OpenAI Tests Computer-Using Agents on Healthcare Portals Across 325 Tasks
OpenAI described healthcare trials in which computer-use agents interacted with electronic health record inboxes and insurance portals. The reported tasks include navigation, information collection, and preparing administrative actions for review.
What the reported numbers cover
OpenAI reported that clinicians judged 99% of 325 reviewed EHR inbox tasks clinically safe and rated 93% of a separate set of 646 evaluations at the top two helpfulness levels. On the HealthAdminBench administrative benchmark cited in the article, GPT-6 Astra scored 94.3% and GPT-6 Luna scored 91.8%. The article reports a mean cost of $0.085 per task for Luna in that benchmark context.
Limits of the evidence
These figures come from selected tasks and evaluation settings, with different denominators. They do not establish safety across all EHR workflows or autonomous medical decision-making. Real deployments must account for errors, authorization, oversight, and escalating cases that the system cannot resolve.
What organizations can evaluate
Healthcare teams can measure whether a narrowly defined administrative workflow saves time under existing staff review and privacy requirements. Useful metrics include cycle time, correction rates, missed escalations, and total cost including reviewers.