AI triage collects a structured symptom intake from a patient, estimates urgency, and routes them — to self-care guidance, a scheduled visit, or urgent escalation — before any clinician time is spent. The goal is to put the right patient in front of the right level of care quickly, and to keep low-acuity cases from consuming scarce clinical capacity. Mechanically it is a guided intake plus a decision model that maps reported symptoms to a recommended pathway.
Its regulatory posture has to be deliberate. Framed and built as decision support — with transparent logic, clinician oversight, and recommendations a human can review and override — AI triage stays out of FDA medical-device classification. Framed or built as something that diagnoses the patient, it walks into FDA Software as a Medical Device (SaMD) territory, with the validation and clearance obligations that brings. The line is about intended use and how the tool presents itself, so product, clinical, and regulatory decisions here are inseparable.
The design implication is to engineer guardrails around the failure that matters most: under-triage, meaning telling a patient with an emergency that they can wait. The asymmetry is stark — over-triage wastes a clinician's time, but under-triage can cost a life — so red-flag symptoms should trigger conservative, fail-safe escalation rather than relying on the model's confidence. The common mistake is optimizing the system for efficiency and average accuracy while neglecting the rare, catastrophic miss; in clinical triage the tail risk, not the average case, is the one that defines whether the tool is safe.
