An AI's refill decisions agreed with doctors 97 to 98 percent of the time in a Utah pilot
The privacy-built service splits a visit into steps, sends each to the best model, and uses guidelines rewritten for AI.
Lightspeed Venture Partners · How AI Seeks to Change Every Step of the Patient Journey | LightworkIn a Utah pilot reported by the company, an artificial intelligence system's prescription refill decisions were judged to agree with what a human doctor would have done in 97 to 98 percent of cases. That agreement means outside review found the same choice, not that the AI acted alone, and the company says the technology is built with guard rails and human oversight.
The system is Doctronic, an online service for getting seen for symptoms and moving from intake through triage toward treatment. A Doctronic executive said it was built from day one to comply with HIPAA, the federal law that protects health information. The executive said most early AI systems were not, and that privacy protection matters for consumer trust.
Instead of using one model for everything, the visit is split into steps, such as taking the history and drafting a refill decision, and software sends each step to whichever AI model handles it best. It also uses company-written clinical guidelines rewritten to be easily read by AI and to follow a condition over time. The company says that match rate keeps improving compared with general-purpose base models.
Questions people ask
Does 97 to 98 percent agreement mean the AI prescribes on its own?
Concordance means reviewers compared the AI's refill choice with the choice a physician would have made and found agreement. The company presents the Utah result as evidence of quality, and says the system uses guard rails and human oversight because health decisions carry high stakes.
Why does HIPAA compliance matter for a health chatbot?
HIPAA is the main U.S. law for privacy of personal health information, with strict rules on storage and sharing. Many consumer AI tools were not designed to meet those controls. Meeting them from the start helps earn patient trust and work with clinics, health systems, and regulators.
Why use several AI models instead of one?
Different models can be better at different jobs, such as gathering history, checking guidelines, or drafting language. Routing each task to the strongest model can lift accuracy above what a single general-purpose model delivers on its own.