameworks
A compliant approach to AI in people decisions includes: clear definition of job-relevant criteria, assessment of whether AI should be involved at all or whether human judgment alone is more appropriate, bias testing before deployment, ongoing monitoring for discrimination, transparency to affected individuals, and human involvement in all decisions that affect significant interests (pay, promotion, redundancy).
For many organisations, this means that AI has an advisory role (providing analysis or recommendations) but not a decision-making role. A manager reads an AI-generated performance assessment but makes the final rating decision based on their judgment and the individual’s circumstances.