Nearly a third of healthcare organizations have adopted AI agents, partly without IT approval
The American Medical Association launched a governance framework for agentic AI in healthcare on October 15 — an attempt to set standards for tools that can carry out multi-step tasks in clinical care without human oversight. The need is evident: In a survey of 250 US healthcare executives, 28% say their organization has already adopted such tools, 44% are running pilots, and 72% report that tools are sometimes implemented without IT approval. The framework, however, arrives into a regulatory vacuum, with no federal AI regulation and alongside parallel work by the Federation of State Medical Boards.
Four questions for clinical AI agents
The framework, as reported by Healthcare Dive (republished via Yahoo News), is organized around four questions the AMA will use when evaluating clinical AI agents:
- What clinical function does the AI perform?
- What evidence shows it can perform that function, and what happens if it fails?
- What should the AI be allowed to do, and under what conditions?
- Who bears responsibility for the risks and outcomes?
The questions cover the entire lifecycle of an agentic tool: what it is supposed to do, the documentation that it can actually do it, the limits on its use, and finally the core of the whole matter — who answers when something goes wrong. According to Healthcare Dive's reporting, the framework is meant to offer insight into how the AMA plans to govern agentic AI as the tools are increasingly used in clinical care, including ambient documentation, diagnostics and patient monitoring.
The principle behind the framework is summarized in a quote Healthcare Dive reproduces from the document: "As AI capabilities advance, evidence, authority, oversight and accountability must remain aligned in service of better patient care."
It is worth being precise about what the framework is: a governance instrument from a professional association, not binding regulation. The AMA is setting a standard for what the organization considers good practice, but the framework carries no authority to enforce its requirements. The reader should also know that the details of the framework here come from Healthcare Dive's reporting; the original AMA document has not been verified directly, so the full text and any nuances cannot be checked against the sources.
Why agentic AI is different
What separates agentic AI from earlier clinical AI tools is the degree of autonomy. According to Healthcare Dive's analysis, these tools can complete multi-step tasks without human oversight — which means an inaccurate or biased decision can harm patients before a clinician has a chance to intervene.
Existing accountability patterns in healthcare are built precisely on a human standing between the tool and the patient: a decision-support system that suggests a diagnosis presupposes that the physician validates it. When an AI agent carries out multi-step tasks on its own — for example retrieving data, assessing it and taking action — that intermediary disappears, along with the question of who is responsible when the outcome is wrong. This is the accountability gap that the framework's fourth question targets.
AMA leadership's skepticism
The framework is not an expression of enthusiasm. AMA chief executive John Whyte was fairly scathing in an interview with CBS News and KFF Health News on October 9, six days before the launch (CBS News):
"Some of these tools just aren't ready for primetime," Whyte said. "We don't have enough data on them. We don't understand how they work, how they use information. There's no liability if these tools get things wrong."
The quote points to the same three weaknesses the framework attempts to address: insufficient evidence, unclear workings and an absence of accountability. That the association's top leader expresses such skepticism while the organization launches a governance framework suggests the goal is to set requirements for a market that is already moving — not to slow it down.
Adoption is outrunning the controls
The numbers from the survey of 250 US healthcare executives illustrate the timing: 28% of organizations already have agentic AI tools in operation, and 44% are running pilots or proofs of concept. Perhaps most striking is that 72% say such tools are sometimes implemented without IT approval.
That last figure is particularly problematic from a governance perspective. If clinically relevant tools are adopted outside the formal IT and quality-assurance chain, then in practice no body has assessed evidence, failure modes or accountability — the framework's first three questions — before the tool meets patients. The survey, however, is not named in Healthcare Dive's reporting, so its sponsor, methodology and date cannot be verified; the numbers should be read with that caveat.
A regulatory vacuum
The framework arrives in a situation where, according to Healthcare Dive, there is no substantial federal AI regulation in the US. Regulation is instead happening state by state: According to the law and consulting firm Manatt, 47 states introduced more than 250 AI proposals affecting healthcare in 2025. The result is a patchwork of local rules with no unified standard — which is the backdrop for the professional associations trying to fill the vacuum themselves.
The FSMB workgroup is underway
In parallel, the Federation of State Medical Boards (FSMB) is developing its own guidelines for AI tools that perform clinical functions with limited or no direct physician oversight. The workgroup has 22 members, and the AMA participates. According to STAT (October 14, 2026), FSMB president Humayun Chaudhry said a draft of the recommendations could be released within "weeks to months."
The two initiatives are, however, separate, and the sources do not establish any coordinated relationship between them. It would stretch the analysis to claim that the AMA framework is an input to, or preparation for, the FSMB draft — even though the AMA has a seat on the workgroup, which makes overlap in content likely.
Open questions
Several questions remain unanswered in the available source material:
- How will the FSMB draft relate to the AMA framework? When the draft arrives within weeks to months, it will become clear whether the two documents complement each other or represent different priorities.
- How binding will the framework be in practice? As professional governance it has no enforcement power; its effect depends on whether hospitals, vendors and insurers adopt the four questions as procurement and implementation requirements.
- How solid are the adoption figures? Since the survey of the 250 executives is unnamed, we do not know who conducted it or how.
- What happens when no one can answer the fourth question? The framework poses the question of accountability, but it is only through regulation and case law that the question will actually be answered.
The clear picture, nonetheless, is this: Agentic AI has already reached clinical practice at significant scale, partly bypassing IT controls, while the rules — federal, state and within the industry — are still playing catch-up. The AMA framework is an attempt to set minimum requirements before the vacuum is filled in other ways.