Oscar Health

Healthcare administration in the U.S. is among the most documentation-heavy industries in existence, and Oscar Health faced this complexity at scale. Medical records for patients with acute conditions regularly exceed 500 pages of unstructured clinical text — making manual review slow, inconsistent, and expensive. Documenting a single patient-clinician conversation took nurses and care managers more than 20 minutes on average. Meanwhile, claims processing required teams to manually trace millions of contractual variables to resolve escalations, creating bottlenecks and error risk. The combined burden was driving clinician burnout, slowing resolution times, and inflating administrative costs across the organization.

Oscar partnered with OpenAI to deploy large language model-based NLP tools via the OpenAI API across two core workflows. First, they built automated clinical documentation tooling that summarizes care conversations and lab results, reducing the manual transcription burden on nurses and clinicians. Second, they developed a claims assistant that ingests detailed claim trace logs — the full decision history of a given claim — and uses NLP to surface answers and resolve escalations without manual review. To enable this at speed, Oscar became the first health insurer to sign a HIPAA-compliant Business Associate Agreement (BAA) with OpenAI, removing the compliance barrier that typically delays healthcare AI deployments. A centralized internal AI Pod governs cross-team adoption and maintains responsible-use standards.

The claims assistant cut escalation resolution time by 50%, with accuracy matching or exceeding human agents. Oscar projects the system will automate investigation for 4,000+ tickets per month — approximately 48,000 annually. Clinical documentation time dropped by nearly 40%, directly reducing burnout and freeing clinicians for higher-complexity patient work. R&D benchmarks show GPT-4 can deliver up to 90% productivity gains in certain documentation scenarios, pointing to further headroom.

• 50% reduction in claims escalation resolution time
• 40% reduction in clinical documentation time
• 4,000+ tickets automated per month (48,000 annually)
• Accuracy on par or better than human agents for claims triage

• Securing a HIPAA-compliant BAA with your AI vendor before scoping use cases removes the compliance bottleneck that stalls most healthcare AI initiatives.
• Decomposing complex workflows into discrete, well-defined NLP tasks — rather than attempting end-to-end automation — is the primary driver of reliable model performance.
• A centralized AI governance function (Oscar's 'AI Pod') accelerates cross-team adoption while ensuring ethical and regulatory standards are maintained consistently.
• Benchmarking foundation models against proprietary, domain-specific datasets — not general benchmarks — is essential for selecting the right model in regulated industries.
• Addressing health equity gaps (e.g., patients with the longest, most complex records) should be an explicit design criterion, not an afterthought.