Operationalizing Clinical Knowledge for Triaging Assistance with GenAI

As more people turn to generative AI for health questions, the advice they act on must be both specific to their circumstances and calibrated on the axis that matters for an uninsured patient: conservative enough to catch an emergency without routing a routine complaint to a four-figure bill. Existing triage work assigns acuity to patients already inside an emergency department; the uninsured face a prior question: which door to enter, when the wrong one is the difference between a \$40 retail-clinic visit and an ER bill. We present CareFinder, a working system that operationalizes a medical model's clinical knowledge for this decision: MedGemma identifies the required tests and interventions; a general orchestrator (Gemini Flash) then makes a Google Places tool call to find nearby facilities that offer them and prices the answer against live data, with no stored database. Building it uncovered a new, generalizable failure mode: asked to pick a setting directly, MedGemma was risk-averse in risky cases and over-conservative in routine ones. It identified the correct tests, then escalated the venue. We report the data and evaluation, and (because the obvious alternative is a frontier chatbot) an audit of one: it is emergency-safe and stable, yet fabricates real-looking facilities and over-triages the poor, failure modes an open, verifiable pipeline is built to catch.