What Physicians Actually Use AI For Isn’t What You’d Guess

When it comes to the healthcare sector, AI adoption is already driving productivity and efficiency gains. And physicians are leading the way. A new AMA survey confirms that AI has moved from pilot project to daily habit — 81% of physicians now use AI professionally, more than double the adoption rate from 2023. But the number worth studying isn’t the topline. It’s the breakdown of what physicians actually doing with AI.

The 2026 Physician Survey on Augmented Intelligence, run by the AMA’s new Center for Digital Health and AI, polled nearly 1,700 physicians across specialties and practice settings. The top use cases: summarizing medical research (39%), drafting discharge instructions and care plans (30%), and documenting billing codes and chart notes (28%). Lower on the list: drafting patient portal message responses (19%) and assistive diagnosis (17%).

That ordering runs against the usual assumption, which is that adoption starts with “safe,” low-stakes work and only later moves into anything touching the clinical record. Instead, physicians are leaning hardest on AI for research synthesis and documentation, tasks that feel more central to their workflow, and less on drafting seemingly lower-stakes patient communications.

The likely explanation: those higher-use tasks share a built-in checkpoint. A physician reads and signs off on a chart note or a research summary before it becomes final. A patient portal message goes out under the physician’s name with far less buffer, so it carries more liability and is far more dependent on tone. Adoption is tracking reviewability, not perceived complexity.

This is exactly the kind of gap that shows up when organizations go straight to AI deployment and skip the step of identifying where AI actually generates value in their specific workflows. Intuition about which tasks are “safe” to automate is often incorrect, and the AMA data reinforces that in a given profession. The physicians who are getting the most out of AI right now aren’t the ones who guessed correctly; they’re the ones whose actual use patterns, whether by trial and error or deliberate evaluation, found the tasks where AI output is reviewable, high volume, and genuinely time-consuming to do manually.

That’s the core of use case identification work: mapping where a workflow has a review checkpoint, real-time cost and enough volume to matter, and then prioritizing accordingly, instead of acting based on assumption. It’s a big part of how we approach AI engagements at Netrio, and this survey is a useful reminder of why the discovery step matters as much as the tool itself.