|
3, 2, 1: Health AI Brief
Every Friday
February 27, 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. |
|
NVIDIA's second annual healthcare AI survey finds that 70% of organizations now actively deploy AI, up from 63% in 2024. GenAI and large language models are the top workload for 69% of respondents (up from 54%), and 47% are already using or assessing agentic AI. On the money side: 85% of executives say AI is helping increase revenue, 80% report cost reductions, and 85% expect AI budgets to rise this year — with 46% planning increases above 10%. Payers and providers cite administrative workflow optimization as their top ROI use case. So what?
The headline is 70% adoption, but the real story is where the AI dollars are being spent. For payers and providers, admin workflows (scheduling, documentation, claims processing) top the ROI rankings. Clinical AI will likely catch up, but it's not there yet. CMS announced the Medicare App Library — a curated directory of digital health tools for Medicare enrollees, with a full Medicare App planned before July 4, 2026. Apps must clear a six-step process: sign an interoperability pledge, partner with ID.me or CLEAR for identity verification, connect to CMS Aligned Networks, pass evaluation by the Digital Medicine Society (DiMe) or CARIN Alliance, and survive CMS review. Three initial focus areas: eliminating manual check-in forms via FHIR-based data exchange, conversational AI assistants, and diabetes and obesity management tools. So what?
This is the government building an "app store" for seniors. What's notable is the vetting bar: CMS is requiring independent evaluation (DiMe Seal or CARIN Alliance), identity verification through ID.me or CLEAR, and interoperability pledges before anything goes live. That's a higher standard than Apple or Google set for health apps in their stores. Knowing this world, I'd assert that every health tech company would happily jump through these hoops if it means getting access to a 68 million member distribution channel! Chris Klomp, Director of the Center for Medicare and senior HHS counselor, told the AMA's national advocacy conference this week: "There should be no human working on prior authorization, period." He urged physicians to pressure their tech vendors to standardize the process and said reform should come in "double-digit months." The very next day, Elevance Health (#20 on the Fortune 500, $197.6 billion in revenue) detailed its own AI playbook: AI automates approvals and claims processing, and handles call center wrap-ups across roughly one million monthly interactions. But claim denials remain exclusively under human review. As Elevance Health's chief digital information officer, Ratnakar Lavu, put it: "If there is some additional kind of review that needs to happen, we'll actually send it to a clinician." So what?
The federal government says automate everything. The nation's largest payer says yes, just NOT denials. Context: MA insurers made 52.8 million prior auth determinations in 2024, per KFF. Of the 4.1 million that were denied, only 11.5% were appealed — but 80.7% of those appeals were overturned. Until this math improves, it's very hard to defend total automation here. |
|
Mount Sinai researchers tested OpenAI's ChatGPT Health (launched January 2026) across 60 clinical scenarios in 21 medical specialties, each with 16 contextual variations for race, gender, and access barriers (960 total interactions). Three independent physicians established ground truth using 56 medical society guidelines. The result: ChatGPT Health under-triaged more than half of cases physicians deemed emergencies. It handled textbook situations well — stroke, severe allergic reactions — but failed on nuanced cases requiring clinical judgment. Its suicide-risk alerts were inverted relative to clinical risk, appearing in lower-risk scenarios while missing cases where users described specific self-harm plans. Why it matters
More than 230 million people already use ChatGPT for health questions weekly, per OpenAI. This study suggests the tool is least reliable precisely when the stakes are highest during emergency conditions. While this will change over time, patients should know that current models are strong on textbook presentations, but weak on ambiguity. Researchers tested a vision-enabled AI scribe (built on Google's Gemini model and Ray-Ban Meta smart glasses) for documenting medication histories, a task that requires both hearing and seeing (pill bottles, labels, dosing devices). Across 100 simulated medication history interviews with 10 clinical pharmacists (2,160 data points), the vision-enabled scribe achieved 98% overall accuracy. Audio-only processing hit just 81% — a 17-percentage-point gap driven almost entirely by omission errors: 10 omissions with video versus 358 with audio alone (p<0.001). Accuracy ranged from 96% for patient details to 99% for dosing directions. Why it matters
Every major health system investing in ambient AI scribes today is typically deploying audio-only tools. This study says those tools are missing roughly one in five medication data points — because they can't see what the patient is holding. It's a simulated study (not yet real-world), but the 358-versus-10 omission gap is hard to ignore. Medication errors are a leading cause of preventable harm; documentation accuracy is the first line of defense. (Side note: I love that the code is open-sourced and available on GitHub.) |
|
The Role of the Human in an AI-first World
AI growth is unrelenting. NVIDIA's survey finds 70% of healthcare organizations now deploy AI. The key question now: where does the human fit in? One battleground is prior authorization. HHS wants to move towards a world with AI-only, and no humans. Elevance disagrees — at least on denials, where 80.7% of appeals get overturned. For even more serious areas, the research reinforces the need for humans. ChatGPT Health under-triaged more than half of emergencies, and audio-only AI scribes missed one in five medication data points. The pattern that's emerging: AI is earning its place in administrative automation, where the cost of error is a delayed claim. It's proving useful as an augmentation layer, where clinicians stay in the loop. But it's not ready to be the last line of defense in safety-critical decisions — triage, denial adjudication, anything where a wrong call could mean serious harm. Takeaway
As AI growth is unabated, the question for health leaders is where the human fits. AI running entire admin workflows — good. AI augmenting clinicians, not replacing them — good. Humans owning the most critical safety judgments — essential. These lines will shift, perhaps very rapidly, so tracking these lines carefully is now part of our job. |
|
|
Know someone who'd find this useful? Forward to a Colleague |
The Human Role in an AI-first World
3 signals, 2 studies, 1 insight for health leaders
Feb 28, 2026
•
1 min read