AI agent cursor moving through abstract hospital record screens and authorization forms in a modern medical office

OpenAI just put its AI agents inside the software hospitals actually run on. Announced on November 16, the company’s computer-use agents can click, type and move between applications to handle the paperwork that sits between care and payment — referrals, prior authorizations, discharge planning and outreach between visits — and they are available today through the OpenAI API, ChatGPT for Healthcare and ChatGPT for Clinicians.

This is not a chatbot answering health questions. It is an agent doing the retyping: pulling information out of a patient’s record in one system and entering it into an insurer’s portal in another, exactly the manual step OpenAI says still blocks care today.

The safety numbers behind the launch

OpenAI evaluated GPT-6 Astra on the multi-step tasks clinicians handle in their electronic health record inboxes. Physicians rated 99% of 325 reviewed tasks clinically safe, and rated the work “very helpful” or “extremely helpful” in 93% of 646 assessments.

On Stanford’s independent HealthAdminBench, which covers prior authorizations, appeals and medical-equipment orders, Astra earned 94.3% of available rubric points. The cheaper GPT-6 Luna was close behind at 91.8% — at just 8.5 cents per task, which OpenAI says is 98% less than Astra.

As OpenAI put it in its announcement: “Computer-use agents carry out multi-step workflows by clicking, typing, and moving between applications.”

CVS cleared a months-long backlog in a weekend

The launch partners are Abridge, Ambience, Cigna Healthcare and CVS Health. CVS is the clearest production case: one agent cleared a months-long document-sending backlog over a single weekend, and another handles the system steps behind tens of thousands of specialist consults every day. CVS is also open-sourcing its Agent Kit, built with the OpenAI Agents SDK.

Cigna’s care managers previously had to navigate more than 10 separate clinical and administrative systems just to gather context on a member. The agents now pull that context together in seconds. Ambience, meanwhile, built the control layer — when a record is ambiguous, the agent stops, asks the clinician, and waits for approval before submitting.

OpenAI says organization data is not used to train its models, and HIPAA-covered use requires a Business Associate Agreement and the required security settings, with role-based access, audit logs and human approval steps built in.

Why it matters

Healthcare is the hardest test yet for agent software, because the work is fragmented, regulated and consequential — the same week Apple asked 200,000 iPhone users to help test AI health answers before it trusts them.

It is also the second time this month OpenAI has given its agents a real-world job with a human approving every consequential step — the same pattern as ChatGPT Wallet, where an agent can pay only if you say yes every time. If the 99% safety figure holds up in production, the humble authorization form may be where AI agents quietly become infrastructure.

Source: OpenAI — “Advancing computer-use agents for healthcare teams” (Nov 16, 2026): https://openai.com/index/advancing-computer-use-for-healthcare-teams/

Mohsen Daemi