OpenAI is targeting administrative processes such as insurance authorizations, referrals, discharge coordination, and patient outreach. These activities often require healthcare workers to retrieve information from medical records and manually enter it into separate systems before care can proceed.
In one demonstration involving a fictional patient experiencing lower back pain, an OpenAI agent examined medical records, previous imaging, and physical therapy documentation. It used that information to prepare an MRI order, clinical notes, and a proposed response to the patient. The agent also completed a draft insurance authorization request using the insurer's online portal. The clinical documents and authorization materials were left for human review rather than being submitted independently.
OpenAI evaluated GPT-6 Astra on similar workflows involving messages received through electronic health record inboxes. Reviewers assessed both the model's actions inside the software and the clinical materials it generated, looking for errors or omissions that could potentially harm patients.
Of the 325 tasks examined for safety, 99% received a clinically safe rating. In a separate assessment involving 646 physician evaluations, 93% rated the model's work as either very helpful or extremely helpful.
The company also tested its models on HealthAdminBench, a benchmark developed by Stanford researchers to measure AI performance on healthcare administrative work. The evaluation covers 135 tasks comprising 1,698 individual steps, including insurance appeals, prior authorization requests, and medical equipment orders.
GPT-6 Astra achieved 94.3% of the available scoring points, while the smaller GPT-6 Luna model reached 91.8%. Both were tested using their highest reasoning settings. According to OpenAI, Luna completed tasks at an average cost of approximately 8.5 cents each, representing a 98% reduction compared with Astra.
Those results measure performance against individual scoring criteria rather than the percentage of entire workflows completed successfully. The original HealthAdminBench research reported a maximum end-to-end completion rate of 36.3% across seven configurations and used simplified conditions for certain authentication and session-management challenges. OpenAI did not provide a comparable full-task success rate for its new models.
Beyond benchmark testing, OpenAI highlighted several organizations already applying computer-use technology to healthcare operations.
CVS Health is using agents to automate document delivery, referrals, and specialist consultation processes. In one production deployment, an agent eliminated a document-processing backlog accumulated over several months during a single weekend. Another system performs software-related steps associated with tens of thousands of specialist consultations daily.
CVS Health has also released an open-source Agent Kit that uses OpenAI's technology to identify and automate workflows in existing applications. The system can translate successful agent interactions into executable scripts, reducing the need to repeatedly invoke an AI model for the same procedure. CVS Health says its deployments include validation checks, human supervision, and records of agent activity.
At Cigna Healthcare, the technology is being used to consolidate information that case managers previously gathered from more than 10 different clinical and administrative systems. The agents retrieve details about insurance coverage, healthcare providers, appointments, and previous interactions, helping staff prepare for conversations with members.
Cigna says the deployment progressed from its initial concept to operational use in less than eight weeks as part of its broader collaboration with OpenAI.
"Moving from concept to deployment in under eight weeks shows what’s possible when we bring together AI, our expertise in care management, and the right guardrails for responsible use," said Madhu Nutakki, Cigna Healthcare's chief information officer.
Other partners are developing applications that coordinate activities across multiple systems while retaining clinician oversight.
Abridge is building agents that can assist with hospital discharge planning, including medication access and home healthcare arrangements. Its approach allows multiple agents to work on related tasks simultaneously, bringing clinicians into the process when medical judgment or decisions are required.
Ambience Healthcare is focusing on insurance authorization workflows. Its system extracts supporting information from patient records, prepares requests through insurer websites, and pauses when clinical documentation requires clarification. Healthcare organizations can define permitted actions, interrupt agent activity, and review detailed logs of each step.
OpenAI also identified Assort, Bunkerhill, Forus, Gravity, and Tennr among healthcare startups exploring computer-use applications. Assort, for example, has developed an agent that completes referral documentation in an electronic health record when the necessary action is not supported by its existing API connection.
The expansion introduces additional considerations around patient privacy, software permissions, and clinical accountability. OpenAI says ChatGPT for Healthcare includes access restrictions based on user roles and logging capabilities that allow organizations to monitor agent activity. Healthcare providers can also configure approval requirements and limit which applications, records, and actions an agent can access.
Organizations handling protected health information must use eligible OpenAI services covered by a business associate agreement and apply the required security and data-retention settings. OpenAI says customer organization data is not used to train its foundation models.
The computer-use capabilities are available to healthcare organizations, individual clinicians, and developers through OpenAI's existing products. Verified U.S. clinicians can access ChatGPT for Clinicians at no cost, while organizations can discuss deployments with OpenAI and developers can build applications using its API.
The release gives healthcare teams a way to automate work across existing software without replacing those systems or building a dedicated integration for every task. OpenAI's early deployments demonstrate applications in document processing, care coordination, and insurance administration, while its safety evaluations and configurable approval controls reflect the need for human oversight when AI agents interact with clinical information and healthcare workflows.
This analysis is based on reporting from OpenAI.
Image courtesy of OpenAI.
This article was generated with AI assistance and reviewed for accuracy and quality.