|
3, 2, 1: Health AI Brief
Every Friday
June 12, 2026
|
|
AI is reshaping healthcare fast. Below are 3 key AI developments, 2 studies, and 1 takeaway for this week to help you better lead with AI. Target read time: 5 minutes. |
|
3
Market Signals
On June 11, Abridge, the AI clinical‑documentation company, announced a partnership with Nvidia to build an AI model designed specifically around clinical interactions between providers and patients. The model will be developed using Nvidia's Nemotron family of open models and deployed exclusively within Abridge's platform, supporting tasks like clinical documentation and decision support. Nvidia is already an investor in Abridge. No financial terms were disclosed. So what?
The chipmaker that sells everyone's AI infrastructure is now co‑building the clinical model itself. Abridge earns its seat in this partnership by providing the incredibly valuable training ground from the millions of clinician‑patient conversations happening on its platform. Read the Abridge announcement → | Read the Yahoo Finance coverage → On June 9, Stepful announced a $55M Series C led by Oak HC/FT. Stepful uses AI to train people into allied health roles like medical assistants, pharmacy technicians, and phlebotomists, and has graduated more than 32,000 practice‑ready workers for its 35+ health system clients, including Mount Sinai, Ochsner, and Providence. The new funding expands its programs into registered nursing, respiratory therapy, and advanced medical imaging. So what?
Most health-AI money chases the clinical or administrative work itself. This bet is instead on the human pipeline: using AI to create healthcare workers, faster. This week the AMA detailed revisions to CPT Appendix S, accepted by the CPT Editorial Panel at its May 2026 meeting. Appendix S is the taxonomy that classifies AI‑enabled medical services as assistive, augmentative, or autonomous, and the revisions draw on over 4 years of real‑world application data. Autonomous AI now spans 3 levels of physician involvement: from Level I, where the output still requires physician judgment to implement or reject, to Level III, where the software automatically initiates management actions on its own. So what?
While FDA clearance gets the headlines, CPT classification shapes what gets paid. Level III, in particular, is interesting: the payment system now has official vocabulary for software that acts without a human signing off before action is taken. |
|
2
Research Studies
An RCT in JAMA Network Open randomized 1,297 parents across the US, UK, and Canada to a no‑message control, official public health materials, or a 3‑minute conversation with 1 of 2 GPT‑4o chatbot styles, then tracked HPV vaccination intentions. Immediately afterward, every intervention worked: public health materials raised intent (d=0.53) similar to the chatbot arm (d=0.48). By day 45, neither chatbot's effect persisted, while the static materials held a modest effect. No arm increased self‑reported vaccination uptake. Why it matters
In this study, public health materials matched the chatbot on day 1 and outlasted it at day 45. Takeaway is to ask about durability over time, not just the immediate delta, in any conversational‑AI engagement pilot. And, even then, this also calls out that intent isn't behavior, i.e., nothing here moved actual uptake. This meta-analysis pooled 48 randomized controlled trials of conversational agents for mental health, covering 28,071 participants and spanning both AI‑based and rule‑based systems. The agents produced small‑to‑moderate but significant reductions in symptoms of depression (SMD −0.27), anxiety (−0.20), and stress (−0.26). Most trials tested scripted or NLP/ML‑based bots, only 2 used generative AI, and risk of bias across the evidence base was largely low. Why it matters
Small‑to‑moderate impact at chatbot prices, with around‑the‑clock availability, matters for any plan with a months‑long behavioral health waitlist. The quality will only get better (this paper had only 2 trials with generative AI; the rest were scripted or NLP/ML‑based). And still, long‑term follow‑up data is largely lacking. |
|
1
Key Insight
The gains are real. The scaffolding isn't.
Two surveys published 8 days apart measured the impact of AI in the eyes of clinicians. In Philips' Future Health Index (2,000+ clinicians, 10 countries), 46% report saving at least 132 hours a year, more than 3 full working weeks, and 39% have watched AI catch or prevent a medical error at least 3 times in the past 3 months. In Wolters Kluwer's US survey, nearly three‑quarters of doctors now use AI at least weekly, and the share using it multiple times a day tripled in a year, from 10% to 38%. The same clinicians also described what's missing. 70% in the Philips survey call their AI training inadequate, inconsistent, or unavailable. In the Wolters Kluwer data, 74% of clinicians name deskilling among AI's greatest risks, only 27% understand their organization's approach to AI governance, and just 22% say anyone has defined where the clinician's responsibility ends and the AI's begins. Ultimately, the adoption ran ahead of the scaffolding. Takeaway
The adoption of AI by clinicians has happened; yet the scaffolding remains behind: specifically, the AI training, the governance structure, and clear accountability. Funds for companies like Stepful and the AMA's sharper vocabulary for autonomous AI will help, but we still have a long way to go. |
|
Know someone who'd find this useful? Share |
Nvidia moves up the healthcare AI stack
3 signals, 2 studies, 1 insight for health leaders
Jun 12, 2026
•
3 min read