OpenAI is putting computer-use agents into clinical software, so AI can work the same screens that nurses, case managers and billing staff click through every day. The company's healthcare update, reported on November 16, makes the agents available through the OpenAI API, ChatGPT for Healthcare and ChatGPT for Clinicians. Early customers describe them handling document backlogs, referrals and insurance authorizations, the kind of administrative work that consumes large parts of clinical operations.

What the Agents Actually Do

Computer-use agents operate existing software through clicks and keyboard input rather than relying on a clean integration. That matters in healthcare, where many electronic health record and insurer systems offer limited or slow API access. OpenAI's demonstrations show an agent reviewing a patient's visit notes, physical therapy records and prior imaging in an EHR while preparing an MRI request. Another example prepares prior authorization requests through insurer portals. OpenAI's featured demo uses synthetic data.

The distinction between screen-based automation and API integration matters for compliance teams. A screen-operating agent can reach systems that have no connector, but it also inherits every permission the human user has. That makes access controls and logging central to any rollout rather than an afterthought.

Early Customers

CVS Health is the most detailed example in the coverage. One CVS agent cleared a months-long backlog of document sending over a single weekend, and another handles software steps for tens of thousands of specialist consultations each day. CVS also published its CUA Agent Kit under an Apache-2.0 license, which lets other developers study and reuse its approach.

Cigna says its case managers previously had to navigate more than 10 systems to gather the information they need. Its deployment assembles that material for member outreach and went from concept to deployment in under eight weeks. Abridge and Ambience are also named. Abridge's cross-system assistant is live only in early environments with a few select partners, while Ambience's agent prepares insurance authorization requests and asks for clinician input when records are unclear.

The Safety Numbers

OpenAI reports that 99% of 325 reviewed GPT-6 Astra tasks in electronic health record inboxes were rated clinically safe. Reviewers asked whether inaccuracies, omissions or actions could plausibly lead to patient harm. Physicians also rated 93% of 646 assessments as very or extremely helpful. Coverage also cites a benchmark built by Stanford researchers, HealthAdminBench, where GPT-6 Astra scored 94.3% of available points compared with 91.8% for GPT-6 Luna. OpenAI also reports that Astra cost about 8.5 cents per benchmark task, 98% less than Luna.

These are vendor-reported figures on limited samples. A 325-task review shows how the system performed in one evaluation. It does not show how it will behave across every health system, payer portal and documentation style. Readers should treat the headline percentages as starting points for their own testing.

Oversight Is Part of the Product

OpenAI frames the agents as assistants for care teams rather than decision-makers. ChatGPT for Healthcare includes role-based access controls and audit logs. Organizations can add approval steps and limit an agent to the applications, records and actions each task requires. OpenAI also says data from ChatGPT for Healthcare is not used to train its foundation models.

Deployments show how those controls work in practice. Ambience's agent pauses when a record is ambiguous and waits for approval before submitting. Health systems can restrict permitted actions, pause an agent and review a step-by-step audit log. CVS relies on validation checks, human review, screenshots and action records. Its repository also warns that run artifacts are not sanitized and that redaction is best-effort, a useful reminder that logs can themselves contain sensitive data.

Why It Matters

Healthcare administration has long been slowed by fragmented systems that do not talk to each other. Agents that work through the user interface could reduce that friction faster than waiting for every vendor to build integrations. The same design also widens the error surface: an agent that can click submit can also submit the wrong form, to the wrong payer, at scale.

For readers following enterprise agents, the healthcare rollout is a useful test case. The strongest claims concern speed and workload. The most important questions concern audit trails, error rates and who is accountable when an agent gets a step wrong. Those answers will matter more than any single benchmark score.

I could not open OpenAI's own announcement page during reporting, so model naming and some figures rest on secondary coverage. Check the company's original posts before relying on specific numbers in a procurement or clinical decision.

Sources