Introduction

Emergency departments serve as the safety net for urgent and unscheduled care. They also take on failures in the health system. This includes issues such as limited access to primary care, shortages of inpatient beds, gaps in behavioral health capacity, delays in skilled nursing placement, and bottlenecks at discharge. When these pressures build up, crowding occurs. This leads to longer waits, delays in ambulance offloading, and extended boarding times for admitted patients.

The phrase “no-wait ED” is appealing but unrealistic. Emergency care is inherently variable. A low-acuity waiting room can get unsafe quickly if many critically ill patients arrive at once. Boarding may persist even when triage and front-end throughput are efficient. For this reason, the concept should be reframed. The ideal target is a low-wait ED. It identifies risks early, minimizes avoidable idle time, and escalates bottlenecks before they become dangerous.

AI-supported patient flow tools may contribute to this goal. These systems can forecast demand, estimate admission probability, identify patients likely to need specific resources, monitor waiting-room risk, and improve visibility into bed status. Smart bed management platforms can aggregate information about room availability, cleaning status, isolation needs, telemetry capacity, ICU availability, sitter needs, transport delays, and pending discharges.

These tools are not replacements for clinicians. They are operational decision aids. Their usefulness hinges on four factors: accuracy, timeliness, understandability, and a link to someone who can act.

Why ED crowding persists

Emergency department crowding is commonly described in terms of input, throughput, and output.

Input refers to patient arrivals, acuity, seasonal illness, local events, ambulance arrivals, and referral patterns. Throughput refers to triage, clinician assessment, diagnostic testing, medication administration, procedures, consultation, reassessment, and disposition decisions. Output refers to discharge, observation placement, admission, transfer, and movement to an inpatient bed.

AI can assist in each domain, but it cannot fix all three on its own. Arrival forecasting may help staffing, but it will not create inpatient capacity. Admission prediction may alert bed management earlier, but it will not help if no staffed beds exist. Smart bed dashboards can cut down phone calls and enhance visibility. However, they can’t make up for delayed discharges, closed beds, or not enough nursing staff.

This distinction is important for publication accuracy. ED flow technology should be presented as one component of a hospital-wide capacity strategy. It should not be framed as a cure for overcrowding.

AI-supported use cases in the emergency department flow

Demand forecasting is one of the most practical applications. Models may use hour of day, day of week, holidays, seasonality, weather, respiratory virus activity, historical volume, local events, and hospital census to estimate arrival patterns. These forecasts can support staffing, radiology planning, laboratory readiness, and surge preparation. They should be interpreted as probabilistic estimates, not exact predictions.

Admission prediction can help bed management teams prepare earlier. A model may estimate the probability of admission based on presenting complaint, age, vital signs, comorbidities, prior utilization, initial laboratory tests, imaging orders, or clinician documentation. This can be useful if it triggers real action, such as early bed huddles, inpatient discharge prioritization, or escalation when boarding thresholds are exceeded.

AI-assisted triage is more clinically sensitive. Triage tools use both structured data and free-text notes. They help estimate acuity, risk of deterioration, resource use, and the need for quick clinician evaluation. These systems can help find patients who need earlier reassessment. However, they shouldn’t replace proven triage processes or the skills of experienced nurses. The major safety concern is undertriage. Overtriage also has consequences because it can divert attention and resources away from sicker patients.

Waiting-room surveillance is another potential use case. Patients waiting for care are not static. Pain, breathing, bleeding, risk of sepsis, intoxication, mental state, and vital signs can change while the patient waits. Automated reassessment prompts and continuous monitoring tools can catch issues early. But they only work if staff know how to respond clearly.

Smart bed management systems focus on output. They may show bed status and cleaning progress. They also display isolation requirements and unit staffing. They list pending discharges and transport delays. Additionally, they note special bed needs. These systems can reduce fragmented communication and make boarding visible across the hospital. Visibility alone is not enough. The hospital must define who owns the bottleneck and what escalation steps follow.

Evidence summary

The evidence for AI-supported ED flow is encouraging but not definitive. Many predictive models show good results in past tests. However, just because they were accurate in the past doesn’t mean they will improve patient outcomes when used. ED operations are complex. A model may be statistically accurate yet fail clinically if it is poorly integrated into the workflow or if no one can act on its output.

