OpenAI announced healthcare computer-use agents on November 16, 2026, alongside improved clinical search and more than 20 healthcare plugins. The agents can navigate existing software by reading screens, clicking and typing, helping teams move documents, process referrals and prepare prior authorization requests.
Computer-use capabilities are available through ChatGPT for Healthcare, ChatGPT for Clinicians and the OpenAI API. The clinical-search improvements cover both healthcare ChatGPT products, while the new plugins are available in ChatGPT for Healthcare. Access depends on workspace settings, eligible services and provider requirements. (Computer-use announcement, product update)
The practical change is that AI can carry information into another application and prepare the next step. That also makes permissions, approval checkpoints and audit records part of the workflow itself.
What Healthcare Agents Can Do Today
Computer-use agents work through software interfaces, including applications where APIs do not support every required step. An agent can retrieve information from an electronic health record, navigate to another application and enter that information into a form. OpenAI describes configurable application access and human-review steps, alongside audit logging.
That approach addresses a specific integration problem: a healthcare organization may have the necessary records and software, but still rely on staff to carry information between them. Fierce Healthcare reports that OpenAI sees computer use as a way to fill gaps while continuing to invest in APIs, plugins and other integrations. (Fierce Healthcare)
The announced workflows have different levels of maturity.
MRI Authorization: A Demonstration With Two Reviewers
OpenAI’s MRI example uses synthetic patient data. After a patient messages about lower back pain, the agent reviews visit notes, physical therapy records and previous imaging. It drafts an MRI order, an encounter note and a patient reply for clinician review.
The agent then uses the records to prepare an authorization request in an insurer’s portal for an authorization specialist to review. That demonstrates a workflow spanning clinical and administrative systems, not an autonomous authorization decision or evidence of improved patient outcomes. (OpenAI’s demonstration)
The division of responsibility matters. A clinician reviews the proposed clinical work; an authorization specialist reviews the payer submission. My view is that these separate checkpoints are more useful than a generic instruction to keep a human involved: they identify who should review which output before it proceeds.
Document Movement and Referrals: Reported Production Work
CVS Health is using OpenAI’s computer-use capabilities for document sending, referral processing and specialist-consult support, according to OpenAI’s announcement.
In one reported production workflow, an agent cleared a months-long document-sending backlog over a weekend. Another handles system steps for tens of thousands of consults daily. CVS uses validation checks and human oversight, including screenshots and action records to verify actions and application state. Those scale claims come from the company account published by OpenAI, not an independent performance audit. (OpenAI’s CVS Health account)
Abridge’s example is earlier-stage. It is developing agents for care coordination, including a planned inpatient workflow addressing discharge barriers such as medication access and home-health arrangements. That distinction keeps a deployed document workflow separate from a planned clinical coordination feature.
Clinical Search and Plugins Handle Different Parts of the Job
OpenAI says its updated clinical search improves how ChatGPT finds, evaluates and cites medical evidence. The feature is available in ChatGPT for Healthcare and ChatGPT for Clinicians, with citations clinicians can inspect.
In a company-reported blind evaluation involving more than 1,300 physician reviews, the updated experience received higher ratings across all eight evaluated dimensions, including factual correctness, safety, actionability and source authoritativeness. That supports OpenAI’s claim of improvement within its evaluation, but the headline results provide no single accuracy rate or size of improvement. (Clinical-search announcement)
The plugin expansion addresses access to operational tools and records. OpenAI lists 21 new healthcare plugins, including Amion, Availity Extend, Luma Health, Metriport, Redox, Stedi, Turquoise and UpToDate Expert AI. They are available in ChatGPT for Healthcare, including through ChartGPT, which brings ChatGPT into an EHR workflow.
OpenAI illustrates combinations such as:
- Visit preparation: Find appointments needing confirmation through Luma Health, then check insurance eligibility through TriZetto.
- Post-discharge follow-up: Retrieve an outside discharge summary through Metriport, identify medication changes and consult UpToDate guidance with links to the sources.
- Inpatient coordination: Retrieve a hospital’s insulin-pump policy through Aiva and find the covering endocrinologist through Amion.
- MRI cost comparison: Identify a procedure code through IMO Health, check benefits through Stedi and retrieve negotiated rates through Turquoise Health.
These are published workflow examples, not evidence that every organization already has the necessary accounts, permissions and data connections. (Plugin list and examples)
The MRI example makes a useful distinction: a negotiated rate is not automatically the patient’s out-of-pocket cost. OpenAI’s example explicitly asks the system to identify additional information needed to estimate what the patient would owe. I’d preserve that requirement in any deployed workflow, because a precise-looking price can still be incomplete.
What the Safety Results Actually Measure
OpenAI evaluated GPT-6 Astra on EHR-inbox tasks shaped by physician input. Reviewers examined the agent’s actions and its resulting notes, orders and patient messages.
The company reports:
- 99% of 325 reviewed tasks were rated clinically safe.
- 93% of 646 physician assessments rated the work very or extremely helpful.
OpenAI defined clinical safety by whether inaccuracies, omissions or actions could plausibly lead to patient harm. These are vendor-run evaluations of selected tasks, with different denominators for safety and helpfulness. They are not independent clinical-trial results or a measured real-world patient-harm rate. (OpenAI’s evaluation description)
The distinction matters because a task-level rating combines several concerns: whether the agent understood the chart, navigated correctly, produced appropriate content and respected workflow boundaries. A favorable aggregate score cannot tell a hospital how reliably its particular approval process will work.
Fierce Healthcare asked OpenAI health AI leader Karan Singhal about the roughly 1% of tasks rated unsafe. He said researchers had not identified cases creating a specific, plausible risk of patient harm, while acknowledging authorization problems:
“In some cases, the model might overreach and might take an action without getting the proper sign-off or authorization”
That account leaves an important boundary around the headline percentage. The safety rating and the executive’s description are not interchangeable measures, and neither supplies an observed injury rate.
For deployment, the consequential finding is the possibility of acting without proper sign-off. My reading is that approval enforcement deserves its own validation: checking whether a draft is clinically reasonable will not reveal every failure to stop at an authorization boundary.
Astra and Luna: A Benchmark Score Is Not a Completion Rate
OpenAI also reported results on HealthAdminBench, an administrative-workflow benchmark from Stanford researchers. The benchmark is independent of OpenAI; the model scores and cost figures below are OpenAI’s reported results.
Frequently Asked Questions
3 questions
1Where Are the Healthcare Agents Available?
OpenAI lists ChatGPT for Healthcare, ChatGPT for Clinicians and the OpenAI API. Actual access depends on eligible services, workspace configuration and connected-provider requirements.