The American Medical Association (AMA) issued a new framework for governing artificial intelligence agents in healthcare, in an effort to define the conditions for using tools capable of carrying out multiple-step clinical tasks with a degree of autonomy. The framework emphasizes that these solutions must be based on clear evidence and maintain high-quality performance before being allowed to perform functions that affect patient care.
Artificial intelligence agents differ from traditional generative artificial intelligence tools in their ability to execute a sequence of actions without continuous human oversight. Potential uses include ambient clinical documentation, diagnostic assistance, and patient monitoring. However, an error or bias in an autonomous decision could harm a patient before a physician intervenes.
Four Questions Before Delegating a Task
The AMA framework proposes that the evaluation of any clinical agent be based on four key questions:
- What clinical function does the artificial intelligence system perform?
- What evidence demonstrates its ability to perform this function, and what happens if it fails?
- What actions should the system be allowed to perform, and under what conditions?
- Who bears responsibility for the risks and outcomes?
The association summarizes its position by stating that evidence, permissions, oversight, and accountability must remain aligned with the goal of improving patient care, while emphasizing the central role of physicians in developing and deploying these technologies.
Why Does This Framework Matter?
The initiative comes as trials of artificial intelligence agents expand within healthcare institutions. According to a survey of 250 healthcare leaders in the United States, 28% of organizations said they had already deployed tools of this kind, while 44% reported that they were conducting trials or proof-of-concept projects. Most participants expect a transformative impact on clinical workflows.
At the same time, 72% of participants said that artificial intelligence agent tools are sometimes implemented without approval from the information technology department. This points to a gap in institutional oversight, as more than half of the participants identified the potential for violations of compliance or regulatory requirements, and unauthorized access to protected health information, among the leading concerns.
Guidance in the Absence of Comprehensive Federal Regulation
The healthcare sector is operating amid incomplete federal regulation of artificial intelligence, prompting professional and industry groups to develop interim principles and controls. Across the United States, laws vary between states that are tightening restrictions and others that are testing artificial intelligence-supported healthcare applications. According to the law and consulting firm Manatt, 47 states introduced more than 250 bills affecting artificial intelligence in healthcare during 2025.
The framework alone does not determine how each agent should be approved or eliminate operational risks, but it provides a practical basis for raising questions about responsibility, permissions, and evidence before introducing autonomous tools into patient-care workflows. How to translate these principles into measurable requirements, and how to distribute responsibility among physicians, institutions, and system developers, remain open questions.