OpenAI Adds Computer-Use Agents to ChatGPT for Healthcare, With Human Oversight
November 16, 2026
OpenAI has expanded ChatGPT for Healthcare with computer-use agents, improved clinical search and more than 20 integrations, aiming to help healthcare organizations complete clinical and administrative work across disconnected systems.
Unlike tools that only draft or summarize, the agents use clicks and keyboard inputs to operate existing software—including systems without complete API support—while leaving clinicians and staff in control.
The offering brings together three distinct capabilities: clinical search retrieves evidence, plugins connect to specific services, and computer-use agents operate software interfaces; their permissions and supporting evidence are not interchangeable.
The integrations span tasks such as checking insurance eligibility with TriZetto, confirming appointments through Luma Health, accessing discharge information through Metriport, and estimating procedure prices with Turquoise Health.
One example is Elation Health’s read-only beta, which helps practice managers find next-day patients with missing information so staff can resolve it in the electronic health record; clinical decisions and patient outreach remain with practice teams.
The 21 plugins have different access and action limits: some are read-only, some support administrative transactions, and certain clinical tasks require professional authorization or verification.
Organizations can limit which applications agents access, require human review for selected actions and inspect activity logs; a demonstration showed an agent preparing MRI follow-up orders and patient responses for clinician review.
Because agents act through browser or desktop interfaces, changing screens, timeouts, unexpected data and validation errors can derail a workflow, and a screenshot or confident completion message does not prove that another system accepted the action.
Teams must distinguish between work that is prepared, signed, submitted, received, approved and scheduled, then verify each system’s status and assign owners to unresolved steps to avoid missed or duplicate actions.
OpenAI reported that GPT-6 Astra was rated clinically safe on 99% of 325 EHR inbox tasks and very or extremely helpful in 93% of 646 physician assessments; on HealthAdminBench, Astra scored 94.3% and GPT-6 Luna 91.8% of available rubric points.
Those benchmark results need context: HealthAdminBench covers 135 tasks in synthetic environments that omit obstacles such as CAPTCHA, multi-factor authentication and session timeouts, and the article says it could not locate raw traces, confidence intervals, full task-selection methods or independent replication for the new runs.
OpenAI said tasks rated not clinically safe did not involve an identified, specific and plausible risk of patient harm, though some raised the possibility of action without appropriate sign-off or authorization.
Summary based on 12 sources
Get a daily email with more AI stories
Sources

OpenAI • Nov 16, 2026
More ways to move care forward with ChatGPT for Healthcare
Fierce Healthcare • Nov 16, 2026
HLTH 2026: CVS Health, Cigna and Abridge collaborating with OpenAI to test computer-use agents
Interesting Engineering • Nov 16, 2026
OpenAI agents enter US healthcare systems after scoring 99% on clinical safety reviews