The Healthcare AI Guy's July 28 issue carries 44 items. Most of them are funding rounds and partnership announcements, which is what the genre is for and which is not what this newsletter is for.
Ten of them are worth something. They are ranked below by how much of a WAiR piece is actually in there, not by how important they are to the industry. The distinction matters, because the biggest news in the issue is a 1.5 billion dollar acquisition and it is near the bottom of this list.
The first serious liability test of a general purpose model giving health advice. A PE is exactly the diagnosis where reassurance kills, because the presentation is often unremarkable and the mortality is front loaded.
Why it is a piece: it forces the question nobody in the industry wants asked. If the model is a medical device it needs clearance. If it is a search engine it has Section 230. It is being marketed as neither and used as both.
Nursing organizations opposing emerging billing codes on the grounds that they would generate revenue from machine-produced services. This is the first organized labor position on AI reimbursement in American healthcare.
Why it is a piece: it is the PHTI report's abstract argument arriving as an actual fight between actual people. PHTI says new payment models are needed. This is what it looks like when someone proposes one.
Eighteen percent of US consumers now use AI for health information. The framing the report chooses is "first opinion," which is a genuinely good phrase and a genuinely alarming one.
Why it is a piece: the number is the denominator for the lawsuit above. Also worth checking whether 18 percent is measuring people who ask a chatbot a health question once a year, which would make it a much less interesting figure. The methodology is in the PDF.
Argues that the sixty year path NPs took through scope-of-practice regulation, payment parity, and evidence generation is the same path clinical AI has to walk. It is a venture firm's blog post and it is the best analytic framing in the entire issue.
Caveat worth stating in any use: the analogy has a hole. Nurse practitioners accumulated authority partly by accumulating professional liability. Nobody has worked out what a model is liable for.
All US adult users can now connect Apple Health data and medical records to ChatGPT. Announced in the same window as the lawsuit above, which is either bad timing or a very deliberate demonstration of confidence.
Why it matters: connecting the record changes the liability picture. Advice given without your labs is a search result. Advice given with your labs in context is something else, and the something else does not currently have a name in regulation.
Medicare proposing to restrict what third-party RPM vendors can bill for. Dry, and directly relevant to every chronic care AI business model that assumes RPM codes will fund the monitoring layer.
Why it matters: read alongside PHTI. This is CMS narrowing the one existing fee-for-service pathway that AI-enabled chronic care companies have been using, at the same moment CMS is opening ACCESS. That is not a coincidence, it is a policy direction.
Children's Hospital of Philadelphia running an internal "School of AI" with a numeric workforce target. Notable mostly because a specific percentage target is unusual and falsifiable, which most health system AI announcements are careful not to be.
A health system deploying AI symptom assessment to route patients to the right level of care. Peer-reviewed implementation write-up rather than a press release, which makes it more useful than most deployment news. Same paywall problem as the other Catalyst piece.
A million encounters is a real scale number for ambient documentation, and the claim attached to it is three hours of documentation time saved per clinician per week.
Treat the three hours carefully. It is a vendor figure in a press release. The independent literature on ambient scribe time savings is considerably more mixed, and this is exactly the sort of number that gets repeated until it becomes a fact.
The largest transaction in the issue by a wide margin. Tempus buying minimal residual disease testing capability to fold into its precision oncology platform. Important to the sector, thin as a newsletter item, included here because leaving out the biggest number would be strange.
The remaining thirty-odd items are funding rounds (Candid Health at 120 million, TytoCare at 25 million, Karoo, Prosper, Inner Logic, Cheiron), vendor product launches, and partnership press releases. Individually they are unremarkable. Collectively they are a data point, which is that the money in health AI right now is going into revenue cycle management, credentialing, prior authorization, and scheduling.
That is the observation worth keeping. Almost none of this capital is going toward treating anybody. It is going toward the paperwork between the treatment and the payment, which is a rational place to put it and a fairly damning description of where American healthcare's costs actually live.
Three stories in this issue are about the same thing without saying so. Consumers are asking AI first. A man is suing because it told him he was fine. And nurses are fighting the billing codes that would pay for the machine's opinion.
That is one section, not three items. The connective tissue is the question of what happens when a first opinion becomes a billable one.
Sources
Scanned issue: "Healthcare AI Guy Weekly, 7/28." healthcareaiguy.com
All ten items above link directly to their primary sources.