🤖 When AI Stopped Asking for Permission

Healthcare AI has crossed a threshold that does not get enough attention. The tools entering clinical workflows right now are not just smarter scribes or better search engines. They are systems that initiate actions, coordinate across platforms, and complete tasks without waiting to be asked. This week, we look at what that shift actually means for clinicians, what the evidence says, and where the governance gaps are.

In today's newsletter:

Plus: 5 use cases for Agentic AI that can help your clinic run smoother

Let’s dive in.

To understand why agentic AI is different, it helps to understand what has changed under the hood. Traditional AI tools, including the ambient scribes, are reactive. You trigger them, they respond, the interaction ends. An agentic AI system works differently: it is given a goal, and then it figures out the steps required to achieve it. It can use tools, query databases, send requests to external systems, evaluate its own outputs, and course-correct along the way. In clinical terms, the difference is something like this: a reactive AI is the medical student who writes what you dictate. An agentic AI is the senior resident running the care team while you are in the OR.

Agentic systems are built on the same large language models driving today's AI assistants, but they are paired with something called a reasoning loop. The model receives a goal, breaks it into sub-tasks, executes each one using connected tools (EHR APIs, scheduling systems, payer portals, lab interfaces), evaluates whether the result moved it closer to the goal, and repeats until the task is complete. Some systems add memory layers so they can learn from prior interactions, and multi-agent architectures where one AI supervises and checks the work of another.

What makes this relevant for clinicians right now is where these systems are being pointed first. Not at diagnosis or treatment planning, but at the administrative layer that consumes roughly forty percent of the clinical day: prior authorization, referral coordination, inbox triage, insurance verification, scheduling, documentation routing. These are high-volume, rule-based, cognitively draining tasks that AI handles well and that carry real consequences when they go wrong. The promise is significant. So are the governance gaps. As you read through this week's news and research, the pattern that keeps surfacing is this: the technology is moving faster than the frameworks designed to oversee it.

And here’s the best part: with the technology we have today, clinicians no longer have to rely on third party companies taking a piece of our earnings, we can do it ourselves! Stay tuned to the end and I’ll show you how.

On June 12, 2026, the US Commerce Department issued an emergency export control directive forcing Anthropic to disable Claude Fable 5 and its restricted sibling Mythos 5 globally, just 72 hours after release, citing national security concerns over the model's ability to identify previously undiscovered software vulnerabilities. The directive suspended access for all foreign nationals, including non-citizen Anthropic employees, and Anthropic complied immediately. Healthcare institutions in the UK and Europe that had already begun piloting Fable 5 for clinical and research applications lost access overnight.
What it means: AI tools that health systems depend on can be removed without notice by government order. Any organization building clinical workflows on top of frontier AI models needs a contingency plan for exactly this scenario.

The biggest bets on healthcare AI agents this month are not coming from clinical AI startups. They are coming from enterprise platforms like UiPath, Salesforce, Microsoft, and Oracle, all of which have existing EHR integrations, payer connectivity, and workflow hooks that take years to build from scratch. UiPath's ViVE 2026 launch and Salesforce's partnerships with HealthEx, Verily, and Viz.ai are the latest examples of clinical AI being layered on top of infrastructure that was already embedded in health systems. A startup may have better clinical logic; an enterprise platform has the connections to actually run it.
What it means: Ask not just "does this work?" but "can it connect to everything it needs to connect to?" Integration debt is where most AI pilots quietly fail.

ARPA-H's ADVOCATE program will fund two systems: a patient-facing agent that autonomously adjusts cardiovascular medications, appointments, and lifestyle recommendations, and a supervisory AI overseer to monitor it, with FDA engagement built into the process from the start rather than treated as a post-development hurdle. The evidence base for agentic AI in complex chronic disease management is essentially nonexistent at scale, but ARPA-H is betting that building under a rigorous regulatory framework is better than waiting for the evidence to catch up. We are about to learn a lot about what autonomous clinical AI can and cannot safely do in real patients.
What it means: The authorization pathway FDA develops here will shape how agentic clinical AI is regulated across every specialty. Clinicians in all fields have a stake in watching how this unfolds.

npj Digital Medicine, 2026

Researchers reviewed five databases and found seven studies meeting criteria for agentic AI in clinical care, spanning emergency medicine, oncology, radiology, and rehabilitation. Performance results were strong: high accuracy in cancer diagnosis, treatment planning, and workflow optimization. One detail worth reading closely: only one of the seven studies involved real patients. The others used retrospective data or simulated environments. Read the review.

