Oscar Health cut claims resolution time by 50% using GPT-4

Health insurance is a paperwork nightmare masquerading as a business. Oscar Health has spent years trying to digitize that mess, and for its CTO Mario Schlosser, large language models finally cracked the code. He put it bluntly: language models were the first tech that could translate the chaos of real-world healthcare into a structured plan. That realization pushed Oscar to go all-in on OpenAI, becoming the first insurer to sign a Business Associate Agreement with the company—a move that locked in HIPAA compliance while competitors were still wringing their hands about regulatory risk.

Nikhita Luthra, who leads AI R&D at Oscar, says the company constantly benchmarks every major model against proprietary healthcare datasets. OpenAI’s models win every time. That consistent performance is now showing up in hard numbers. Clinical documentation that once ate up 20-plus minutes per patient interaction has been slashed by nearly 40% across the board, and R&D tests suggest GPT-4 can push those gains to 90% in certain scenarios. The less obvious win? Nurse burnout, which drops when clinicians stop drowning in administrative busywork.

Claims processing is where things get genuinely transformative. Understanding a claim’s full lifecycle means untangling millions of contractual variables—work that requires human agents to dig through dense, sprawling logs. Oscar built an assistant with OpenAI’s API that navigates those claim traces automatically. The result: resolution time for escalations dropped by half, with accuracy matching or beating human agents. The company expects to automate investigation for 4,000 tickets per month, scaling to 48,000 by year-end.

Then there’s the medical records problem. For the sickest patients—those with records stretching to 500 pages—finding relevant information has always been a needle-in-a-haystack exercise. Oscar is now using AI to summarize those records before provider visits, extract signals for care decisions, and flag diabetic patients who might benefit from continuous glucose monitoring. Luthra frames this as an equity play: the patients with the most complex cases have historically been the ones most poorly served by unwieldy documentation. The six-person AI pod driving all this—five of them women, all in their twenties and thirties—operates on the principle that curiosity about real-world problem-solving matters more than publication records. Oscar’s not dabbling in AI. They’re restructuring around it.