What the AMA actually released
In mid-October 2026, the American Medical Association's Center for Digital Health and AI published a governance framework for agentic AI assuming clinical functions — a response to AI tools that no longer just generate a single output for a clinician to read, but can act autonomously across multiple steps. The framework draws the line plainly: agentic AI solutions "can operate autonomously, completing complex tasks that generative AI cannot," and that autonomy is precisely what creates the risk that an inaccurate or biased decision reaches a patient before a clinician intervenes.
The framework's organizing idea is that clinical authority has to be earned, not assumed: "for agentic AI solutions to perform clinical functions, they must be backed by evidence and maintain high-quality performance," and authority or oversight should shift as evidence, performance, risk, or the technology itself changes. The AMA puts it directly: "Evidence, authority, oversight and accountability must remain aligned in service of better patient care."
The four questions
The framework is built around four questions any healthcare organization should be able to answer before letting an AI agent touch clinical work:
1. Function
What clinical function is the AI performing?
2. Evidence & failure
What demonstrates that it can perform that function, and what happens if it fails?
3. Authority & oversight
What should the AI be permitted to do, and under what conditions?
4. Accountability
Who is responsible for the risks and outcomes?
None of this is radiology-specific — the framework is written for agentic AI across clinical settings, including ambient documentation and patient monitoring. But imaging is one of the places where "clinical function," "failure," and "authority" are already concrete enough to answer precisely, which makes the framework unusually easy to operationalize for imaging AI purchasing decisions.
Why governance can't wait for the next pilot
The AMA's framework lands amid a documented governance gap. A survey of 250 U.S. healthcare leaders conducted by Vanson Bourne for Imprivata found 28% had already deployed agentic AI tools and 44% were running pilots or proof-of-concept projects — but 72% said their organization had deployed AI tools without formal IT approval at least some of the time. Half cited compliance or regulatory violations as a top concern, and 48% cited unauthorized access to protected health information. Adoption, in other words, is running ahead of the governance structures meant to contain it — exactly the gap the AMA's framework is trying to close.
Turning the framework into a vendor checklist
The AMA's four questions translate directly into four things to ask an imaging AI vendor before you sign a contract — and four answers that should give you pause.
| AMA question | Ask the vendor | Red flag |
|---|---|---|
| Function | Exactly which task does the AI perform — triage flagging, measurement, draft report generation — and where does its autonomy end? | "It handles the read" with no clear boundary |
| Evidence & failure | What validation evidence supports this specific use, and what is the fallback when confidence is low or the output is wrong? | No published validation; no defined escalation path |
| Authority & oversight | Can the system ever finalize or transmit a report before a radiologist reviews it? | Yes, under any routine condition |
| Accountability | Is there a documented audit trail tying each report to the specific AI version and clinicians who touched it, and who is responsible if it is wrong? | No audit trail; liability is undefined or deflected to the client |
Governance by design, not by policy memo
The practical upshot of the AMA's accountability question is that an imaging AI workflow needs a traceable chain: which system produced which part of a report, which physician reviewed it, and who ultimately signed it. That is not a feature an organization can bolt onto an agentic AI product after deployment — it has to be built into the workflow itself. xAID's model is structured around exactly that chain: the AI produces a structured report draft, xAID's in-house radiologist reviews every preliminary, and the report is delivered ready-to-sign — the client's reading radiologist signs the final. The AMA's framework describes, almost point for point, why that division of labor matters.
Frequently asked questions
What is the AMA's new governance framework for agentic AI?
In October 2026, the American Medical Association's Center for Digital Health and AI published a governance framework for agentic AI assuming clinical functions. It is built around four questions: what clinical function the AI performs, what evidence demonstrates it can perform that function and what happens if it fails, what it should be permitted to do and under what conditions, and who is responsible for the risks and outcomes. The framework's premise is that clinical authority should be earned through evidence and maintained through demonstrated performance.
What makes AI "agentic" rather than generative, according to the framework?
Coverage of the framework describes agentic AI as software that can operate autonomously and complete complex, multi-step tasks without human oversight at each step — something generative AI tools, which produce a single output for a human to review, do not do. That autonomy is exactly why the AMA argues agentic tools need explicit limits on what they're permitted to do and under what conditions.
What should imaging AI buyers ask vendors based on the AMA framework?
For each of the four AMA questions, buyers should get a specific, non-marketing answer: exactly which task the AI performs and where its autonomy ends; what validation evidence supports that claim and what happens on a low-confidence or incorrect output; whether the system can ever finalize or transmit a report without a clinician reviewing it first; and whether there is a documented audit trail tying every report to the specific clinicians — AI and human — who touched it.
How does xAID's reporting model align with the AMA's accountability question?
xAID never signs a report. The AI produces a structured draft, xAID's in-house radiologist reviews every preliminary, and the report is delivered ready-to-sign so the client's reading radiologist signs the final. That sequence creates the documented chain of custody — which system acted, which physician reviewed it, who signed — that the AMA's accountability question asks every agentic AI deployment to be able to answer.
Source: American Medical Association governance framework for agentic AI, as reported by MedTech Dive (Oct. 16, 2026) and Healthcare Dive. Survey data on agentic AI deployment: Vanson Bourne for Imprivata, "The Agentic AI Trust Gap: Why Healthcare Needs Identity-Led Governance" (Sept. 2026), as reported by TechTarget. Figures are rounded as reported.