New evidence shows that AI-based triage support can boost triage performance and improve patient flow. However, this should not be taken as a blanket claim that AI reduces ED wait times by 20% to 40% across all settings. Outcomes depend on many factors. These include baseline performance, staffing, and patient mix. EHR integration, hospital occupancy, and inpatient capacity also matter. Behavioral health resources and governance also play a role.

The broader medical machine-learning literature also supports caution. There are still a few randomized trials for machine-learning interventions. This is small compared to the many models that have been published. Reporting quality, reproducibility, inclusivity, and generalizability are recurring concerns. For ED flow applications, stronger evidence should include prospective pragmatic studies, predefined safety outcomes, equity-stratified analyses, and post-deployment monitoring.

Clinical and medication-safety implications

ED crowding is not merely an inconvenience. It can delay analgesia, antibiotics, imaging, anticoagulation decisions, medication reconciliation, procedural sedation, psychiatric evaluation, consultation, and continuation of essential home medications. Patients boarding in the ED may miss scheduled medications. They might also receive duplicate therapies. Incomplete handoffs can occur, leading to confusion. Some patients may remain in areas not meant for long-term inpatient care.

Pharmacists should be included in ED flow redesign when boarding is common. High-yield pharmacist contributions include medication reconciliation for admitted boarders, renal-dose adjustment, antimicrobial review, anticoagulant assessment, high-alert medication oversight, continuation of time-sensitive home medications, and transitions-of-care planning.

AI dashboards should not focus only on beds and timestamps. A good flow system should help identify medication-safety risks that accumulate during long ED stays. For example, an older adult getting admitted might need a review of their medications. This could include anticoagulants, insulin, antiepileptics, Parkinson’s drugs, immunosuppressants, opioids, sedatives, and antimicrobials. Operational efficiency and medication safety should be designed together.

Practical implementation requirements

Implementation should begin with a specific operational problem. “Reduce ED crowding” is too broad. Better targets include reducing left-without-being-seen rates, reducing time from decision to admit to ED departure, improving door-to-provider time, shortening bed assignment delays, reducing behavioral health boarding, improving ICU boarding, or decreasing medication delays for admitted boarders.

The tool should match the problem. Arrival forecasting supports staffing and surge planning. Admission prediction supports bed planning. Smart bed management supports placement and transfer coordination. AI-assisted triage supports risk recognition. Waiting-room surveillance supports reassessment. These tools are related but not interchangeable.

Data quality must be audited before deployment. ED timestamps are often inconsistent. “Decision to admit,” “bed requested,” “bed assigned,” “bed ready,” “patient departed ED,” and “clinician first contact” may be noted in different ways. This can happen across departments and shifts. If these data elements are unreliable, model training and performance reports will also be unreliable.

Workflow integration is equally important. Clinicians need to know what the system recommends. They should understand the expected actions and how urgent they are. Also, they must know how to override these recommendations. A model that sends alerts to a rarely used dashboard is unlikely to improve care. A model that generates frequent low-value alerts may worsen alert fatigue.

Governance, regulation, and accountability

AI tools used in ED operations may fall into different categories. A bed-status dashboard is not the same as a triage recommendation tool. A model that predicts room turnover is not the same as a model that influences clinical prioritization or deterioration detection. Health systems should evaluate each product according to its intended use, FDA status when applicable, EHR integration, cybersecurity posture, audit logging, data-use agreements, and clinical governance requirements.

Clinical decision support tools require special scrutiny when they influence patient care. Institutions should define who is responsible for reviewing, acting on, overriding, and monitoring outputs, and for ensuring clinician override is preserved. The tool should support judgment, not displace it.

Transparency is also essential. Clinical users should have access to plain-language information on the model’s purpose, input variables, training population, known limitations, validation results, update schedule, subgroup performance, and local monitoring plans. A black-box model is especially risky in triage, deterioration detection, and bed-prioritization workflows.

