How Much Should Artificial Intelligence Do Without a Physician? An Autonomy Scale for AI Agents in Solid Organ Transplantation.

Abstract Background: Artificial intelligence in surgery and transplantation is moving from tools that give a single output toward software agents that retrieve data, plan several steps, use other tools, and take actions. Existing autonomy classifications and device frameworks address physical autonomy and fixed-function outputs well, but they do not specify oversight for systems whose behavior depends on intermediate decisions and tool use. Methods: We developed a governance framework by combining Yang and colleagues’ six-level autonomy scale with an agentic overlay covering tool use, data access, action scope, and persistence. We mapped the framework against current U.S. guidance for AI-enabled devices and the American Medical Association’s 2026 framework for agentic AI. We illustrated its application with four worked examples from solid organ transplantation: donor procurement, machine perfusion, and post-transplant monitoring with cell-free DNA and donor-specific antibodies, assigning each plausible agent function a level and governance tier and identifying the controls that apply. Results: The framework assigns each agent function to one of four governance tiers based on autonomy level, reversibility, and consequence. Tier A (levels 0 to 1) requires accuracy testing and query logging. Tier B (level 2) requires bounded limits, full action logs, and drift monitoring. Tier C (level 3) requires clinician approval of each consequential action, prospective evaluation, and named accountability. Tier D (levels 4 to 5) is not recommended for high-consequence decisions in the current evidence period. In the transplant example, the organ-acceptance decision is assigned to Tier C with clinician decision authority, allocation pairing is excluded from agent action, and a protocol-defined perfusion-time limit is enforced deterministically outside the model. Across the four examples, clinicians approve each consequential action, no viability marker drives acceptance alone, and automatic immunosuppression changes are excluded. Conclusions: Oversight of agents in transplant care should scale with autonomy, reversibility, and consequence rather than with product category. Binding rules should be enforced in deterministic code rather than left to model reasoning. Prospective evaluation is required before any agent in a high-consequence transplant decision advances beyond Tier C.