OpenAI says it is bringing computer-use AI agents to healthcare workflows, where they can navigate existing electronic health record (EHR) systems and insurer portals to help with repetitive, multi-step tasks. In its November 16 announcement, the company described work with CVS Health, Cigna Healthcare, Abridge and Ambience, whose efforts range from production administrative flows to development and demonstrations. The announcement is about assisting care teams across software—not replacing clinicians.

What a computer-use agent does in a healthcare system

Instead of relying only on a software integration, a computer-use agent can read what is on screen, click, type and move between applications. That could matter in healthcare, where staff often have to transfer information between clinical records, referral systems and payer portals.

OpenAI’s MRI example is a research demonstration using synthetic patient data. The agent reviews fictional visit notes, physical-therapy records and prior imaging, then drafts an MRI order, an encounter note and a patient reply for a clinician to review. It also prepares a prior-authorization request for a specialist to review in a separate insurer portal. This illustrates the proposed workflow; it is not evidence that an AI has independently made a real patient’s treatment decision.

Partners are at different stages

OpenAI says CVS Health is using the capabilities in production for repetitive tasks such as sending documents, processing referrals and supporting specialist consults. One workflow cleared a months-long document-sending backlog over a weekend, the companies say; another handles system steps for tens of thousands of consults each day. CVS says its approach uses validation checks and human oversight. Its public CUA Agent Kit describes a process in which an AI agent discovers a workflow, then successful repeatable steps can be converted into inspectable browser or desktop automation code.

OpenAI says Cigna Healthcare is using its models to gather member information from more than 10 clinical and administrative systems for care managers. Abridge is developing an inpatient use case to help coordinate care-plan steps such as medication access and home-health support. Ambience has demonstrated controls that pause when a record is ambiguous and wait for clinician approval before submitting a prior-authorization request. These are not identical deployments: some are in use, while others are being developed or demonstrated.

Availability and patient-data safeguards

OpenAI says healthcare computer-use capabilities are available through the OpenAI API, ChatGPT for Healthcare and ChatGPT for Clinicians. Organizations can limit which applications agents access, configure review and approval steps, and keep audit logs. OpenAI says data from ChatGPT for Healthcare is not used to train its foundation models.

For workflows involving protected health information, OpenAI says organizations must use eligible services covered by its Business Associate Agreement and apply the required security and retention settings. That is a deployment requirement—not a blanket guarantee that any use of an AI agent is automatically HIPAA-compliant. The organization still needs to govern access, human review and the connected systems.

How to read OpenAI’s safety and benchmark figures

In its evaluation of EHR inbox tasks, OpenAI says reviewers rated 99% of 325 GPT-6 Astra tasks as clinically safe. Physicians rated the work “very” or “extremely” helpful in 93% of 646 assessments. OpenAI also reports that GPT-6 Astra earned 94.3% of available rubric points on HealthAdminBench, while GPT-6 Luna earned 91.8%.

Those are evaluation results, not clinical outcomes. The independent HealthAdminBench project describes 135 administrative tasks across prior authorizations, equipment orders, and appeals. Its paper defines full task success as passing every subtask, a stricter measure than earning a share of rubric points. Its public leaderboard also evaluates different model configurations, so the figures should not be treated as a direct comparison or as proof that the same share of real-world cases will be completed correctly. Fierce Healthcare’s independent report says OpenAI’s health AI chief acknowledged that models can still overreach—for example, by taking an action without the proper sign-off.

What the launch could mean for care teams

The clearest near-term opportunity is reducing time spent copying information and navigating disconnected software. CVS’s production examples suggest some administrative workflows have moved beyond demos, while the synthetic MRI scenario shows how a clinician could review drafted work before it advances.

The harder test is dependable performance across real systems, exceptional cases and changing screens. Healthcare organizations will need to define which actions an agent may take, when a person must approve them, how errors are caught, and how audit records and sensitive data are protected. OpenAI’s launch makes computer-use agents available for healthcare settings, but the quality and safety of a deployment will still depend on the models, workflow design and oversight around them.

Frequently asked questions

What are computer-use agents in healthcare? They are AI systems that interact with software screens—reading, clicking and typing—to help complete multi-step tasks across applications such as EHRs and payer portals.

Can they make medical decisions on their own? OpenAI’s featured example uses synthetic data and has the agent prepare drafts for clinicians and authorization specialists to review. Organizations configure access and approval requirements for their own workflows.

Are the agents available now? OpenAI says they are available through its API and healthcare ChatGPT products. Healthcare organizations using protected health information need eligible services, a Business Associate Agreement and required security and retention settings.

Sources: OpenAI’s computer-use announcement; ChatGPT for Healthcare update; Becker’s Hospital Review; CVS Health’s Agent Kit.