Four things happened in the last two weeks and together they answer one question, which is who gets to say what counts as qualified. The FDA proposed testing generative AI the way medicine tests a new physician, and opened it up for public comment. The AMA and the Digital Medicine Society wrote down five responsibilities that belong to physicians, one of which is deciding which tasks stay with physicians and which go to everybody else. Epic gave its customers a way to build AI agents in an afternoon and told them to bring their own policies. Google and Zocdoc made it possible to find a practitioner and book the appointment without ever leaving a chatbot. Not one of the four covers the AI your client actually used last night, and only one of them is asking anybody outside the room what they think.
Signal 1
The FDA Proposed Testing Generative AI the Way Medicine Tests a New Physician
On August 18, 2026 the FDA published a discussion paper on how it might regulate medical devices built on generative AI. The paper comes from the Digital Health Center of Excellence inside the Center for Devices and Radiological Health. It lays out a two-axis way of sorting risk, and it proposes something the agency has not proposed before: before a product goes on sale, it has to pass a competency assessment, which the paper says was inspired at a high level by how physicians are trained and evaluated. There are two parts to that, benchmarking the system against set tasks and then confirming in clinical use that it does what it claims. The paper also asks how closely these products should be watched after release, and it flags foundation models and agentic systems as problems it has not solved. This is not a rule and it is not guidance. It is a list of questions, filed under docket FDA-2026-N-7874, open for comment until October 19, 2026. The FDA says who it wants to hear from: device manufacturers, clinicians, consumers, researchers, and the public. Six days later, on August 24, the director of the Digital Health Center of Excellence told STAT that formal policy guidance is the goal, and that broad guidance on generative AI and narrower guidance on specific topics are both expected.
Sources: FDA news release, August 18, 2026 | STAT, August 24, 2026
What this means for you
Read the word competency, because it decides what actually gets tested. The FDA is not proposing to look inside the code. It is proposing to check whether the system can do the job before anyone is allowed to sell it, roughly the way a training program checks whether a person can do the job. Issue 012 covered the Joint Commission's first healthcare AI certification and made the point that what got graded was the governance around the tool, not the tool. Now both ends are being graded, the system and the practice that runs it. Here is the part that matters on a Tuesday afternoon. All of this applies to medical devices, and a medical device is something sold to diagnose or treat. The app your client opened at eleven at night is not one. When she tells you she asked an AI about her exhaustion and it told her to go get her ferritin checked, that tool was tested by nobody, is not covered by this paper, and will not be covered by whatever guidance comes after it. The products getting a competency review are the ones sold into hospitals. The one in your client's hand is the one nobody is checking, and you are the coach who hears what it told her.
Signal 2
The AMA Wrote Down Five Things That Belong to Physicians, and One of Them Is Deciding What Belongs to Everyone Else
On August 20, 2026 the American Medical Association and the Digital Medicine Society put out a framework defining the physician's role in what they call the digital and AI era of medicine, with a companion roadmap. The framework lists five responsibilities it calls enduring. Physicians hold trust and stay responsible for decisions that fit a person's goals and circumstances. Physicians exercise clinical judgment. Physicians help shape technology-enabled care models and, in the framework's own words, define the tasks that must remain in physicians' hands rather than other healthcare practitioners. Physicians make sure the tools get used safely, effectively and fairly. Physicians look after the profession and train the people coming after them. The document says individual tasks will change as the technology changes while these five stay remarkably stable, and that the goal is not to keep every task physicians do today. The roadmap hands work to the people receiving care, to technology companies, to policymakers and to payers as well. It came out the same week as a viewpoint in JAMA, written by researchers at the University of Pennsylvania's Department of Medical Ethics and Health Policy with the chief executive of Curai Health and the founder of Khosla Ventures, arguing that AI on its own will probably beat both physicians alone and physicians working with AI on thinking tasks, and that the gap will grow because physicians are losing skills to the software.
Sources: AMA and DiMe framework, August 20, 2026 | TechTarget, August 20, 2026
What this means for you
Four of the five are about what physicians owe the person sitting in front of them. The third one is different, because it points sideways instead of at the software. It says physicians define the tasks that must remain in physicians' hands rather than other healthcare practitioners. You are other healthcare practitioners. Issue 013 covered the AMA opposing autonomous AI in coverage and prior authorization decisions, and argued that professional bodies move first to protect their own seat at the table. The third responsibility is doing the same thing, except the line it draws is between physicians and you. This is a document that gets quoted at you, so it helps to know what it says before that happens. When a client asks whether she ought to be seeing a doctor about this instead of a coach, or a physician's office wants to know why you are asking about her thyroid panel, the answer people reach for comes out of papers like this one. Nothing in it says you may not do what you do. What it does put in writing is that physicians decide where the line sits. Read the five responsibilities once, so the next time the question comes up you are answering from the text itself instead of from what somebody told you it said.