Model drift should be expected. ED populations change over time. So do coding practices and staffing models. Respiratory virus patterns also shift. Hospital capacity changes, too. A model that performed well during development may degrade after implementation. Ongoing monitoring should include calibration, false negatives, false positives, missing-data rates, override rates, downtime, alert burden, subgroup performance, and adverse events.

Equity and bias

AI systems can repeat existing unfairness. This happens if they train on biased data or if their results are used without context. Triage and patient-flow tools should be evaluated across age, sex, race, ethnicity, language, disability, payer category, behavioral health presentation, housing instability, and other locally relevant factors when data are available and appropriate.

Equity review should focus on clinical consequences. A model might seem accurate overall. However, it can perform poorly for certain groups. This includes non-English speakers, older adults, those with psychiatric issues, patients with unusual symptoms, or those with symptoms that aren’t well documented. False reassurance in these groups may worsen existing disparities.

Override patterns should also be reviewed. Frequent overrides could indicate model weakness, a poor workflow fit, or insufficient training. They may also reflect clinicians recognizing risks that structured data misses.

Economic considerations

The business case should be local and conservative. Potential benefits include fewer patients leaving without being seen, improved bed utilization, shorter boarding times, reduced overtime, better throughput, improved ambulance availability, and improved patient experience. These benefits should not be assumed. They rely on baseline performance, implementation quality, staffing, inpatient capacity, and leaders taking action on the system’s recommendations.

Costs include software licensing, integration, cybersecurity review, analytics support, staff training, validation, governance, maintenance, and change management. Institutions should think about hidden costs. These are workflow disruption, alert fatigue, bad bed placement, and the burden of documentation. They should also watch out for overreliance on vendor-reported metrics.

A pilot is usually safer than an immediate enterprise-wide deployment. A good pilot sets clear metrics, finds the target group, and defines go-live criteria. It also sets safety stop rules, names clinical and operational owners, and includes a plan for independent evaluation.

Clinically useful metrics

ED flow tools should be judged by more than dashboard activity. Useful access metrics include door-to-triage time, door-to-provider time, left-without-being-seen rate, and elopement rate. Throughput metrics include ED length of stay for discharged patients, diagnostic turnaround time, consultation delays, and time to disposition decision. 

Boarding metrics cover: 

• Time from decision to admit to ED to departure
• ICU boarding time
• Behavioral health boarding time

Safety metrics should cover:

• Undertriage
• Adverse events in the waiting room
• Medication delays
• Time to analgesia
• Time to antibiotics when needed
• Medication reconciliation for admitted boarders
• Unplanned escalation of care

Model-specific metrics should cover:

• Calibration
• False-negative review
• False-positive burden
• Missing-data rate
• Override rate
• Downtime
• Alert burden
• Model drift
• Subgroup performance

A model that improves throughput while increasing undertriage is not acceptable. Flow must be measured alongside safety.

Table 1. AI-supported ED flow tools

Tool	Best use	Main safeguard
Arrival forecasting	Staffing and surge planning	Use prediction ranges, not exact counts
Admission prediction	Earlier bed planning	Validate locally before operational use
AI-assisted triage	Acuity support	Preserve clinician override
Waiting-room monitoring	Deterioration detection	Define who responds to alerts
Smart bed management	Bed matching and transfer	Audit bed-status accuracy
Command dashboard	Hospital-wide visibility	Assign accountable escalation leaders

Table 2. Metrics before and after implementation

Domain	Measures to track
Access	Door-to-triage, door-to-provider, left without being seen
Throughput	ED length of stay, consult delay, diagnostic turnaround
Boarding	Decision-to-admit to ED departure, ICU boarding, behavioral health boarding
Safety	Undertriage, waiting-room adverse events, unplanned escalation
Medication safety	Medication reconciliation, time-sensitive drug delays
Equity	Performance by language, age, race, ethnicity, disability, behavioral health status
Governance	Drift, missing data, override rate, downtime, alert burden

Table 3. Publication-safe language

Avoid	Use instead
“No-wait ED”	“Low-wait, high-reliability ED”
“AI reduces waits by 20% to 40%”	“Selected implementations have reported improved flow, but results vary”
“AI optimizes triage”	“AI may support triage when locally validated and clinician override is preserved”
“Smart beds solve boarding”	“Smart bed tools improve visibility, but boarding requires hospital-wide action”
“ROI within 18 to 24 months”	“Financial benefit should be evaluated locally”
“Automated decision-making”	“Decision support with clinical accountability”

