The American Medical Association released a governance framework for agentic AI in clinical functions, aiming to keep clinical authority and accountability aligned as these systems become more capable.

 

What happened

The framework starts with the principle that AI should earn clinical authority through evidence and keep it through demonstrated performance. It focuses on four areas: function, evidence and risk, authority and oversight, and accountability. The AMA said these questions will guide how it works with physicians and other stakeholders to advance approaches to liability, licensure and regulation. The AMA is also bringing physicians together with those who build, deploy and regulate AI to shape responsible governance.

 

The backstory

Last year, the AMA created the Center for Digital Health and AI to ensure physicians help shape how AI is developed, implemented and regulated across the healthcare system. This past September, the AMA named Jennifer Goldsack, MChem, the center’s inaugural leader and a senior vice president. Goldsack has experience at the Digital Medicine Society, a global nonprofit she founded in 2019.

The AMA is not the only group to act. Earlier this year, the Health Sector Coordinating Council (HSCC) Cybersecurity Working Group published its own AI governance guide, the “Health Industry AI Cyber Governance Framework Implementation Guide.” The guide helps healthcare organizations set up cyber governance frameworks for secure AI implementation.

 

Going deeper

The AMA framework breaks governance into four parts:

  • Function: defines the task the AI performs.
  • Evidence and risk: demonstrates that the AI can perform the function and what happens if it fails.
  • Authority and oversight: sets the permissible scope of action and the human or institutional oversight required for safe use.
  • Accountability: names who is responsible for the risks and outcomes.

 

What was said

The AMA said the framework is meant to ensure “increasingly capable AI improves care while clinical authority and accountability remain aligned.”

The AMA also said, “As agentic AI systems become capable of assuming increasingly consequential clinical functions, clear approaches are needed to determine what these systems should be permitted to do, under what conditions, and who is accountable for their actions and outcomes.”

Dr. Fahad Razak told Healthcare Finance News that physicians need a way to review AI for patient safety and to protect themselves from liability.

Why it matters

Two major healthcare groups have now published AI governance frameworks from different angles, and both address agentic AI. The AMA asks who is accountable when an AI system takes on a clinical function. The HSCC asks how organizations protect those same systems from threats like data poisoning and adversarial attacks.

This matches Razak’s warning that AI is arriving so fast that the ability to review it against a reliable standard does not yet exist. The AMA’s accountability question speaks to his concern about physician liability. The HSCC guide gives organizations a structure for the security side.

 

The bottom line

The AMA wants clinical authority for AI to be earned and kept through evidence, not assumed. Healthcare organizations can start by reviewing the AMA framework which recommends beginning with an inventory of the AI systems already in use. Physicians, AI builders, deployers and regulators should also watch how the AMA turns this framework into approaches to liability, licensure and regulation.

Related: HIPAA Compliant Email: The Definitive Guide

 

FAQs

What is an AI governance framework?

It is a structured set of policies, oversight roles, and risk controls that determines how an organization approves, uses, monitors, and audits AI tools.

 

What is human oversight of AI?

It means a qualified person or institution can review, question, and override an AI system’s output or actions before they affect patient care.

 

What is model drift?

Model drift happens when an AI system’s performance declines over time because real-world data no longer matches the data it was trained on.

 

What is data poisoning?

Data poisoning is an attack in which someone corrupts the data used to train or run an AI system so it produces flawed or manipulated results.

 

Why is an AI inventory a good first step?

An organization cannot govern or secure AI systems it does not know it has, so listing every tool in use shows where oversight is needed.