An OpenAI-backed study conducted with Kenya-based Penda Health found that deploying an AI “clinical copilot” inside electronic health records significantly improved the accuracy of frontline medical care. The research analysed 39,849 patient visits across 15 Nairobi primary-care clinics between January and April 2025, comparing outcomes for clinicians who used the GPT-4o-powered tool, called AI Consult, with peers who did not. Clinicians assisted by AI Consult recorded a 16% reduction in diagnostic errors and a 13% reduction in treatment mistakes. Errors in history-taking fell by 32%, while investigation-ordering missteps declined 10%. In visits where the system issued high-severity “red” alerts, diagnostic errors dropped 31% and treatment errors 18%. Extrapolated over a full year, the authors estimate the tool could avert about 22,000 diagnostic and 29,000 treatment errors at Penda Health alone. The software runs in the background during consultations, flagging potential lapses but leaving final decisions to clinicians. The deployment received approval from the AMREF Health Africa Ethical and Scientific Review Committee, and no patient harm attributable to the AI’s recommendations was reported. OpenAI and Penda said the findings demonstrate that large-language-model copilots can act as real-time safety nets, though they stressed the importance of tailored implementation and ongoing clinician training.