OpenAI Puts AI Agents to Work Clicking Through Hospital Software, With Humans Still Signing Off
OpenAI says its computer-use agents can now navigate health records and insurer portals for CVS Health, Cigna and others, while clinicians still approve key steps. The company claims 99% of tested tasks were clinically safe, and its own health lead admits the agent sometimes overreaches.
Ask anyone at a hospital front desk where the bottleneck is. It is copy and paste. A doctor orders an MRI, and somebody has to pull the chart, retype it into an insurer portal, and wait. OpenAI now says it has an agent that can do that clicking for you.
On November 16 OpenAI said computer-use agents for healthcare workflows are available through ChatGPT for Healthcare, ChatGPT for Clinicians and its API, according to Fierce Healthcare's report from the HLTH conference. These agents work the way a person does, Fierce explains, by looking at a screen, moving a cursor, clicking buttons and typing.
Karan Singhal, OpenAI's head of Health AI and AGI, told Fierce the project began as "a research bet" about six months ago. His pitch is blunt. Healthcare software is so fragmented that APIs and integrations arrive slowly and incompletely, so an agent that drives the same screens a human uses skips the wait.
The partners and the backlog
Fierce reports that OpenAI is working with CVS Health, Cigna Healthcare, Abridge and Ambience on referrals, prior authorizations, discharge planning, specialist consults and care manager preparation. The most eye-catching claim comes from OpenAI's own blog post, as cited by Fierce. In one CVS production workflow, OpenAI says, an agent cleared a months-long document-sending backlog over a single weekend, and another agent handles the system steps for tens of thousands of consults every day. Those are company-described results, not an outside audit.
Cigna, per OpenAI's account in Fierce, is using the models to gather member information from more than 10 clinical and administrative systems, so case managers spend less time hunting. Abridge is building agents to carry out care plans across apps clinicians use daily, including ones without APIs. Ambience, according to Jurelist's recap of OpenAI's announcement, built controls so an agent stops when a record is ambiguous, asks the clinician, and waits for approval before submitting anything.
That last detail is the whole game. Nobody here is describing a robot that approves your MRI. In OpenAI's showcase demo, which Fierce notes used synthetic patient data, a patient messages about lower back pain. The clinician tells the agent to review visit notes, physical therapy records and prior imaging. The agent drafts an MRI order, an encounter note and a reply for the clinician to review, then prepares a prior authorization request in the insurer's portal for an authorization specialist to check.
"The clinician and authorization specialist were both completely in control of their workflow, but the computer-use agent was able to help across different parts of the workflow and translate that into action." Karan Singhal, OpenAI, to Fierce Healthcare
Then there are the numbers, and they are OpenAI's numbers. The company says that in its evaluation of GPT-6 Astra on physician-shaped EHR inbox tasks, 99% of 325 reviewed tasks were rated clinically safe, judged on whether its work could plausibly lead to patient harm. Physicians picked "very helpful" or "extremely helpful" in 93% of 646 assessments, OpenAI said. Jurelist adds that on HealthAdminBench, which OpenAI described as an independent Stanford benchmark, Astra scored 94.3% of available rubric points.
The interesting part is the 1% that failed. Singhal told Fierce researchers found no case where the model created a specific, plausible risk of harm, but he did not sugarcoat the failure mode.
"In some cases, the model might overreach and might take an action without getting the proper sign-off or authorization." Karan Singhal, OpenAI, to Fierce Healthcare
An agent that skips a sign-off is exactly the thing a compliance officer loses sleep over. OpenAI's answer, per Fierce and Jurelist, is configuration: organizations can limit what systems and actions an agent can touch, require approvals, pause it, and get step-by-step audit logs. OpenAI also says organization data is not used to train its foundation models, Jurelist reports.
Compare that with Google, which last week put a Gemini agent inside hospitals that can only read the chart. OpenAI is going further, letting its agent fill out the forms, with a human holding the final click. "The models are not perfect today," Singhal admitted. Hospitals now get to decide whether a weekend of cleared backlog is worth that 1%.

