AI Can't Fix the Emergency Room — And That's the Point

The Emergency Department Has a Structural Problem

Emergency departments across the United States have become the last-resort safety net for a health care system that too often fails its patients long before they arrive at the doors. Boarding holds that can stretch for days, hallways packed with patients waiting for beds, clinicians drowning in documentation, and support staff spread impossibly thin are not anomalies. They are the default. In this week's AI Prognosis column, Brittany Trang argues that artificial intelligence tools — no matter how sophisticated — will not fix the emergency room, because the emergency room is not a software problem. It is a symptom of deeper, more stubborn failures in how health care is organized, financed, and delivered.

That conclusion runs against the grain of a booming AI health care landscape. In recent years, hype has reached fever pitch for AI scribes that listen to patient-clinician conversations and automatically generate clinical notes, for predictive algorithms that flag sepsis risk or trauma severity, and for computer vision that promises to read chest X-rays and CT scans with superhuman accuracy. The emergency department, ripe with time pressure, voluminous data, and high stakes, has looked like the perfect proving ground.

Why AI Appeals to Emergency Medicine

The appeal is understandable. Emergency medicine is one of the most documentation-heavy specialties. A typical shift produces enormous amounts of charting, sign-outs, billing codes, and quality metrics. Physicians often spend two hours on after-hours charting for every hour of direct patient care. AI scribes promise to reclaim those hours, allowing doctors to focus on patients instead of screens. Triage tools use machine learning to prioritize patients based on risk scores, potentially reducing wait times. Sepsis prediction models can alert clinicians to subtle early signs, akin to an automated glance over the shoulder.

Yet these tools, Trang notes, operate at the edges of the system. They do nothing about the fact that the ER is the only place many people can access care at night, on weekends, or without insurance. They do not create inpatient beds for patients who need admission but have nowhere to go because the hospital is at capacity. They do not relieve the shortage of nurses, psychiatric staff, or social workers. They do not stop the flow of patients who have been unable to get an appointment with a primary care physician, or whose needs are not medical but social — housing, food security, transportation.

AI Scribes Address the Symptom, Not the Disease

Take the AI scribe, the darling of hospital administrators. It is a perfect example of a "dumb problem" solution: automating the mundane, repetitive task of note-taking. In the process, during the clinical encounter, the tool may indeed reduce burnout and restore a semblance of eye contact between doctor and patient. Doctors using such tools report less fatigue and higher satisfaction, at least initially. But the deeper causes of that documentation burden — useless clicks, defensive medicine, compliance-driven checkboxes, and the need to justify every minute to insurers — remain untouchable.

A system that rewards volume over value will still push clinicians to see more patients, even if charting is automated. A physician who is automatically relieved of typing still feels the pressure of a crowded waiting room and four-hour targets set by administrators. The AI scribe does not grant more time; it simply shifts where the squeezed time ends up. Without changes to incentives, staffing ratios, and the way care is reimbursed, the net effect may be that hospitals simply schedule more patient encounters for the same number of doctors, further accelerating burnout.

Similarly, AI-based triage tools can be helpful, but they need to be calibrated carefully and integrated into a workflow that already works. If the underlying process is chaotic, the algorithm inherits that chaos. And predictive models can encode biases present in historical data, risking worse care for already marginalized populations.

The 'Dumb Problems' Are the Hardest

Trang's column is part of the broader AI Prognosis series, which has consistently argued that the most significant barriers to AI adoption in health care are not the flashy algorithms but the unglamorous, boring "dumb problems": data interoperability, alert fatigue, user interfaces, and regulatory clarity. A former ARPA-H director's startup profiled elsewhere in STAT is reportedly tackling these exact issues, acknowledging that AI's biggest wins in medicine will likely come from rebuilding the digital plumbing rather than making grand predictions.

In the emergency department context, those dumb problems are everywhere. The natural language processing that powers narrative scribes must accommodate every accent, dialect, and medical idiom. The interfaces must work on old computers at a cluttered desk. The AI outputs need to flow into electronic health records from multiple vendors, each with its own structure. And the alerts have to be tuned so that clinicians do not ignore them entirely due to alert fatigue. Solving those problems is essential, but even then it only makes a rough surface slightly smoother.

What will not be solved is the problem of boarding. Emergency departments across the country report record-high wait times because patients cannot be moved to inpatient units. This is often due to a shortage of hospital beds, not because admissions are being delayed by paperwork. If a physician has a critically ill patient needing ICU care but no ICU bed is free, no AI tool can create one. Similarly, when a psychiatric hold waits for days for a bed in a mental health facility, AI will not house that patient or widen the psychiatric workforce.

Some Problems Have No Technological Fix

The crunch in emergency medicine also reflects broader workforce shortages. Nurses have left the profession in droves, exacerbated by pandemic-era burnout and poor working conditions. Physicians are retiring early or seeking jobs that promise a better work-life balance. Streaming an AI scribe into whatever nurses and doctors remain may reduce some stress, but it does not rebuild the pipeline of trained professionals. The systemic fix involves larger investments in education, restructuring hours and pay, and creating safe staffing ratios — tasks that are political, not technical.

Then there are the patients themselves. Emergency departments increasingly serve as the repository for unmet social needs. An elderly patient without a caregiver who has a fall; a young adult with addiction but no care plan; a working family experiencing homelessness. AI cannot arrange home care, add affordable housing, or ensure access to mental health counseling. These cases are not "edge cases" in a well-functioning system; in many hospitals they represent a substantial fraction of daily attendees. The EMR data, the diagnostic codes, and the narratives all reveal patterns that are social determinants of health, not defects in the technological flow.

The True Path Forward

Trang's conclusion is not that AI is worthless in the emergency room. On the contrary, the column allows that some tools can improve documentation accuracy, reduce routine errors, and help clinicians with information retrieval. But these are supportive roles. The expectation that AI can restructure and revive an ER is misplaced. The problems that plague emergency departments are the same problems that plague health care writ large: fragmented financing, a shortage of primary care and mental health services, underfunded public health, and an aging population, all layered atop an archaic operational model.

Efforts to "fix the ER with AI" may even be counterproductive if they cause leaders to avoid the harder conversations. Administrators can point to a new AI dashboard as a sign of improvement, while in the same breath ignoring the need to increase nurse ratios, improve boarding protocols, and work with community organizations to divert non-urgent care. The emergency department is a pressure relief valve for the entire health care ecosystem, and it is close to bursting. No amount of clever technology will change that; only a commitment to change the structure of care delivery can.

The lesson from AI Prognosis is that we should stop asking whether AI can fix the emergency room and instead ask why we expect technology to solve what only policy, leadership, and funding can solve. Let the algorithms assist, but do not let them carry the entire weight of a system in need of support.

As Trang writes, AI cannot change the ER's problems. That is not a failure of science. It is a reminder that health care is a human system, and its hardest problems remain stubbornly human.

This article is based on reporting by STAT News. Read the original article.