Key Finding: Agentic AI architectures reduced clinical cognitive workload by up to 52% compared to traditional decision support, but prospective validation in actual clinical settings is essentially absent.

Clinical Implication: Strong retrospective numbers are a starting point, not a finish line. The right question to ask vendors is not "what was your accuracy on the test set?" but "where has this been validated prospectively, in what setting, and what happened when it was wrong?"

Microsoft / The Health Management Academy, New England Journal of Medicine, January 2026

Microsoft and The Health Management Academy published original NEJM research on whether health systems are actually ready to deploy agentic AI at scale. The short version: most are not. Leaders expressed confidence in AI's potential while simultaneously lacking the governance structures, data infrastructure, and change management capacity that safe deployment requires. Fewer than one in three organizations had a formal AI governance committee.

Key Finding: The limiting factor for agentic AI adoption is not the technology. It is organizational readiness.

Clinical Implication: If your organization is deploying AI tools without a governance structure, you are absorbing more risk than you may realize. Clinicians who push for evaluation frameworks before go-live, not after, are doing something that protects patients and their institution's liability exposure.

In January 2026, the FDA released revised oversight guidance exempting most clinical decision support tools from medical device regulation, as long as clinicians can independently review the AI's output. The line between "presenting options" and driving decisions is blurrier in practice than the regulatory text suggests: if you are seeing forty patients a day and accepting the top recommendation nine times out of ten, the distinction matters less than it appears. A clinician-focused breakdown is worth reading for the practical implications.
Why This Matters: More AI tools will reach clinical workflows without formal federal review. That shifts more due diligence responsibility to health systems and, by extension, to clinicians. What was true of the regulatory framework last year may not be true next year as these systems take on more autonomous functions.

What It Does: Glass Health combines ambient documentation with real-time clinical reasoning in a single interface. It captures patient-clinician conversations, generates structured notes, and surfaces differential diagnoses and relevant guidelines during the encounter, not as a reference you open separately, but as a layer on top of the visit itself.

Best For: Outpatient clinicians (internists, hospitalists, primary care) who want something beyond a documentation tool and are looking for AI that contributes to clinical thinking, not just charting.

One Practical Use Case: During a complex undifferentiated presentation, Glass Health's reasoning engine works through the differential as you talk, flagging diagnoses and guidelines in real time. By the time the visit ends, you have both the note and a structured clinical summary that reflects the reasoning you actually did, rather than reconstructing it from memory afterward.

One Limitation: Designed for outpatient settings. In fast-moving environments like the ED or ICU where the clinical picture shifts mid-conversation, the tool's performance is less established.

The scribe metaphor has served healthcare AI well. It gave clinicians a mental model for what these tools do: they record, they assist, they reduce friction. And it kept the role of judgment clearly with the clinician. That framing made AI feel manageable.

Agentic AI requires a different mental model, and we do not quite have one yet. These systems are not scribes. They are closer to autonomous coordinators, initiating actions, navigating systems, completing tasks that used to require a human at each step. That is genuinely useful for the administrative burden that consumes clinical time. It is also a real governance challenge, because the moment AI moves from recording decisions to driving workflows, questions of accountability get a lot more complicated.

What I keep coming back to is this: the clinicians who will have the most influence over how agentic AI operates in healthcare are the ones who engage with it now, not by adopting every tool uncritically, but by asking the questions that actually matter. Where was this validated? Who is responsible when it is wrong? What does our governance structure look like for overseeing something that does not wait for permission? Those are not obstructionist questions. They are the right ones.

Send me an email with any feedback or questions and I will get back to you. What topics would you like to see me cover?