Rapid NHS AI adoption is driving organisational reconfiguration of clinical work, decision authority, and accountability allocation, yet governance clarification remains incomplete. This study advances a configuration-specific explanation of human-AI interaction by analysing how alternative AI configurations redistribute judgement, responsibility, and escalation control, and how such redistribution shapes trust formation, role restructuring, and equity exposure. A comparative multiple-case mixed-design investigation across four NHS trusts examines collaborative, decision-support, autonomous, and adaptive/escalation-aware configurations, focusing on override enactment, handover specification, auditability, and accountability assignment in practice. The analytical integration of socio-technical theory, human-AI collaboration theory, and technology trust theory supports mechanism identification linking configuration design to organisational adaptation. The principal output is an empirically refined Minimum Viable Accountability Control (MVAC) matrix, specifying the minimum documentation and oversight control-set required for safe, trusted, and equitable AI implementation at scale.