Matt dives into a specific healthcare topic to help those in the industry, and those outside of it, better understand the market drivers causing today’s healthcare challenges.

On August 5, STAT reported that federal health regulators invited industry and lobbyists to closed-door meetings on clinical AI. Staff from FDA, CMS, and HHS sat down with the companies building AI doctors, got a firsthand look at the products, and started shaping the policy that will govern them. The meetings ran invite-only. No public notice went out. Everyone agreed to Chatham House rules, so participants can describe what happened but cannot say who said it. Then the group opened a one month sprint to write consensus principles for evaluating patient-facing clinical AI.

Read that last sentence again. Principles for evaluating patient-facing AI. Drafted in thirty days. By the people who sell it.

Somewhere in Washington state, a Medicare patient waits two to four times longer than she used to for a procedure, because an algorithm now screens her prior authorization request. Her hospital reported that delay to Senator Maria Cantwell’s office in April. She did not get an invitation either.

Pew Research Center surveyed 3,488 US adults from June 22 to June 28 of this year. Seventy-two percent said it matters extremely or very much that a provider tells them when AI touches their care. More than eight in ten want notice when AI analyzes their scans, generates a diagnosis, or explains their lab results. The demand does not stop at the high-stakes uses. Seventy-two percent want to know when AI takes notes during the visit. Sixty-four percent want to know when it handles a refill. Fifty-six percent want to know when it books an appointment.

Then Pew asked about control, and the floor gave out. Fifty-three percent of Americans say they have not too much say or no say at all over whether a provider uses AI on them. Sixty-three percent want more input, roughly three times the share who feel fine with the input they have. Forty-six percent cannot tell you whether AI has already touched their care. Among the people who know it did, only 22 percent understood how.

KFF found the same population reaching for these tools anyway. In a poll fielded February 24 through March 2, across 1,343 adults, 32 percent said they had used an AI chatbot for health information in the past year. Seventy-seven percent said they worry about the privacy of medical information they hand to AI tools, and that worry includes 65 percent of the people who handed it over regardless. Read that as desperation, not enthusiasm. People with a good option do not use one they distrust.

Meanwhile the technology kept moving. The Office of the National Coordinator recorded 71 percent of non-federal acute care hospitals using predictive AI in 2024, up from 66 percent the year before. An American Medical Association survey early this year found 72 percent of physicians using at least one AI application, nearly double the 2023 share. So the tools are in the building, most patients cannot tell you whether one touched them, and the majority who want a say do not have one. Nobody built the consent infrastructure. The industry skipped that part and kept shipping.

Doximity pulled from more than 250,000 compensation responses over seven years, including roughly 23,000 US physicians surveyed last year. More than 65 percent of physicians now use AI daily or weekly. When Doximity asked who should capture the savings if AI lets a physician work faster, more than 40 percent said physicians should keep most of it. Twenty percent said share it. Seventeen percent said the savings should lower what patients pay.

Seventeen percent. Every actor in that survey is behaving rationally inside the incentives somebody handed them, and the incentives point away from the patient at every branch.

The money confirms it. Business Group on Health surveyed 127 employers covering 8.7 million Americans and found 64 percent reporting a cost impact from providers’ AI-driven revenue optimization. Translate that from benefits-consultant into English: health systems aimed AI at the billing department and the bills went up. Those same employers project a median 9.2 percent increase in health costs for 2027. The first at-scale, documented financial effect of clinical AI in this country did not reduce anyone’s costs. It optimized the claim.

CMS ran the other live experiment. The WISeR model put AI behind Medicare prior authorization in six states. Washington state hospitals reported patients waiting two to four times longer for some affected procedures. GAO then determined the pilot qualifies as a rule subject to congressional review, and Senate Democrats filed a Congressional Review Act resolution to end it. A pilot meant to cut fraud and unnecessary spending produced a queue, and it took an act of Congress to get a look under the hood.

Now the advocacy side, and this part should sting, because we earned it.

Patient advocacy has spent three years fighting the paper version of prior authorization while the decision rules quietly migrated into code. I have sat through enough coalition calls to know exactly which slide gets skipped, and it is always the one with the technical detail on it. Most patient organizations still have no published position on clinical AI. Most cannot name the person at their partner company who owns AI deployment decisions. When those invitations went out in August, nobody on our side noticed the omission until a reporter published the story.

Under-resourced is a true description of patient advocacy and a lousy excuse. Being under-resourced explains a delay. It does not explain an absence. The organizations that represent patients did not lose a fight over a seat at that table. They never asked for the seat.

Archo has fielded the ELAVAY study every year since 2015, across more than twenty disease states, asking patient advocacy organizations and community groups how biopharmaceutical companies actually behave as partners. The pattern that shows up year after year is a gap: companies grade their own advocacy performance well, and the organizations they name as partners describe something meaningfully different. Clinical AI is about to widen that gap faster than any issue we have measured, because it moves at product-release speed while advocacy relationships move at annual-planning speed.

Stanford Medicine built patient advisory panels that review AI tools before those tools reach patients. The panels look at diagnostic suggestions and care navigation features. Patients function as co-designers instead of end users who find out later.

Look at the behaviors underneath, because the values statement is not the thing that works.

Stanford convenes patients before deployment rather than after a bad outcome. Somebody owns that panel, by name, with a calendar. The review sits at a decision gate in front of go-live, which means a tool can fail it and stall. And the process generates a record, so a year from now somebody can reconstruct who approved what and on what evidence.

Compare that to the median organization running the same activity with none of the effect. The median organization publishes a patient-centered AI principle in a values document, seats a patient advisory board that meets twice a year, and shows that board nothing with a deployment date attached. The board reviews the mission statement. Engineering ships the model. Both groups leave the meeting satisfied.

FDA is circling something adjacent. Rick Abramson, who directs the agency’s Digital Health Center of Excellence, has promised formal guidance for generative AI medical devices and floated a discussion paper proposing that regulators evaluate AI tools roughly the way hospitals credential physicians. Credentialing is a genuinely useful frame. It also happens entirely without the patient in the room, which is precisely the habit that produced the August meetings.

Three moves, each small enough to start this quarter and specific enough to assign to a person.

The consensus principles from those August meetings will surface soon, and they will carry the fingerprints of everyone who sat at that table. Companies get one narrow window to decide which side of that document they want their name on. The ones who bring patients in before the principles land will spend the next five years explaining a decision they are proud of. The rest will spend those years explaining a room they sat in with the door closed.

Take the Advocacy Influence Diagnostic. It takes five minutes, is free and anonymous, and returns an Advocacy Visibility Index and a Strategic Risk Score, showing whether your advocacy function sits close enough to the decision to matter in a fight like this one. The 2025/2026 ELAVAY Report is available now, and fielding for 2026/2027 opens in October. Reach me directly at [email protected].