AI Governance at the Point of Care: A Scenario-Based Analysis of Governance Composability in High- Stakes Maternal Care

Abstract BackgroundArtificial intelligence (AI) is increasingly incorporated into maternal and obstetric care through risk prediction,clinical decision support, monitoring, diagnostic interpretation, patient communication, and generative AI.Existing governance literature provides mechanisms for risk management, human oversight, verification,monitoring, accountability, and intervention, but these mechanisms are distributed across frameworks withdifferent scopes and governance functions. ObjectiveThis study examined whether governance requirements originating from heterogeneous AI-governanceframeworks could be composed into coherent and actionable governance decisions when applied to thesame high-stakes maternal care scenarios. MethodsI conducted a comparative, scenario-based qualitative document analysis using four heterogeneousgovernance lenses: the NIST Generative AI Profile, a meaningful-oversight framework for medical AI, anormative tiered framework for generative AI governance, and a maternal and child health AI implementationframework. I applied the four lenses to six standardized scenarios and generated 24 framework–scenario analytical cells. I extracted trigger, action, authority, condition, timing, verification/control,reconfiguration, and reversibility requirements, and then compared the resulting decision profiles. ResultsI found substantial compatibility at the level of broad governance principles and complementary coverageacross governance functions. I did not identify a substantive conflict among the four governance lenses. Thecompound high-stakes scenario produced the greatest coordination challenge: multiple legitimategovernance requirements became simultaneously relevant, but they did not automatically resolve into ashared cross-framework decision procedure specifying precedence, authority, activation conditions,sequencing, and control transitions. ConclusionThe findings distinguish governance availability from governance composability. Existing frameworks canprovide substantial governance coverage without automatically specifying how their requirements should becoordinated at a single high-stakes decision point. The finding supports empirical testing rather than a claimof universal governance failure.