Who Frames, Who Acts, Who Learns? Allocating Decision Rights for Agentic AI in Graduate Medical Education, Clinical Research, and Clinical Care

Background: Health care organizations are deploying artificial intelligence (AI) agents faster than they are deciding who holds authority over the decisions those agents touch. In June 2026, the MIT Center for Information Systems Research (CISR) published the AI Decision Matrix, which assigns humans and AI to three decision components (frame, act, and learn) along two dimensions: ambiguity and risk. The framework was derived from executives in telecommunications, financial services, manufacturing, and law, and it has not been examined in health care. Methods: We conducted a structured conceptual analysis. We operationalized the two dimensions with health care anchors, assembled an inventory of 18 recurring decisions across graduate medical education (GME), clinical research operations, and clinical care, and classified each decision by quadrant. We then compared the framework’s default human and AI assignments with the accountability requirements of the ACGME, 21 CFR, ICH E6(R3), FDA clinical decision support guidance, the ONC HTI-1 rule, and the EU AI Act. Results: Five decisions were classified as Routine, four as Consequential, four as Exploratory, and five as Strategic. Health care regulation altered the framework’s defaults in two systematic ways. A licensure floor requires a named professional in the act component of 5 of the 18 decisions, including two in which the source framework assigns the act to AI. A learning floor makes outcome ownership a human responsibility in every decision that affects a patient, trainee, or research participant, including Routine ones. We also identified quadrant drift, in which tools introduced as low-risk pilots migrate into high-risk use without reclassification. Conclusions: The AI Decision Matrix offers health care leaders a practical vocabulary for allocating decision rights, but it requires two domain-specific constraints and a reclassification trigger before adoption. We propose an adapted matrix and recommend empirical validation through a modified Delphi study and prospective institutional audits.