References

Agency for Healthcare Research and Quality. (2025). AHRQ summit to address emergency department boarding [Technical report]. https://www.ahrq.gov/sites/default/files/wysiwyg/topics/ed-boarding-summit-report.pdf

Ahmadzadeh, B., Patey, C., Norman, P., Farrell, A., Knight, J., Czarnuch, S., & Asghari, S. (2025). Artificial intelligence solutions to improve emergency department wait times: Living systematic review. The Journal of Emergency Medicine, 75, 174–187. https://doi.org/10.1016/j.jemermed.2025.05.031

Boudi, Z., Lauque, D., Alsabri, M., Östlundh, L., Oneyji, C., Khalemsky, A., Lojo Rial, C., Liu, S. W., Camargo, C. A., Jr., Aburawi, E. H., Soucy, N., Singer, A. J., & Bellou, A. (2020). Association between boarding in the emergency department and in-hospital mortality: A systematic review. PLOS ONE, 15(4), e0231253. https://doi.org/10.1371/journal.pone.0231253

Canada’s Drug Agency. (2025). Technologies to address wait times in the emergency department. Canadian Journal of Health Technologies. https://www.ncbi.nlm.nih.gov/books/NBK617247/

Centers for Medicare & Medicaid Services. (2026). Hospital Outpatient Quality Reporting Program. U.S. Centers for Medicare & Medicaid Services. https://www.cms.gov/medicare/quality/initiatives/hospital-quality-initiative/hospital-outpatient-quality-reporting-program

El Arab, R. A., & Al Moosa, O. A. (2025). The role of AI in emergency department triage: An integrative systematic review. Intensive and Critical Care Nursing, 89, 104058. https://doi.org/10.1016/j.iccn.2025.104058

Emergency Nurses Association. (n.d.). Emergency Severity Index handbook (5th ed.). https://media.emscimprovement.center/documents/Emergency_Severity_Index_Handbook.pdf

Food and Drug Administration. (2025). Marketing submission recommendations for a predetermined change control plan for artificial intelligence-enabled device software functions: Guidance for industry and Food and Drug Administration staff. U.S. Food and Drug Administration. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence

Food and Drug Administration. (2026). Clinical decision support software: Guidance for industry and Food and Drug Administration staff. U.S. Food and Drug Administration. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software

Health Level Seven International. (2023). FHIR Release 5: Fast Healthcare Interoperability Resources specification. https://hl7.org/fhir/

Morley, C., Unwin, M., Peterson, G. M., Stankovich, J., & Kinsman, L. (2018). Emergency department crowding: A systematic review of causes, consequences and solutions. PLOS ONE, 13(8), e0203316. https://doi.org/10.1371/journal.pone.0203316

National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework, AI RMF 1.0 (NIST AI 100-1). U.S. Department of Commerce. https://doi.org/10.6028/NIST.AI.100-1

Office of the National Coordinator for Health Information Technology. (2024). Health Data, Technology, and Interoperability: Certification Program Updates, Algorithm Transparency, and Information Sharing Final Rule. U.S. Department of Health and Human Services. https://www.federalregister.gov/documents/2024/01/09/2023-28857/health-data-technology-and-interoperability-certification-program-updates-algorithm-transparency-and

Plana, D., Shung, D. L., Grimshaw, A. A., Saraf, A., Sung, J. J. Y., & Kann, B. H. (2022). Randomized clinical trials of machine learning interventions in health care: A systematic review. JAMA Network Open, 5(9), e2233946. https://doi.org/10.1001/jamanetworkopen.2022.33946

Taylor, R. A., Chmura, C., Hinson, J., Steinhart, B., Sangal, R., Venkatesh, A. K., Xu, H., Cohen, I., Faustino, I. V., & Levin, S. (2025). Impact of artificial intelligence-based triage decision support on emergency department care. NEJM AI, 2(3), AIoa2400296. https://doi.org/10.1056/AIoa2400296