Governing work delegated to AI through dynamic control fit

Abstract When organizations delegate work to artificial intelligence (AI), controls that remain formally in place may no longer achieve their purposes. We combine systematic identification and selection with integrative analysis of 171 papers across five research streams and 26 supplementary sources. The synthesis develops dynamic control fit: an arrangement's ability to achieve a specified control function while sustaining or restoring the conditions on which that function depends. We explain this ability through a framework of three interconnected requirements: Deliver, Renew and Adapt. Deliver concerns whether coverage, information, expertise and opportunities to intervene support effective control in current work. Renew concerns the practice, feedback and maintenance through which these capabilities persist. Adapt concerns recognizing deficiencies, obtaining authorized changes and verifying restored performance. The framework explains how AI may improve immediate enforcement while reducing experiences through which people develop judgment for future oversight, although other activities may compensate. Coverage, reviewer capacity and revision authority impose distinct constraints requiring different responses. Continued reliance must also be assessed separately from performance because beliefs, external expectations and practical constraints can sustain or limit use. By following a specified function across changes in control arrangements, the review develops testable explanations of improvement, erosion, compensation and recovery.