|
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.
|
On June 19, Bloomberg reported that UnitedHealth, the largest US health insurer, plans to invest $3 billion in AI across 2026 and 2027. The company says it already runs more than 1,000 AI use cases, with 20,000 AI engineers and 117 large language models available to staff, and one trial has AI agents phoning physician offices to schedule patient appointments.
So what?
The biggest payer is normalizing AI that doesn't just draft and summarize; it acts, placing calls and booking appointments with no person in the loop. My bet is the "agent makes the call" model spreads across plans quickly, and draws scrutiny just as fast, given how strained payer–provider trust already is.
Read the Bloomberg report →
On June 18, OpenAI said the free default model in ChatGPT, GPT‑5.5 Instant, now performs on health questions about as well as its frontier models, with what the company reports as a 71% drop in factuality problems over two months. In OpenAI's own evaluations, physicians rated ChatGPT's answers higher than answers written by doctors across roughly 3,500 responses. The company says people already ask ChatGPT around 230 million health questions a week.
So what?
Answers to clinical questions are moving quickly from a search box or a nurse line to a chatbot that millions of people already have open. A better model for free means more patients will arrive already "briefed" by an AI. Worth noting the benchmarks here are OpenAI's own and not peer‑reviewed.
Read the OpenAI announcement →
On June 11, Wave Neuroscience announced FDA clearance of its MeRT system (Magnetic EEG‑guided Resonance Therapy) for PTSD, which it calls the first personalized, biomarker‑guided neuromodulation therapy cleared for the condition. The device uses AI‑augmented analysis of each patient's EEG to tailor a transcranial magnetic stimulation protocol to their own brain activity, instead of applying one fixed protocol to everyone. Clearance was supported by a double‑blind, randomized, multisite trial of 158 patients showing clinically meaningful reductions in PTSD symptom severity; the device earned Breakthrough Device designation in 2024.
So what?
This is autonomy of a narrower, more bounded kind, backed by a 158‑patient randomized trial: an AI reads the individual's biology and sets the treatment. This is a constrained version of "AI that acts" that's easier to be more comfortable with.
Read the announcement → | Read the Patient Care coverage →
|
Researchers at TU Dresden and the German Cancer Research Center built MIRA, an AI agent that operates autonomously inside a sandboxed electronic health record, taking histories, ordering and interpreting labs and imaging, generating differentials, and proposing treatments. Across hundreds of real emergency-department cases, it matched or exceeded physician performance on many of the conditions tested. It also ordered roughly twice as many blood tests as the doctors did.
Why it matters
This agent can string a whole encounter together on its own, not just answer narrow questions. The over-ordering, though, is interesting in that a model turned loose defaults to do more. Worth noting also that this was a retrospective simulation on past records, not a live deployment.
Read the Nature study →
Google DeepMind extended AMIE, its medical AI agent, from one-shot diagnosis into multi-visit disease management, grounding it in clinical guidelines and national drug formularies. In a randomized, blinded virtual exam across 100 multi-visit cases, specialist raters found AMIE non‑inferior to 21 primary-care physicians on management reasoning, and rated it higher on the preciseness of its treatment and investigation plans and on guideline alignment. It also outperformed the physicians on the hardest medication-reasoning questions.
Why it matters
Medical management by AI is impressive, but it's important to note this is still based on simulated exams, not real patients in a clinic. Still, this is directionally telling.
Read the Nature study → | Read the Google Research write-up →
|
The week AI stopped assisting and started acting
The working model for AI in healthcare has largely been as a copilot: the AI drafts, summarizes, and suggests, and a human decides and acts. This week's updates highlight this is slowly changing.
An autonomous agent ran an entire EHR encounter and kept pace with physicians (MIRA). Google's AMIE managed disease across multiple visits. UnitedHealth said its agents are now placing calls to doctors' offices to book appointments. Even the FDA's clearance of Wave's brain-stimulation therapy is a machine reading the individual's biology and setting the treatment.
The verb keeps shifting from "suggest" to "do."
Takeaway
The ability for AI to act has arrived fast. The question for leaders is shifting from "is the AI accurate?" to "what is AI allowed to do with and without a human?" This is a non-obvious dilemma to resolve because not all AI are created equal, and the indications can be narrow like Wave or broad like MIRA — with different implications. Ultimately, health leaders have to move faster than we typically do to keep pace.
|
|