A real-world trial of an AI-powered clinical support tool in Kenya has shown promising results in reducing medical errors, marking a breakthrough in the application of artificial intelligence in frontline healthcare settings.
In a joint study by OpenAI and Penda Health—a network of 16 primary care clinics in Nairobi—researchers found that an AI tool known as AI Consult reduced diagnostic errors by 16% and treatment errors by 13% among clinicians using it. The study, involving data from approximately 20,000 patient visits, is one of the first large-scale, prospective evaluations of AI used alongside health professionals in live clinical settings.
Unlike previous AI studies based on simulations or retrospective data analysis, this project embedded the tool directly into day-to-day operations at Penda Health, where clinicians often act as primary care providers across a wide range of medical conditions. These clinicians, similar to physician assistants in the U.S., operate under national and in-house clinical guidelines but face constant challenges due to the broad scope of knowledge required.
“This tool acts as a safety net,” said Dr. Robert Korom, Chief Medical Officer at Penda. “It’s not dictating care, but guiding clinicians and providing corrective nudges when necessary—like a senior physician supervising a junior colleague.”
Initially, AI Consult offered optional text-based guidance during consultations, but clinicians used it inconsistently. The tool was then adapted to run silently in the background, prompting clinicians only when their clinical choices deviated from standard practice—such as overprescribing antibiotics. The result was greater use and improved outcomes.
Beyond improving clinical decisions, AI Consult has become a learning resource. “It was a surprise,” said Karan Singhal, OpenAI’s Health AI lead. “Clinicians viewed it as an educational partner, helping them build knowledge and confidence.”
Dr. Sarah Kiptinness, Head of Medical Services at Penda, said the tool also provided immediate performance feedback through a color-coded system—green, yellow, or red—for clinical actions. This has helped the company tailor future training. “We plan to create personalized learning modules,” she said. “For example, if someone is strong in obstetrics but weaker in pediatrics, we can target those gaps.”
The success of AI Consult has prompted calls for broader real-world testing. Dr. Isaac Kohane of Harvard Medical School described the study as a long-awaited shift toward practical, evidence-based AI use in healthcare. Dr. Ethan Goh of Stanford added that such findings could pave the way for even more proactive AI systems that prevent, rather than simply correct, medical mistakes.
As small clinics and underserved regions continue to face provider shortages, AI tools like this may play a key role in expanding access and improving the standard of care. “We probably wouldn’t want our clinicians to be completely without this,” Korom said.
