Accountability in action: Exploring ethical and legal risks in AI- supported employee relations in the UK railway industry

HR practitioners increasingly deploy Artificial Intelligence [AI] in employee relations [ER] activity that is high-stakes, procedurally sensitive, and legally bounded, yet empirical research into how they navigate the associated ethical and legal risks remains limited. This study gathered practitioner insight into those risks and developed theory- and practice-based recommendations for this nascent field. Semi-structured interviews with 11 senior HR practitioners in the UK railway industry were analysed within an interpretivist approach using NVivo, with reflexivity maintained throughout as the researcher is employed by the same organisation. Coding developed four themes: professional judgement at the human-AI boundary; ethical and legal risks and guardrails; strategic oversight and digital maturity; and tensions across system ownership, governance, and accountability. From these, four empirical and four conceptual contributions were advanced. The empirical contributions are an accountability gradient between tactical and strategic practitioners; the passive positioning of HR in AI governance; context-sensitive anthropomorphisation; and information-flooding as a pathway to weakened meaningful human involvement and increased Art. 22 GDPR exposure. The conceptual contributions are a three-axis framework [kind/wicked × simple/complex × AI/human], a four-domain nexus, the alignment of kind/wicked rulesets with the jagged frontier, and an ER-specific case-law synthesis reinforcing substantive Human-in-the-Loop [HITL]. Findings suggest variation in senior practitioners' technical understanding, which may constrain their ability to retain meaningful ethical accountability for ER decision-making, with HITL frequently symbolic rather than substantive. HR must shift from passive policy recipient to active contributor in AI governance. This appears to be among the first empirical studies focused specifically on AI-supported Employee Relations practice, and no directly comparable study within the UK railway industry was identified by the author. Generative AI use is disclosed in full in the Declarations section [editorial/grammar support only; see section 4.5 and Appendix E for the reflexivity comparator methodology].