OpenAI brings computer-use agents to healthcare paperwork and clinical workflows
The agents can navigate medical records and insurer portals. OpenAI pairs access with company-run safety tests, customer deployment claims and configurable human review.
OpenAI’s Nov. 16 release moves computer-use agents beyond drafting in one interface: they can carry administrative work across EHRs and insurer portals, including systems that lack useful API connections. CVS Health has live workflows and Cigna uses agents to gather member context, while Abridge’s discharge-coordination work is still in development. Agents can prepare actions, but organizations can restrict permissions, require approvals and inspect audit logs; HIPAA-compliant use also depends on eligible services and safeguards in connected systems.
01
CVS Health says one agent cleared a months-long document-sending backlog over a weekend; another handles system steps for tens of thousands of consultations daily, with validation and human oversight.
02
OpenAI rated 99% of 325 GPT-6 Astra EHR inbox tasks safe; physicians rated the work very or extremely helpful in 93% of 646 assessments.
03
At maximum reasoning effort, Astra scored 94.3% and Luna 91.8% of HealthAdminBench rubric points; OpenAI reports Luna costs 8.5 cents per task, 98% less than Astra.
OpenAI made computer-use agents available for healthcare on November 16 through its API, ChatGPT for Healthcare and ChatGPT for Clinicians. The agents click, type and move between applications, letting teams delegate steps in referrals, insurance authorizations and clinical documentation rather than manually transfer information between disconnected systems.
An MRI request illustrates the job. A clinician may recommend a scan, but someone still has to retrieve supporting information from the electronic health record, or EHR, and enter it into an insurer’s separate portal. OpenAI’s agents are designed to carry that work across the interfaces healthcare staff already use.
In a demonstration built with synthetic patient data, an agent reviews visit notes, physical therapy records and prior imaging after a patient reports lower back pain. It drafts an MRI order, an encounter note and a patient reply for clinician review, then prepares an authorization request for an authorization specialist to review.
From demonstrations to daily operations
OpenAI described a mix of production deployments and work still in development. At CVS Health, agents automate document sending, referral processing and steps supporting specialist consultations. The company uses them to navigate applications where software connections known as APIs do not fully support the required work.
According to the announcement, one CVS agent cleared a months-long document-sending backlog over a weekend. Another handles system steps for tens of thousands of consultations daily. CVS uses validation checks and human oversight, with screenshots and action records to verify what happened in the applications.
Cigna Healthcare’s use centers on gathering context for member outreach. Case managers previously navigated more than 10 clinical and administrative systems; Cigna says computer use can bring relevant information together in seconds. Abridge, meanwhile, is developing agents for tasks such as resolving discharge barriers through medication-access and home-health coordination.
Safety ratings and administrative scores measure different things
OpenAI evaluated GPT-6 Astra on multistep EHR inbox tasks shaped by physician input. Reviewers examined actions alongside generated notes, orders and messages. They assessed safety by asking whether inaccuracies, omissions or actions could plausibly harm a patient. Separately, physicians rated the work very or extremely helpful in 93% of 646 assessments.
OpenAI’s reported evaluation results
99%Astra clinical safety rating
OpenAI says 99% of 325 reviewed EHR inbox tasks were rated safe.
94.3%Astra administrative rubric score
OpenAI reports that Astra earned 94.3% of available HealthAdminBench rubric points at maximum reasoning effort.
91.8%Luna administrative rubric score
OpenAI reports that Luna earned 91.8% of available HealthAdminBench rubric points at maximum reasoning effort.
The administrative test covers work such as prior authorizations, appeals and medical equipment orders. HealthAdminBench is an independent benchmark from Stanford researchers, but these model results were reported by OpenAI. Its percentages represent available rubric points, not shares of tasks completed. OpenAI also reports Luna cost 8.5 cents per task, 98% less than Astra.
Who controls the next click
In Ambience’s demonstration, an agent prepares an authorization request from chart evidence. When the record is ambiguous, it stops and asks the clinician, then waits for approval before submitting. Health systems can set permitted actions, pause the agent and inspect a step-by-step audit log.
ChatGPT for Healthcare provides role-based access controls and audit logs. Organizations can add review steps and limit agents to the records, actions and applications each task needs. OpenAI says organization data is not used to train its foundation models. HIPAA-compliant workflows require eligible services covered by a Business Associate Agreement, required security and retention settings, and protections extended to connected systems.
Individual clinicians: Eligible, verified U.S. clinicians can sign up for free ChatGPT for Clinicians accounts and access computer use in Work and Codex.
Developers: Computer-use capability is available through the OpenAI API for building healthcare applications.
Sources
openai.comMore ways to move care forward with ChatGPT for Healthcare
openai.comAdvancing computer-use agents for healthcare teams
Reader comments
Newest comments first. Replies stay oldest first.