AI in the Emergency Department: What Actually Works

AI in the emergency department is already in production at scale in four places: imaging triage (large-vessel-occlusion and intracranial-hemorrhage detection that pages the neuro team before a radiologist opens the study), ambient documentation (scribes that draft the note from the room audio), flow and forecasting (arrival volume, boarding, and admission-likelihood models feeding bed management), and demand-side intake (AI phone and chat handling the calls, follow-ups, and scheduling that surround the visit). The FDA has now authorized more than 1,000 AI-enabled medical devices, roughly three-quarters of them in radiology. What has *not* worked reliably: sepsis prediction, acuity scoring, and anything sold as clinical decision-making rather than clinical prioritization. Below is what each category actually returns, what it costs, and the specific cases where buying it is a mistake.

The Four Categories, With Real Numbers

Imaging triage is the most defensible ROI in the building. Stroke-detection AI (Viz.ai's LVO tool cleared via De Novo in 2018, Aidoc's ICH triage, RapidAI) doesn't diagnose — it reorders the worklist and fires a push notification. Published door-in-door-out reductions for transfer patients in the 20–40% range are typical, and CMS granted Viz LVO a New Technology Add-on Payment of up to $1,040 per case before it expired in FY2023. The mechanism is boring and that's why it works: it moves a phone call up by 15–30 minutes.

Ambient documentation is the fastest-adopted and the least contested. A 2024 *NEJM Catalyst* report on Kaiser Permanente's ambient-scribe deployment covered more than 300,000 encounters in ten weeks across thousands of physicians. Reported effects cluster around a 20% drop in time spent in notes and 40–60 minutes per day returned to clinicians who use it heavily. In an ED specifically, the gain is smaller than in clinic — interrupted, multi-patient workflows break the single-encounter recording model — but the burnout math still holds. Emergency medicine has repeatedly topped Medscape's burnout rankings at roughly 63–65%, and replacing an EM physician costs a hospital somewhere between $250,000 and $1,000,000 in recruitment, ramp, and locum coverage.

Flow models are real but bounded. Arrival forecasting is a solved statistical problem — you can hit ±5–8% on daily volume with two years of history and a calendar. The constraint is that knowing you'll have 214 arrivals on Monday does nothing if you have no inpatient beds to move admits into. In ACEP's 2022 survey, 97% of emergency physicians reported boarding admitted patients in the ED. AI does not create beds.

Sepsis prediction is the cautionary tale, and every vendor demo skips it. The external validation of the Epic Sepsis Model published in *JAMA Internal Medicine* in 2021 found an AUC of 0.63, sensitivity of 33%, and a positive predictive value of 12% — while alerting on 18% of all hospitalized patients. It missed two-thirds of sepsis cases and generated an alert on nearly one in five patients. If a vendor won't show you PPV and alert burden alongside AUC, you're being sold on the wrong number.

The Part Nobody Benchmarks: Demand Before the Door

US EDs handle roughly 155 million visits a year. Clinical AI touches the patient after arrival. Almost nothing in the standard ED AI stack touches the 48 hours *before* arrival — and for freestanding EDs, hospital-affiliated urgent cares, and multi-site emergency groups, that window is where the volume and the payer mix are actually decided.

Left-without-being-seen rates tell you how much of this is a service problem rather than a medical one. National LWBS sat near 1% before 2020 and has run above 2% since, with individual systems reporting 4–10% during surge months. A 60,000-visit ED at 3% LWBS walks out 1,800 patients a year. At a conservative $1,200 in net contribution per treated visit, that's about $2.2 million leaving through the front door, and those patients tell people.

The pre-arrival layer that moves this:

The ED is the only service line in American healthcare where more demand can make you poorer. Every other business we work with wants the phone to ring more. An ED that markets indiscriminately can add boarding hours, LWBS, and uncompensated care simultaneously.

That's the sentence most agencies won't put in writing.

What It Costs

Integration is the line item people underestimate. Budget 3–6 months and real IT hours for HL7/FHIR work, SSO, and security review. A $60,000 tool with a nine-month integration queue has an effective first-year cost far above its license fee.

When This Is a Bad Investment — Read This Section Twice

Do not buy clinical AI if your ED is under 20,000 annual visits. Per-facility licensing doesn't scale down. At 15,000 visits, an imaging triage contract costs $4–$8 per visit against a stroke-alert volume that may be under 50 cases a year. Buy a better transfer agreement instead.

Do not buy demand-side marketing if you're already boarding. If your median admit boards 6+ hours and your inpatient occupancy runs above 90%, adding ED volume converts directly into LWBS, ambulance diversion, and press-ganey damage. Fix throughput first. We have told prospective clients this and lost the engagement; it was still the right call.

Do not buy it if your payer mix is the actual problem. Marketing raises total volume; it does not meaningfully re-sort acuity or coverage. A campaign that grows visits 12% while your self-pay share is 22% grows your bad debt 12% too. Model contribution margin by payer *before* the campaign, not after.

Do not buy AI chat that triages. Any pre-arrival tool that tells a patient whether to come in is practicing medicine, and EMTALA plus state law leaves you very little room. Pre-arrival AI should answer logistics — hours, location, wait time, insurance, parking — and route everything clinical to a human or to 911 language. Vendors who blur this line are handing you liability, not leverage.

Do not buy anything you can't audit. Ask every vendor for PPV, alert volume per 1,000 patients, and subgroup performance by race, age, and sex. Pulse oximetry bias is the standing example of how a device validated on one population underperforms on another; AI models trained on that same data inherit the flaw.

The Failure Modes That Actually Show Up

Alert fatigue is the first. A model with 12% PPV firing on 18% of patients trains staff to dismiss it within about six weeks, and dismissal generalizes to your *good* alerts. Second is drift — a model tuned on 2022 case mix degrades quietly as your catchment, EHR templates, or documentation habits change, and almost nobody re-validates on a schedule. Third is the shadow workflow, where nurses build a parallel process around the tool and you're paying for software that produces spreadsheets. Fourth, on the marketing side: attribution collapse. If your call tracking, EHR arrival data, and ad platforms never join, you'll spend a year unable to prove whether any of it worked.

How to Pilot Without Betting the Department

Run a 90-day pilot with a pre-registered primary metric and a kill threshold you write down before you start. One metric — door-to-needle, notes-per-shift, LWBS rate, or cost per insured arrival. Shadow-mode the clinical models for the first 30 days: the model runs, nobody acts on it, and you compare its output against actual outcomes. Demand a contract with a 90-day out. On the marketing side, start with one site and one payer-favorable service line rather than a system-wide launch.

For health systems and multi-site groups, the demand layer is where we work — intake, follow-up, local search, and attribution, built to survive a compliance review. Our industries page covers how that stack differs by service line, and case studies shows the actual before/after numbers. If your bottleneck is inside the department, the honest answer is that a marketing agency can't fix it, and you should spend that budget on beds and staffing.