AI News
AI News AgentModel releaseOpenAI4 min read

OpenAI brings AI agents to healthcare teams

OpenAI is introducing AI agents that can navigate medical records, insurer portals and other applications to automate clinical and administrative tasks. Tests report that 99% of clinical tasks were deemed safe, while the more affordable model cost 8.5 cents per task, always with controls and human review.

OpenAI is bringing AI agents that can use computer programs to hospitals, insurers and clinics. These systems can click, type and move between applications to complete tasks that currently force healthcare teams to copy information from one system to another.

The goal is not for AI to make decisions for doctors. It is to handle part of the administrative work and leave people with decisions that require clinical judgment.

From a medical record to an insurer's portal

When a doctor recommends an MRI, for example, the information needed to request authorization is often spread across the medical record, physical therapy reports and previous tests. An employee has to gather that data and enter it again in the insurer's portal.

A computer-use agent can review those records, prepare the MRI order, draft a note for the encounter and suggest a response for the patient. It can then complete the authorization form in the insurer's portal for a specialist to review.

OpenAI demonstrated this process with synthetic data, not real patient information. The doctor and specialist still have the final say.

What results the model achieved

OpenAI evaluated GPT-6 Astra on clinical inbox tasks designed with input from doctors. The tests required understanding a patient's context over time and navigating complex interfaces.

Reviewers analyzed both the model's actions and the notes, orders and messages it generated:

  • 99% of 325 tasks were deemed safe. The evaluation looked for errors, omissions or actions that could harm a patient.
  • Doctors rated the work as “very useful” or “extremely useful” in 93% of 646 ratings.

These figures describe specific tasks evaluated by reviewers. They do not mean the system can work without supervision in every clinical setting.

On administrative tasks such as prior authorizations, appeals and medical equipment requests, GPT-6 Astra scored 94.3% of the points on HealthAdminBench, an independent test developed by Stanford researchers. GPT-6 Luna scored 91.8% and cost 8.5 cents per task, 98% less than Astra. Both models were tested at the highest reasoning level.

How organizations are using it

OpenAI says its partners are already testing or deploying these agents across different workflows:

  • CVS Health uses them to send documents, process referrals and support specialist consultations. According to the company, one agent cleared a document backlog that had been building for months over a single weekend, while another manages system steps for tens of thousands of consultations a day.
  • Abridge is working on agents that help coordinate care plans, such as resolving obstacles to discharging a patient, managing medications or coordinating home care.
  • Cigna Healthcare uses them to gather information from more than 10 clinical and administrative systems. The goal is for case managers to have data on coverage, providers, appointments and previous interactions within seconds.
  • Ambience has built controls that make the agent stop when the record is unclear and ask for a doctor's approval before submitting a request.

In CVS Health's case, the data remains within the company's controlled infrastructure, and the organization says its production data is not used to train foundation models.

The important limit: supervision

Navigating a program does not make AI responsible for a medical decision. In healthcare, every action must be reviewable and stoppable.

The announced tools include role-based permissions, audit logs, screenshots and controls that limit which systems and data an agent can access. Organizations can also require human approval before sending documents, changing information or completing a process.

To work with protected health information under HIPAA rules, organizations must use services covered by the applicable business agreement with OpenAI and enable the required security and retention settings. Availability does not remove those obligations.

Computer-use agents are already available to organizations and developers through ChatGPT for Healthcare, ChatGPT for Clinicians and the OpenAI API, with access conditions that depend on the user type and service.

For you, the most likely change will not be AI handling a consultation on its own, but less time lost to portals, forms and searches across systems. The question to watch is whether that convenience comes with enough controls: in a hospital, a useful agent is not the one that clicks the most, but the one that knows when to stop and ask for review.