1) Agentic AI — Military and health-system rollouts move from pilot to scale

The Defense Health Agency completed a phased global rollout of its ambient-listening Clinical AI Agent across military hospitals and clinics. Mount Sinai and Mayo Clinic are running agentic workflows for scheduling and administrative automation; NHS launched a program on responsible agentic AI deployment. Why it matters: Moves the agentic AI conversation from "what is it" to "who's accountable when it acts outside its bounds" — the harder, more C-suite-relevant question.

ACC Comments: 

I am noticing and hearing that agentic AI in healthcare is slowly progressing from curiosity to capability. An erroneous dividend from generative AI is a nuisance, whereas an erroneous act from agentic AI is a liability. Military health is a closed system with one payer, one enterprise EHR, and one chain of command. It can scale an agent globally perhaps easier  because accountability is vertical. This reminds me of what I observed in China and its health system. Civilian systems, with their federated medical staffs and negotiated autonomy, have no equivalent hierarchical structure. 

Note the chosen beachheads: ambient documentation, scheduling, and administrative automation. Most if not all of these areas are reversible. The accountability question is answerable. Agents should receive delineated privileges, not enterprise licenses, with defined scope, proctoring period, immutable audit trail, a named supervising executive, and a reappointment cycle where performance is monitored closely. An agent operating outside its bounds should trigger peer review, not a press release. The NHS is writing governance while American systems write code.

 

2. Nursing/workforce — AI adoption triples, but training gap blunts the payoff

A new State of Nursing report found AI use among nurses nearly tripled in a year, yet almost half report no time savings. Only 8% of nurses say their employer has a clear AI strategy; nurses with real training save over an hour a day versus 16% without. Nurses who helped select tools use them at far higher rates (81% vs. 62%) than those never consulted. Why it matters: Direct evidence that governance and change-management — not the tool itself — determine ROI. Speaks to retention, a top-named priority for 67% of healthcare leaders this year.

ACC Comments:

I have been very encouraged by the exponential rise in interest for AI in healthcare amongst nurses as well as nurse practitioners and physician assistants, especially in taking the ABAIM courses (including a special edition just for nurses a few months ago). Time liberated inside a hospital does not accrue to the person who liberated it; it is reabsorbed by the next task in the queue. Nearly all of these nurses are using AI inside organizations that have not decided what AI is for , which means the tool is arriving as a personal coping strategy rather than an institutional one. That is shadow adoption, and it is the least governable form.

The consultation finding deserves more attention than it will get. Nurses who helped select the tools use them at 81%. People who know the workflow selected tools that fit the workflow; people who didn't, did not do so. The nurses do not leave over technology. They leave over the steady accumulation of unfunded work. An AI tool deployed without workflow redesign becomes exactly that: one more thing to learn, on your own time, for someone else's efficiency metric. Adoption is a purchase but application is a practice.

 

3. Regulation — CMS builds a payment pathway for clinical AI

CMS signaled plans to rename "Software as a Service" to "Software as a Medical Service," a new interim payment category for clinical AI and software, and stood up a dedicated office to oversee AI, interoperability, and digital health across CMS programs. The ACCESS (Advancing Chronic Care with Effective, Scalable Solutions) Innovation Center model, aimed at AI diagnostics and workflow software, began its first cohort July 1. Why it matters: Reimbursement uncertainty is the single biggest reason "exploring"-stage health systems stall on AI pilots. A named payment pathway gives CFOs and C-suite something concrete to plan around — this is a session that pulls the C-suite in the room, not just IT.

ACC Comments:

I have been stating for more than two years now: AI in healthcare belongs on the boards of directors and the C-suites agendas, and not solely on the CIO's to-do list. In addition, boards of directors and the C-suites need to be educated on AI in healthcare (one cannot govern something one knows little about). The renaming mentioned is a substantive act. "Software as a Service" is a procurement category as a line item negotiated by IT whereas "Software as a Medical Service" is a clinical category, and the word imports the entire apparatus behind it: a furnishing entity, medical necessity, documentation, and eventually audit. Payment confers legitimacy and exposure in the same motion. 

Provisional pricing buys time while evidence accumulates, which means the real standard is not being written now but will be written at conversion, in the criteria CMS uses to make the category permanent. That is where the operative definition of clinical AI value gets set through coding policy rather than legislation. One structural caution: paying per inference recreates volume incentives in a new medium. And a familiar asymmetry: this pathway is Medicare-anchored. Pediatrics inherits it through Medicaid, late and unevenly, as always. Children's systems should plan for the lag rather than discover it.

 

4. Governance/legal — Mayo Clinic whistleblower suit alleges systemic AI governance failure

A federal lawsuit (filed July 6, continuing to generate coverage this week) alleges Mayo Clinic demoted and terminated an employee who raised AI compliance concerns over 18 months — citing bypassed IRB review, a concealed 67% error rate in an internal AI tool, and unauthorized use of an AI-guided procedure abroad. Separately, a suit against Sutter Health and Memorial Healthcare Services alleges an AI ambient-listening tool illegally recorded patient-clinician conversations. Why it matters: This is the first major federal whistleblower case built specifically around AI governance failure at a top-tier academic medical center — exactly the cautionary tale that moves cautious, exploring-stage leaders toward (not away from) structured governance

ACC Comments: 

I find it quite annoying that quite a few authors and organizations publish on governance and regulation as well as ethics but few of those authors are actively on the front lines of AI in healthcare. While the allegations against Mayo are unproven and Mayo will answer them, the shape of the complaint is instructive regardless of verdict. Models degrade; that is expected but  manageable (and should be managed). Governance rarely fails at detection as someone almost always knows. It fails at escalation: the distance between a person who sees a problem and a body empowered to act on it. When that distance is traversable only at personal cost, the organization has not built governance. It has outsourced governance to individual courage, then discovered that courage files in federal court.

The remedy is not novel. We solved this decades ago for patient safety: no-fault event reporting, anonymous submission, just culture, and a register visible to the board. In my opinion, AI concerns should enter that same pipeline. An AI safety event should be reportable the way a medication error is reportable - routinely, without heroism or retribution. Governance is not what you publish. It is what happens to the person who invokes it.