Signal 3
Epic Gave Its Customers a Factory for Building AI Agents, and Told Them to Bring Their Own Policies
At its Users Group Meeting on August 18, 2026, Epic showed Ergo Visit, which pulls together information from the health record, the MyChart portal and Cosmos, Epic's deidentified research database, and turns it into takeaways a clinician reads before an appointment. Judy Faulkner described it as human intelligence and AI together, with people bringing judgment, common sense, empathy and intuition, and the software bringing pattern recognition across very large amounts of data. One piece of it, called Best Care Choices for My Patient, runs against 320 million deidentified records in Cosmos and tells a clinician how common a diagnosis is among similar cases and what treatments those cases got. Sixty early-adopter organizations have used it, and Epic says they sometimes found the most commonly prescribed medication was not the best one for that type of person. Ergo Visit is running now at Ochsner Health and with five physicians in Epic's Garden Plot collaborative. Epic also showed Agent Factory, which lets an organization switch on one of 120 ready-made AI features, adjust it with its own policies and reference material, or build a new agent without writing any code.
Source: TechTarget report on Epic UGM, August 18, 2026
What this means for you
The phrase to underline in Agent Factory is customize it with your own policies. That is a vendor handing the governance question to the buyer along with the software. Issue 020 covered a health system that passed the first AI governance certification in three weeks after four years of keeping records, and made the point that responsibility follows whoever put the tool in place. Building something like this used to take a project and a budget. Now it takes an afternoon, and whoever builds it owns every decision about what it is allowed to do, whether or not anybody there has the job of writing those decisions down. You are never going to buy Epic, so here is where this reaches you instead. The clinicians who refer to you work inside it, and the summary they read before your client walks in is now put together by software rather than written out by a person. You will see it in referral notes and visit summaries that read smoothly, sound confident, and are slightly wrong. Your client tells you she has been worse since March and the summary says she reports improvement. When the paperwork and your client disagree, believe your client, and write down what she actually said somewhere the next person will see it. You are usually the only one in that chain who sat with her for forty minutes.
Signal 4
Google and Zocdoc Made It Possible to Find a Practitioner and Book the Appointment Without Leaving the Chatbot
On August 12, 2026 Zocdoc and Google announced that people using Gemini can now book appointments inside the chatbot itself, using Gemini's agentic features and Zocdoc's network of more than 200,000 healthcare providers. Somebody can describe what they need, have Gemini sort providers by specialty, location and what is available, and finish the booking without opening another site. Google's chief health officer described it as going from knowledge to action and access. Zocdoc's founder said people increasingly start their health journeys with AI. The companies point to a June 2026 survey by rater8 which found 47 percent of people have used an AI chatbot to look for a provider, up from 31 percent at the end of 2025. Zocdoc has been building toward this for a while. Healthgrades added its booking button at the end of 2025 and Yelp added one in April 2026.
Source: TechTarget report on the Zocdoc and Gemini deal, August 12, 2026
What this means for you
The number to hold onto is 47 percent, and how fast it got there. Eight months earlier it was 31 percent. Looking for a practitioner is turning into a conversation with a model instead of a page of search results, and the model is the one deciding who gets mentioned. Issue 021 covered a poll finding that one adult in five keeps something back from a doctor out of shame or mistrust, and that some of them take the question to a chatbot instead. Now the same tool that hears the embarrassing question also answers the question of who to go see about it. This one decides whether you come up at all. Somebody types that she is exhausted and bloated and wants somebody who will actually listen, and the model replies with names. It can only give names it can read about somewhere. A Linktree, an Instagram bio and a booking button hand it almost nothing, because there are barely any sentences in them. Somewhere on a page with your name on it there should be plain sentences saying who you work with, what they usually come to you for, and where you are, written the way a client would say it rather than the way it reads on your certificate. Nobody has to visit that page for it to do its job.
The Pattern
Put the four next to each other and the same gap turns up in every one. The FDA is working out what competent means for AI sold to hospitals as a device. The AMA and DiMe are working out which work belongs to physicians. Epic is handing the governance question to whoever buys the software. Google and Zocdoc are working out who gets named when somebody asks where to go. Issue 018 covered a mid-year count of 84 new state AI laws and argued the rules were showing up faster than anyone could read them. These four are not rules. They are definitions, which is why none of them comes with a deadline or a penalty attached, and why they are easy to miss right up until the day somebody quotes one back at you. The tool your client opened last night is in none of them. Neither are you. The FDA is the only one of the four that left a door open, and that door shuts on October 19.
One Thing You Can Do This Week
Write to the FDA. Docket FDA-2026-N-7874 is open at regulations.gov until October 19, 2026, and the FDA's own announcement lists consumers and the public right alongside device manufacturers and researchers. You do not need a lawyer and you do not need to answer the whole paper. Write about what you actually see. A client came in believing something an AI told her, here is what it said, here is what was wrong with it, here is what it took to work through it with her, and here is the thing nobody testing these systems inside a hospital would have thought to ask about. That is two paragraphs and it is a real comment. Put your name and your practice on it. Comments are public and the agency has to consider them, which means a coach who writes one is in the record and a coach who does not is not. If you would rather not write, read the paper anyway and mark the questions you could have answered. That will show you where your work sits next to a definition being written without you.