OpenAI made computer-use agents available for healthcare on Monday — AI that reads screens, clicks and types inside the software care teams already use — through ChatGPT for Healthcare, ChatGPT for Clinicians and the OpenAI API, with CVS Health and Cigna Healthcare already running them in production. Why it matters: this is computer use moving out of demos and into prior authorizations, referrals and discharge paperwork, some of the most tedious work in medicine.
November 16, 2026
What do the healthcare agents do?
The agents operate existing applications — electronic health records, payer portals and the other systems staff switch between all day — to complete multistep work: authorization requests, referrals, discharge planning and outreach between visits. In OpenAI’s synthetic-patient demonstration, a message about lower back pain prompts the agent to review visit notes, therapy records and prior imaging, draft an MRI order, an encounter note and a patient reply for clinician review, then prepare an insurer-portal authorization request for a specialist to check. OpenAI is working with Abridge, Ambience, Cigna Healthcare and CVS Health, and also added more than 20 healthcare plugins to ChatGPT for Healthcare alongside upgraded clinical search.
What are the safety numbers?
They are OpenAI’s own reported evaluations, and should be read that way. Reviewers rated 99% of 325 reviewed GPT-6 Astra tasks in electronic health record inboxes as clinically safe, and rated the work “very helpful” or “extremely helpful” in 93% of 646 assessments. On HealthAdminBench — an independent benchmark from Stanford researchers covering 135 tasks across prior authorization, medical-equipment orders and appeals — OpenAI reports Astra earned 94.3% of available rubric points and the cheaper GPT-6 Luna 91.8%, with Luna costing 8.5 cents per task, 98% less than Astra.
Who is using it already?
CVS Health is furthest along: OpenAI reports one production agent cleared a months-long document-sending backlog over a weekend, while another handles system steps for tens of thousands of specialist consults each day. CVS has also published its CUA Agent Kit as open source under an Apache-2.0 license. At Cigna, case managers previously navigated more than 10 separate systems to assemble member context; its deployment pulls that information together and went from concept to deployment in under eight weeks. Abridge is developing a cross-system discharge assistant in early environments, and Ambience’s version prepares prior-authorization requests but stops to ask a clinician when the record is ambiguous and waits for approval before submitting.
Access runs through workspace settings: organizations enable computer use in ChatGPT Work and Codex, verified U.S. clinicians can use the free ChatGPT for Clinicians, and developers can build on the API. Role-based access controls, audit logs and approval steps are built in, and HIPAA-covered use requires a Business Associate Agreement and the required security settings; OpenAI says organization data is not used to train its foundation models.
Why it matters
Healthcare paperwork is exactly the kind of high-volume, screen-bound work computer use was promised for — and exactly the setting where a wrong click matters. The honest read: the production deployments so far are administrative, with humans approving the clinical judgment calls, and the headline safety figures come from OpenAI’s own evaluations rather than measured patient outcomes. If the audit-trail model holds up at CVS and Cigna scale, expect every EHR workflow vendor to ship the same thing within a year.
Frequently asked questions
Who can use OpenAI’s healthcare computer-use agents?
Healthcare organizations through ChatGPT for Healthcare, verified U.S. clinicians through the free ChatGPT for Clinicians, and developers through the OpenAI API. Workspace administrators control enablement.
Are the agents making clinical decisions?
In the deployments described, agents handle administrative steps — referrals, authorizations, document work — and draft clinical documents for clinician review. Partners describe approval steps and escalation to humans for judgment calls.
What did the safety evaluation find?
OpenAI reports reviewers rated 99% of 325 reviewed GPT-6 Astra EHR inbox tasks clinically safe. Those are OpenAI’s own evaluation results, not independent patient-outcome studies.
Sources: OpenAI, Unite.AI, Becker’s Hospital Review

