AI agents are being tested as digital assistants for healthcare paperwork, moving information between clinic records and insurance portals through the same screens staff use. OpenAI says its GPT-6 Astra and GPT-6 Luna models can prepare clinical drafts and handle administrative steps, with clinicians and staff reviewing work before release.
Agents work through existing software
Many healthcare tasks involve moving details between electronic health records and external websites. OpenAI’s computer-use agents are designed to click and type in those existing interfaces, which can help when systems lack APIs for a needed step.
For example, when a patient reports back pain, OpenAI says an agent can search relevant prior findings, draft an MRI order and a reply to the patient, and prepare a prior authorization request on the insurer’s portal. Those drafts are reviewed by clinicians and responsible administrative staff before they are sent or finalized.
The features are available through the API and in ChatGPT for Healthcare and ChatGPT for Clinicians. The intended role is preparatory: the agent handles software steps and assembles drafts, while people review the output.
Evaluation results and costs
In an OpenAI evaluation of work from electronic health record inboxes, GPT-6 Astra prepared tasks that included an ibuprofen dosage for a child with a broken arm. Physician reviewers rated 99 percent of the 325 tasks evaluated as clinically safe. In 93 percent of 646 reviews, they said the preparation was very or extremely helpful.
For administrative work such as prior authorizations, appeals and medical equipment orders, the company points to HealthAdminBench, an independent benchmark from Stanford University researchers. GPT-6 Astra scored 94.3 percent of possible points. GPT-6 Luna scored 91.8 percent; OpenAI says it cost about 8.5 cents per task, 98 percent less than Astra. Both models ran at maximum reasoning effort on the benchmark.
These figures describe specific evaluations and workflows. They indicate how the models performed in those settings, but the source does not establish that every clinic task will have the same results.
Healthcare organizations are trying the tools
At CVS Health, agents already process referrals and specialist consultations across separate systems. According to OpenAI’s announcement, one agent cleared a months-long backlog of outgoing documents in one weekend. Another handles system steps for tens of thousands of consultations daily. CVS Health published its software architecture as Agent Kit on GitHub.
Cigna Healthcare uses the models to merge patient information from more than 10 separate systems. Abridge is developing agents for planned use in clinical discharge planning. Ambience has shown a safety mechanism in which an agent stops when a patient’s record is unclear and asks a physician for input. Assort uses the feature in Codex to complete referrals in practice management software where APIs do not support the final click.
Privacy and oversight shape deployment
Because agents interact with electronic health records, access to sensitive information is a central concern. OpenAI says it does not use organizational data to train its base models. According to the company, ChatGPT for Healthcare offers role-based access controls and audit logs, and organizations can configure systems so agents access only the records and actions needed for a task.
For HIPAA-compliant workflows involving protected health information, organizations must sign a Business Associate Agreement with OpenAI, use covered services and apply required security and retention settings. Healthcare facilities remain responsible for proper use, while OpenAI has obligations as a contracting party. Verified US medical professionals using the free individual ChatGPT for Clinicians plan can also sign an agreement, which applies only to covered features.
In the CVS deployment, OpenAI says data stays within infrastructure controlled by CVS, and CVS reviews agent steps through screenshots and action logs. The announcement does not say where those screenshots are processed or how long they are retained. It also does not address the European General Data Protection Regulation (GDPR); the offering currently targets the US market.
Whether these tools gain wider use will depend on how reliably they work with existing systems and whether organizations can monitor their actions in practice. The source also leaves open how human decision-making will be required for steps with clinical or legal consequences. For now, the described workflow keeps people responsible for reviewing drafts before they are released.