Ask who's responsible.

The American Medical Association released a governance framework for agentic AI, the kind that carries out multi-step tasks instead of answering one question at a time. The starting principle is that an agent doing clinical work has to be backed by evidence and keep performing well. Then it boils down to four questions. What clinical function is the AI performing? What shows it can do that, and what happens if it fails? What should it be allowed to do, and under what conditions? Who is responsible for the risks and outcomes?

The timing matters. Healthcare Dive points to a recent survey of 250 U.S. healthcare leaders where about 28% said they've already deployed agentic AI and 44% are piloting it. And 72% said these tools sometimes get rolled out without IT approval.

The Wall Street Journal reports that a run of AI safety incidents, from agents hacking websites to chatbots accused of encouraging self-harm, is exposing a gap in insurance. AI developers face potential liabilities in the billions, and the businesses using their tools are exposed too. Claims could outgrow what existing policies cover.

A narrative review in Clinical and Experimental Emergency Medicine, from a team at Akita University in Japan, lays out fairness for language models in the ED in three layers. The data layer covers biased training and prompt data, like records with non-random gaps or stigmatizing language. The output layer asks whether its recommendations and instructions are equally accurate for every patient group, even when overall accuracy looks strong. The human layer asks whether clinicians over-trust or ignore the tool, and whether patients can access and understand what it tells them.

Confido Health raised a $12.5 million funding extension to expand CareOS, a platform built to automate practice operations like scheduling, insurance verification, billing, and patient outreach. The company frames it as moving practices from systems that store records to systems that take action.

A scoping review in the Journal of Medical Internet Research looked for studies on what generative AI teaching tools actually do for undergraduate medical students. Only 27 made the cut. Of those, 23 measured how students reacted, 16 looked at learning outcomes, and none measured whether behavior changed. Meanwhile, 119 studies the team screened had no student-facing intervention at all, and 110 were about how well the AI performed. The authors propose a road map, called TIED, that asks educators to design for how, for whom, and under what conditions a tool helps students learn.

Google DeepMind released Gemini 4 Argon on October 13, its new top model. It's in the Gemini app for Google AI Pro and Ultra subscribers and in the API at an introductory $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after November 30. In the launch post, Ambience Healthcare says that on more than 200 synthetic patient charts, Argon had the highest recall and ICD-10 code specificity of the models it tested, and that it's now validating on real records.

List every AI tool your practice uses, including the ones staff picked up on their own. For each one, write one line per AMA question: what it does, how you know it works, what it's allowed to do without a person, and who owns the outcome. Any blank line is your to-do list. Then send the list to your insurance carrier with the coverage question from above. It's an hour of work, and you'll know more about your risk than most practices do.