Executive Summary
A composite health system deployed an AI-driven patient triage tool, reducing demographic disparities in high-acuity wait times from an 8% variance to under 1% within six months. This was achieved not by finding a "less biased" model, but by enforcing a cost-per-triage cap of $0.15 and building a real-time bias dashboard before full rollout. The key was refusing to scale the system until the human-in-the-loop audit path and its unit economics were defined, preventing a trade-off between equity and throughput.
The Challenge
The Chief Nursing Officer (CNO) of a multi-hospital network sponsored an initiative to use a large language model (LLM) to accelerate patient triage in overburdened Emergency Departments (EDs). The goal was to reduce patient wait times and improve resource allocation. However, the project immediately faced competing constraints from three key stakeholders.
The CNO’s primary concern was patient safety and equity. Any automated system could not introduce or amplify existing demographic biases in triage acuity, which carried significant clinical and legal risk. The Chief Financial Officer (CFO) saw the potential for a runaway operational expense, as unmanaged LLM inference costs can quickly erase any efficiency gains. The Director of Clinical Informatics, tasked with implementation, was caught between the need for real-time performance (sub-second responses integrated into the Electronic Health Record) and the rigorous demands of model validation, HIPAA compliance, and clinician trust.
The core challenge was not just building an accurate model, but designing an auditable, cost-controlled system that clinicians would actually use. The stakes were clear: a biased model could lead to delayed care for vulnerable populations, while a costly or slow system would be rejected by both finance and frontline staff, leaving the ED with the same throughput problems it started with.
