New research shows AI scribes and triage models are earning real but modest wins in crowded ERs. The bigger opportunity may be the patients who never should have ended up there.
Walk into almost any American emergency department this fall, and you will find artificial intelligence somewhere in the building. It might be listening in the exam room, drafting the physician’s note. It might be scoring the urgency of a patient’s symptoms before a nurse finishes taking vitals. At the American College of Emergency Physicians’ annual meeting earlier this month, researchers asked a question that would have sounded like heresy a few years ago: is the Emergency Severity Index, the five-level triage system that has anchored U.S. emergency medicine for more than two decades, about to be replaced by an algorithm?
The honest answer, based on the evidence piling up in 2026, is: not yet, and maybe not ever. That’s exactly why this moment matters.
The wins are real. They are also small.
Start with documentation, the clerical burden emergency physicians routinely cite as a driver of burnout. A Mass General Brigham study published in Annals of Emergency Medicine in June reviewed more than 198,000 ED encounters across four sites. Ambient AI scribes cut physician documentation time by about 1.6 minutes per note. Human scribes cut it by roughly twice that. And neither one increased how many patients a physician saw per hour.
A Stanford analysis in JMIR AI found a similar pattern: about 72 seconds saved per encounter with Microsoft’s DAX Copilot. Meaningful over a long shift, but hardly a revolution. Another Annals study found that only about 11% of eligible encounters used the AI scribe, and that fewer than one in ten physicians accounted for more than 70% of the use.
Triage tells a more promising story, with an important asterisk. A systematic review published this summer found that purpose-built machine learning models consistently outperformed standard human triage instruments at predicting which patients were truly sick. General-purpose large language models, the chatbots most people picture when they hear “AI,” did worse than trained triage nurses and tended to over-triage, flagging too many patients as urgent. One review found a widely used chatbot correctly identified low-risk patients less than a fifth of the time. In a department already drowning, an AI that sounds the alarm on everyone is not a solution.
The researchers’ conclusion was almost unanimous: AI should augment human judgment, not replace it.
The ER is where the system’s failures show up
One company thinking about that question works far from the trauma bay. TDZ Functional Medicine Academy (TDZ-FMA), founded by Dr. Brandy Zachary, DC, IFMCP, trains clinicians to work up complex, chronic, hard-to-diagnose patients: the people who often cycle through multiple specialists before anyone connects the dots, and who frequently land in the ER along the way.
That vantage point exposes a blind spot in the current debate. Most emergency AI is designed to move patients through the front door faster. Far less attention goes to the upstream decisions that determine whether a patient needs to come through that door at all.
It is not an abstract concern. Emergency departments nationwide are struggling with boarding — admitted patients held in the ED for hours or days because no inpatient bed is free. Federal regulators considered the problem serious enough to finalize a new emergency care access and timeliness measure for 2026. No triage algorithm, however accurate, creates a single new hospital bed.
Generic AI versus clinical AI
TDZ-FMA has built its own clinical decision-support tools, including LabDX, lab interpretation software for practitioners, and DiagnoseThis, a consumer-facing platform. Both are positioned very differently from the general-purpose chatbots patients increasingly consult before, or instead of, seeing a clinician.
Zachary has said her first deep dive into a complex functional medicine case took more than 10 hours of manual work. TDZ-FMA’s software was built to compress that work, so practitioners spend their time on clinical judgment rather than data wrangling, the same trade-off emergency departments are now trying to make with AI scribes.
TDZ-FMA trains clinicians across disciplines; its programs are accredited by more than eight national credentialing bodies for MDs, DOs, physician assistants, nurse practitioners, nurses, pharmacists, and others. Built without outside investment, the company landed on the 2026 Inc. 5000 with 2,095% three-year growth, and Zachary won the inaugural Gold Stevie Award for Best Female Entrepreneur in this year’s American Business Awards.
What hospitals should demand
The lesson from 2026’s research is not that AI in emergency medicine has failed. It is that the technology is only as good as the question it is pointed at. Saving a minute on a chart is helpful. Flagging the septic patient in the waiting room before she crashes is lifesaving. Keeping a patient with a missed chronic condition out of the ER entirely may matter most of all, and it is the hardest benefit to measure.
Hospital leaders evaluating these tools should insist on three things: prospective, multi-site evidence rather than single-hospital retrospective studies; systems that explain their reasoning to the clinician using them; and a clear line of human accountability for every decision an algorithm informs.
Emergency medicine has always been the specialty that catches what everyone else missed. AI may finally help it catch those patients earlier. But the real test will be whether it helps the rest of the health care system send fewer of them through the doors in the first place.
