OpenAI brings computer-use agents to clinical software with CVS, Cigna, Abridge and Ambience, plus 20+ new plugins and improved clinical search.
- OpenAI launches computer-use agents for healthcare that navigate EHRs and payer portals by reading screens.
- CVS, Cigna, Abridge and Ambience are first deployers across referrals, prior auth and discharge planning.
- One CVS agent cleared a monthslong document backlog over a single weekend in production.
- More than 20 new plugins added including Availity Extend, Oracle HDI and UpToDate Expert AI.
- CVS is open-sourcing its Agent Kit built on OpenAI's Agents SDK for other health systems.
- Available now via ChatGPT for Healthcare, free ChatGPT for Clinicians, and the OpenAI API.
OpenAI is expanding ChatGPT for Healthcare with computer-use agents, improved clinical search, and more than 20 new plugins, according to its announcement. The agents interpret visual interfaces and operate mouse and keyboard controls, allowing them to work across electronic health record systems, payer portals, and other clinical software that may lack usable APIs.
Four partners, four workflows #
- CVS Health: Production agents send documents, process referrals, and support specialist consultations. OpenAI says one agent cleared a months-long document backlog over a weekend. CVS is also open-sourcing its Agent Kit, built on OpenAI's Agents SDK.
- Cigna Healthcare: Agents gather context for member outreach. Cigna says case managers previously navigated more than 10 clinical and administrative systems and can now assemble relevant information in seconds.
- Abridge: Agents in development address inpatient discharge barriers, including medication access and home health coordination.
- Ambience Healthcare: A demonstration showed an agent pausing for clinician input when it encountered an ambiguous patient record.
Plugins connect evidence to action #
The expanded plugin catalog includes UpToDate Expert AI, Availity Extend, and Oracle HDI. Collectively, the integrations cover clinical reference material, revenue-cycle connectivity, and health data infrastructure.
OpenAI says supported workflows include:
- Reviewing patient records and retrieving clinical evidence
- Preparing prescriptions for clinician review
- Coordinating care across hospital and payer systems
- Checking insurance eligibility
- Analyzing costs and operational performance
OpenAI presents the plugins and computer-use agents as a connected workflow: ChatGPT can retrieve relevant information, prepare an action, and operate the application responsible for completing it.
Screen control crosses API gaps #
The release follows OpenAI's September 1, 2026 integration with Epic, which brought authorized patient context into ChatGPT and placed ChatGPT within the EHR interface in supported deployments. That integration supplies clinical context, while computer use extends automation into adjacent applications.
Visual automation can reach older software without dedicated integrations, but it introduces different failure modes. Interface redesigns, pop-ups, expired sessions, latency, and duplicate submissions can interrupt a workflow. Teams evaluating these agents will need application-specific tests and recovery paths for every action that changes a patient record, places an order, or submits information to a payer.
Approval gates carry the safety load #
OpenAI says ChatGPT for Healthcare provides role-based access controls and audit logs. Organizations can configure which applications an agent may access, which actions require approval, and which events are recorded.
An MRI follow-up demonstration uses Voice in the ChatGPT app to examine charts and prepare orders and patient responses for clinician review. The clinician remains the approval gate. Ambience's demonstration applies a related control by pausing when the record is ambiguous and routing the decision to a person.
Before an agent can click Submit
- Limit read and write permissions by role, workflow, and patient context.
- Require confirmation for orders, prescriptions, referrals, and payer submissions.
- Prevent duplicate actions through idempotency checks and submission tracking.
- Record source context, proposed actions, approvals, execution results, and model versions.
- Detect interface changes and session failures, then route incomplete work to a human.
The missing numbers #
OpenAI did not identify the agent models, publish workflow-level error rates, provide comparative benchmarks, or disclose pricing. The announcement also omits intervention rates, completion times, and performance after application interfaces change.
Organizations can purchase ChatGPT for Healthcare through OpenAI sales or private offers on AWS Marketplace. Comparing it with an internal implementation built on Claude, Gemini, or another model will require local measurements of task success, clinical errors, latency, review time, and total cost.
Three routes to access #
| User | Availability | Conditions |
|---|---|---|
| Enterprise customers | Computer-use agents in ChatGPT Work and Codex, improved clinical search, and healthcare plugins | Available as of the announcement, subject to workspace settings and administrator controls |
| Individual clinicians | Computer-use agents in ChatGPT Work and Codex, improved clinical search, and selected plugins through a free ChatGPT for Clinicians account | Requires verification as a U.S. clinician |
| Developers | Computer-use capabilities through the OpenAI API, ChatGPT Work, and Codex | Implementation details are available in the developer guide |
The SDK becomes the template #
CVS Health's open-source Agent Kit gives other healthcare teams a reusable starting point for building on OpenAI's Agents SDK. Teams can inspect and adapt the orchestration layer while retaining their own permissions, approval rules, application mappings, and audit requirements.
OpenAI is betting that computer control can automate fragmented healthcare workflows before every vendor exposes a suitable API. Production viability will depend on consistent task completion, low rates of harmful or duplicate actions, reliable human escalation, and costs that compare favorably with manual work and conventional integrations.