Over 14 million patients across more than 1,000 clinics are now served by the healthcare automation platform Anima, which processes more than 800,000 cases a month as medical providers attempt to contain administrative overhead. When backend operations and patient intake channels become saturated, critical information sits in queues, and clinical workflows break down under the weight of manual documentation. Operational margins in primary care depend directly on how fast administrative friction can be removed without expanding headcount.
Clinics and care teams face mounting pressure to process patient requests, clinical letters, and consultation notes before backlogs compromise patient safety. The core economic lever for medical providers deploying artificial intelligence lies in reducing administrative labour hours while maintaining strict deterministic control over patient intake.
Infrastructure Benchmarking and Unit Economics
To power its operations, Anima relies on specialized models selected after rigorous evaluation of available systems. The company uses GPT-6 models in its document workflows and GPT-Live for calls, selecting OpenAI after benchmarking all closed-source and open-source models, according to Shun Pang, Anima’s CEO and co-founder.
“We chose OpenAI for Anima’s mission-critical workloads because of superior intelligence, intelligence per dollar, and steerability with in-context learning.”
This economic calculus determines whether AI deployment remains a costly experiment or becomes a permanent margin-accretive infrastructure layer. On September 22, 2026, OpenAI launched an improved prompt caching system with the GPT-6 family to deliver higher cache hit rates by default and reduce computing overhead.
Voice Automation and Clinical Throughput
Administrative bottlenecks concentrate heavily around telephone intake and document processing. Anima’s AI receptionist, Andi, is built on GPT-Live to handle patient symptom descriptions over the phone, while a proprietary policy engine guides the conversation through deterministic, clinically approved intake workflows.
GPT-5.4 mini then selects the appropriate questionnaire, and the completed medical history appears as a structured case in the clinic dashboard. How do these automated voice pipelines translate into clinic economics?
Clinics using Andi have reduced phone waits from 40 to 50 minutes to 10 minutes, a reduction of at least 75%, according to company reports. In Anima’s patient surveys, 86% said they would use Andi again.
Document processing yields similar operational shifts. Practices that used to process about 15 documents per hour typically process 30 to 45 with Anima's Documents product, while top-performing clinics process 45 to 60 using the system powered by GPT-6 models.
Scalability and Clinical Accuracy
Maintaining reliability across 800,000 monthly cases requires strict adherence to clinical coding standards without requiring custom fine-tuning for every medical practice. Clinical coding proved to be one of the hardest technical problems to solve, as different practices make different decisions about which codes to use among hundreds of thousands of possibilities.
“There are around 800,000 possible SNOMED codes. Only a fraction are used regularly, but different practices still make different decisions about which codes to use.”
To solve this without custom engineering overhead, Anima uses in-context learning to adapt model decisions to each practice. To capture consultation transcripts and draft structured notes, the platform uses GPT 5.6 Luna in its Scribe product, reducing documentation time from several minutes to seconds.
Specialized closed-source models deliver deterministic control and high instruction-following reliability at scale, proving that voice and document automation can fundamentally restructure clinic unit economics without sacrificing clinical accuracy.
Why it matters
Handling over 800,000 monthly cases without fine-tuning separate weights for each provider required pairing OpenAI GPT-Live with strict deterministic policy engines to enforce clinical coding rules. Tripling document processing throughput directly protects operating margins in primary care without adding administrative headcount. The unit economics work solely because deterministic clinical guardrails keep general-purpose model outputs strictly inside protocol bounds.
Our stories: OpenAI GPT-Live Audio Engine Replaces Request-Response Loops (August 3).