Every eight minutes, someone in an American emergency department is harmed by a boarding-related delay. That statistic, drawn from a growing body of research linking ED overcrowding to a 1.3-fold increase in 90-day mortality, is a patient safety emergency unfolding in real time across the country. Emergency departments are absorbing record volume. There will be a 28% increase in emergent visits among adults over 65 over the next decade, a 12% surge in behavioral health presentations, and a 5% baseline increase in overall inpatient utilization. The traditional triage model was never designed to withstand this kind of pressure. More than half of all ED visits in the United States are currently triaged to ESI Level 3, producing an enormous pool of undifferentiated patients that undermines the entire purpose of acuity sorting. Artificial intelligence offers a way forward.

Why the ESI Alone Is No Longer Enough

The Emergency Severity Index has been the backbone of American ED triage for over two decades. It works, but its limitations have become impossible to ignore as patient volumes and acuity have escalated. ESI relies heavily on individual clinician judgment and intuition. Inter-rater reliability ranges from 0.46 to 0.91, a spread wide enough to mean that the same patient could be classified as a Level 2 at one hospital and a Level 3 at the next. When a triage nurse is managing eight patients in the waiting room, fielding ambulance radio calls, and processing a behavioral health patient in crisis, the cognitive load compounds. Crowding, interruptions, knowledge gaps, and time constraints all degrade triage accuracy in predictable ways.

The most damaging consequence is the Level 3 bottleneck. Because ESI's middle tier captures such a heterogeneous group, from the stable ankle sprain to the chest pain patient whose ECG has not yet been run, patients who need rapid intervention can sit for hours alongside those who could safely wait. Machine learning models trained on clinical data have demonstrated the ability to disaggregate this middle tier with significantly greater precision.

None of this means AI should replace the triage nurse. Nurses who agreed with AI-informed triage recommendations more often outperformed both the AI system alone and nurses who frequently disagreed with it. AI triage works best as a decision-support layer that sharpens clinical judgment rather than substituting for it.

Specialty-Specific Triage: Where One Size Fails

Pediatric Emergency Departments

Children are not small adults, and pediatric triage reflects that reality. Vital sign norms shift with age. A heart rate of 140 in a two-year-old is unremarkable; the same reading in a twelve-year-old demands immediate attention. Developmental status affects history-taking; a nonverbal toddler cannot describe the character of their abdominal pain, forcing the triage clinician to rely more heavily on parental report and physical observation. Demand models trained on or validated against pediatric-specific datasets, and ensure that age-adjusted vital sign parameters are embedded in the algorithm's logic rather than bolted on as an afterthought.

Psychiatric and Behavioral Health Presentations

Behavioral health patients already experience the longest ED lengths of stay, an average of nine to ten hours compared with four to five hours for all other ED patients. These patients are also among the hardest to triage accurately using conventional tools, because psychiatric acuity does not map neatly onto the vital-sign-driven ESI framework. A patient presenting with suicidal ideation may have entirely normal vitals yet require immediate intervention.

Urgent Care and Primary Care Settings

Urgent care and primary care clinics face a different triage challenge: distinguishing patients who can be safely managed on-site from those who need emergency referral, often without the diagnostic resources available in a hospital ED. Five life-threatening conditions, 38.5% of patients whose virtual triage AI indicated emergency-level acuity had no pre-triage intent to consult a physician at all, meaning the AI caught emergencies that patients themselves would have ignored.

For primary care practices managing high patient volumes and limited staff, platforms like Sully are demonstrating how AI-driven triage can be operationalized outside the hospital. Sully's AI handles intake, symptom collection, and triage routing before each visit, evaluating severity against clinical protocols and either scheduling an appropriate appointment, directing to a nurse line, or flagging for immediate emergency attention, with safety escalation protocols ensuring that patients showing concerning symptoms are never left in an automated loop. This kind of pre-visit triage layer is particularly valuable in primary care.

Building Your Implementation Roadmap

Step 1: Define the Problem Before Choosing the Tool

The most common implementation mistake is starting with the AI and working backward to the clinical need. Start instead by identifying the specific triage pain point that is measurable, consequential, and addressable. That might be a high rate of undertriage among ESI Level 3 patients who subsequently require ICU admission or avoidable emergency referrals from your urgent care sites.

