Blueprint for Safety: Implementing a Clinically Governed AI Digital Assistant for Patient Guidance

Patients navigating a fragmented health care system may feel increasingly tempted to turn to publicly available large language models for quick answers to clinical questions; however, these tools were not built with patient safety, risk stratification, or escalation pathways in mind. In this case study, the authors describe how Included Health designed, piloted, and clinically governed a risk-stratified artificial intelligence (AI) digital assistant that offered generalized health guidance while reliably routing higher-risk situations to human clinicians. Building on OpenAI's generative pretrained transformer 4 (GPT-4) model, the team created a multitier risk classification engine that separated emergency, high-risk, and standard-risk patient inquiries; developed conservative safety guardrails that blocked AI advice and triggered escalation for concerning symptoms; and ran a continuous human-in-the-loop audit program that reviewed 100% of clinical interactions during the pilot. Using a randomized rollout to half of the patient population, the authors found that the risk-stratified assistant maintained a high level of clinical safety (96% accurate guidance, 0% critical safety events, and no AI-generated diagnoses) while reducing standard-risk queries routed to human support by 65%, shortening average human response times from 9.6 to 3.6 minutes, and improving resolution of health inquiries without additional visits. This blueprint illustrates how health care organizations can pair proactive risk analysis, adversarial testing, and ongoing governance to deploy patient-facing generative AI that is explicitly designed to put safety ahead of convenience and still meet patients' expectations for timely, trustworthy guidance.