OpenAI is expanding computer-use agents into healthcare workflows, where they can navigate electronic health records and insurer portals to prepare work for clinical staff. The significant shift is from AI that drafts or summarises to AI that can carry tasks across disconnected systems, with people still expected to review key actions.
OpenAI Watch analysis
What happened
OpenAI is working with CVS Health, Cigna Healthcare, Abridge and Ambience on workflows including referrals, prior authorisations, discharge planning and specialist consultations. In a research demonstration using synthetic patient data, an agent prepared an MRI order, clinical note and patient reply for a clinician to review, then assembled an insurance authorisation request for a specialist to review.
OpenAI says its GPT-6 Astra model was rated clinically safe in 99% of 325 reviewed EHR tasks, and physicians judged it very or extremely helpful in 93% of 646 assessments. On Stanford’s HealthAdminBench, OpenAI reports Astra scored 94.3% of available rubric points. These are company-reported evaluations, not evidence that every workflow is safe in everyday care.
Why it matters
Healthcare work often crosses systems that do not connect neatly. Computer use could let an agent handle repetitive steps even where no useful software integration exists, potentially freeing staff from moving information between screens. The stakes are higher than a particularly obedient office assistant: a mistaken action in a care workflow can matter.
OpenAI also describes CVS using agents in production. One reportedly cleared a months-long document-sending backlog over a weekend; another handles system steps for tens of thousands of consultations each day. Those examples suggest deployment is already moving beyond demonstrations, though the account does not give independent outcome measures.
Our read
This is a more consequential test of AI agents than getting one to click through a toy task. OpenAI’s figures are encouraging, but “clinically safe” in a reviewed evaluation is not the same as proven safe across the unruly variety of real clinical systems. The useful dividing line is whether people can inspect, approve and correct the work before it affects care. Conveniently, that is also where the less glamorous product details become the most important ones.
What to watch
- Whether partner organisations publish results on accuracy, time saved and errors in routine use.
- Which actions remain subject to clinician or specialist approval.
- How access limits and audit logs work across different systems.
- Whether performance holds up beyond the tasks and data used in evaluations.
Discussion spark: Should healthcare agents be allowed to complete routine actions across systems once staff can review them, or should every step require approval before it happens?
Sources and evidence
- HLTH26: CVS Health, Cigna and Abridge collaborating with OpenAI to test computer-use agents – Fierce Healthcare (16 November 2026, 14:05 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.