Step 2: Assess Data Readiness and EHR Integration Capacity

AI triage models are only as reliable as the data flowing into them. Before procurement, conduct a candid assessment of your EHR data quality. Models that rely on clinical natural language processing need unstructured data to be captured digitally and in sufficient volume. Integration architecture matters enormously. Systems built on interoperable standards like HL7 FHIR allow AI triage tools to pull real-time data from the EHR and push recommendations back into the clinical workflow without requiring nurses to toggle between screens. If your EHR infrastructure cannot support bidirectional data exchange with a third-party application, you are not ready for AI triage. You are ready for an EHR modernization project.

Step 3: Pilot in a Controlled Environment

The effectiveness of AI triage began with a controlled pilot. Your pilot should run for a minimum of 60 to 90 days, capture both AI-generated recommendations and nurse-assigned acuity levels, and measure outcomes including disposition accuracy, time to physician evaluation, and rates of over- and undertriage.

Step 4: Design the Clinical Workflow, Not Just the Software

Even highly accurate AI models have struggled when dropped into complex, human-driven clinical environments. The failure point is almost never the algorithm, but the workflow. Nurses need to understand what the AI recommendation means, how much weight to give it, and what to do when they disagree. The most successful implementations position the AI recommendation as one input alongside the nurse's clinical assessment. Build the workflow so that the AI score appears in the triage documentation alongside the nurse's ESI assignment. When the two disagree, create a structured protocol for escalation review. This disagreement signal, when tracked over time, becomes one of your most valuable quality metrics.

Confronting Bias and Ensuring Equity

No discussion of AI triage implementation is complete without addressing algorithmic bias, and the evidence here demands serious attention from hospital leadership. Biases in healthcare AI often originate from imbalanced training datasets and unequal data collection practices, leading to misdiagnoses and unequal resource allocation that disproportionately affect marginalized communities.

In triage specifically, the risks are concrete. If an AI model is trained on data from a hospital system that historically undertriages Black patients, the model will learn to replicate that bias. Research from the American Academy of Pediatrics has examined racial and language disparities in pediatric ED triage, identifying patterns that could be encoded into ML models unless developers deliberately work to prevent them. AI dermatology tools have performed poorly on darker skin tones when trained primarily on images of lighter-skinned patients.

For hospital leaders, the equity checklist before deployment should include scrutinizing dataset representativeness for your local patient population, identifying proxy variables that could encode racial or socioeconomic bias, requiring vendors to provide performance breakdowns by demographic subgroup, and establishing a prospective monitoring protocol that tracks triage accuracy across race, ethnicity, language, and age groups after go-live.

Navigating the Regulatory Landscape

In January 2025, the FDA issued draft guidance applying a Total Product Life Cycle approach to AI-enabled device software functions. Then, in January 2026, the FDA published additional guidance that reduced oversight of certain digital health products, including AI-enabled software and wearable devices.

The critical distinction for triage systems lies in how the tool is classified. AI tools that assist clinicians may fall outside the scope of medical device regulation if clinicians can independently review the recommendations. However, AI systems that make autonomous clinical decisions or directly determine patient routing without clinician review remain subject to full FDA medical device authority. Choose AI triage tools designed for clinical decision support rather than autonomous decision-making. This aligns with the clinical evidence showing that the human-AI collaboration model outperforms either humans or AI working alone. Ensure your vendor can clearly articulate their regulatory classification and maintain documentation demonstrating that your nurses retain independent clinical judgment for every triage decision.

Measuring What Matters: KPIs for AI Triage

Deploying AI triage without a rigorous measurement framework is like prescribing a medication without follow-up labs. The metrics you track should go beyond the vendor's accuracy claims and capture real-world clinical impact:

• Triage accuracy: Track the rate at which AI identifies patients requiring ICU admission, emergent surgery, or critical care who were initially triaged to lower acuity levels by conventional methods.

• Time-based metrics: These should be measured before and after implementation with sufficient sample sizes to detect meaningful differences.

