Before we automate the consult: What Google's AMIE (Video) study really tells us

A recent announcement from Google shows a glimpse into how the latest AI capabilities can lead to a fully automated telemedicine experience. It's an incredibly impressive display of how AI can lead to better quality health care and shows that the optimistic dream of dramatically increased access to care is close to reality. 

That being said, one of their findings reinforces a key point I made my recent Medscape article: we should be thoughtful about what we automate before we decide if we should automate it. Patients rated Google's multi-agentic approach to a telehealth consult on par with human primary care physicians, except in two domains that trended toward the human physicians: building rapport and building partnership. 

Most clinicians, and many patients, will tell you that rapport – and the trust that results from that – may be the ultimate factor that triggers a health-promoting behavior, or not. Filling a prescription, starting exercise, changing dietary habits – many patients know these are good for their health, but lack the motivation or catalyzing force to carry them out. LLMs are quite good with health information given the right context. But how that information is delivered – by text, by video, at the right time, with empathy – is likely the critical link between a health need and a health action. If a patient isn't ready to hear the advice, they're unlikely to translate it into action. 
 

The gap between the rubric and the room

Dig into the full paper's appendix (not the blog post summary) and the picture gets more interesting. The independent panel of clinical evaluators — physicians scoring transcripts against validated rubrics — actually rated AMIE (Video) higher than the human PCPs on "fostering relationship" (80% vs. 69%) and on empathy. By the checklist, the AI built rapport better. But the patients who actually sat through the consult rated it a toss-up on rapport and gave the edge to the human physicians on partnership. That's the finding I'd flag for anyone building clinical AI: A validated communication rubric and a patient's felt sense of trust are not the same instrument, and right now they're measuring different things. 
 

The cameras-off detail

Here's the part that should give any health system pause. To keep the study fair, the PCPs in this trial had their cameras off, an attempt to control for bias tied to visual appearance. Google's own authors acknowledge this "reduced the ecological validity of the human baseline," since eye contact and facial expression are core channels for building trust. In other words, the human physicians were competing for rapport with one hand tied behind their back — and they still won that category. That's a more compelling signal than the raw stat suggests: rapport in this study turned out to be stubborn, not fragile. 
 

What patients actually said they wanted

Google also ran structured interviews with the patient actors after their sessions. The pattern was consistent: people were comfortable letting AI triage routine, lower-stakes symptoms, but wanted a human voice the moment things felt serious — one participant specifically described wanting a person, not an agent, when discussing an elevated blood pressure reading. That's a real-world articulation of exactly the question I keep coming back to: not can we automate a given task, but should we, and under what conditions. 
 

One more caveat worth naming

This was a simulated objective structured clinical examination (OSCE) study with professional patient actors, not real patients with real stakes in the outcome. The rapport findings are directionally useful, but they haven't yet been tested with people who have something to lose. 

Ultimately, the Google team is showing us the way forward. AI can expand access to information and care, tailored to individuals in a way the existing clinician workforce simply doesn't have the bandwidth to craft. The core question, though, is what specific tasks or functions are we going to automate with agents, and which core competencies must stick with human clinicians. For health systems building their own AI strategy, that's not a philosophical question — it's a design question, one that should shape where agents sit in the workflow and where a human is deliberately kept in the room.

Travis Bias, DO, MPH, FAAFP, is a family medicine physician and deputy chief medical officer of health information systems at Solventum.