• Safety metrics: Monitor undertriage rates and overtriage rates. Post-implementation, the high-acuity identification of patients requiring critical care rose from 78.8% to 83.1%, a specific, measurable improvement that justifies continued investment.

• Equity metrics: If triage accuracy improves overall but the improvement is concentrated among white, English-speaking, commercially insured patients, your AI system is widening disparities rather than closing them.

Clinician Adoption: The Human Side of Technical Change

There are clinicians who are open to AI but need to see evidence and feel confident that the tool supports rather than threatens their professional judgment. Training should be structured around three pillars:

• Clinical rationale: nurses need to understand what data the AI model uses, how it generates its recommendations, and why its performance envelope includes both strengths and blind spots.

• Workflow mechanics: where the AI recommendation appears, how to document agreement or disagreement, and what the escalation protocol looks like when the AI and the nurse diverge.

• Feedback loops: create mechanisms for nurses to flag cases where the AI recommendation was clearly wrong, and demonstrate that this feedback actually influences model refinement.

While AI does not consistently outperform experienced clinicians in overall triage, it demonstrated higher accuracy than nurses, specifically in the most urgent triage category. Frame this finding for your clinical staff, not as a threat but as a safety net. AI is best at catching the cases where the consequences of missing are highest.

The trajectory of AI triage in 2026 is moving from pilot-phase experimentation toward operational integration. Successful health systems are transitioning AI from experimental technology to an essential operational tool within behavioral health programs, and the same pattern is emerging in emergency medicine. The emergency departments that will weather the coming decade of rising volume and escalating acuity are the ones whose leadership teams invest now in intelligent augmentation of clinical triage. The evidence supports moving forward, and the patients waiting in your ED right now cannot afford to stand still.

Sources

• AI-driven triage in emergency departments: A review of benefits, challenges, and future directions — International Journal of Medical Informatics, 2025

• Impact of Artificial Intelligence–Based Triage Decision Support on Emergency Department Care — NEJM AI, 2025

• Artificial intelligence in emergency department triage: perspective of human professionals — Frontiers in Digital Health, 2026

• Doctors and nurses are better than AI at triaging patients — EUSEM 2026 Congress — EUSEM, 2026

• AI Could Help Emergency Rooms Predict Admissions, Driving More Timely, Effective Care — Mount Sinai Health System, 2025

• Artificial Intelligence Models for Predicting Triage in Emergency Departments — JMIR Medical Informatics, 2026

• Predicting triage of pediatric patients in the emergency department using machine learning approach — International Journal of Emergency Medicine, 2025

• Artificial intelligence: Revolutionizing pediatric emergency care — A narrative review — PMC, 2025

• A Qualitative Assessment of an AI-Assisted Psychiatric Triage System — PMC, 2025

• AI assisted triage of UK patients in mental health care services — BMC Psychiatry, 2025

• Mental health AI breaking through to core operations in 2026 — Healthcare IT News, 2026

• The potential of virtual triage AI to improve early detection and emergent care referral — Frontiers in Public Health, 2024

• Improving ED Emergency Severity Index Acuity Assignment Using Machine Learning and Clinical Natural Language Processing — Journal of Emergency Nursing

• From Every Angle: Emergency Department Overcrowding — Vizient, 2025

• Emergency Department Boarding, Crowding, and Error — PMC, 2025

• Clinical Impact of Artificial Intelligence-Based Triage Systems in Emergency Departments: A Systematic Review — PMC, 2025

• Improving Safety and Quality of Emergency Care Using Machine Learning-Based Clinical Decision Support at Triage — AHRQ Digital Healthcare Research

• The bias algorithm: how AI in healthcare exacerbates ethnic and racial disparities — PubMed, 2024

• Using Artificial Intelligence and Machine Learning to Promote Child Health Equity — Pediatrics (American Academy of Pediatrics), 2025

• A scoping review and evidence gap analysis of clinical AI fairness — PMC, 2025

• FDA "Cuts Red Tape" on Clinical Decision Support Software — Arnold & Porter, 2026

• Transforming the Emergency Department with AI — HealthTech Magazine, 2025

• Feasibility of Mental Health Triage Call Priority Prediction Using Machine Learning — PMC, 2024