OpenAI is making a new push into one of healthcare's least glamorous but most expensive problems: moving clinical and insurance paperwork between incompatible software systems. The company says computer-use agents can navigate existing interfaces to help prepare authorizations, referrals and other administrative work. Its new benchmarks look promising, but they measure several different things.

In a November 16 announcement, OpenAI described work with Abridge, Ambience, Cigna Healthcare and CVS Health. Computer-use agents can interact with software through screens, clicks and text entry, rather than requiring a conventional application programming interface for every step. That makes them relevant where one task spans a patient record, an insurer portal and a document system.

From the patient record to the insurer portal

OpenAI's demonstration uses synthetic patient data. The agent retrieves medical information, prepares a draft order, encounter note and message for clinical review, then assembles materials for a prior authorization request. The clinician and authorization specialist were expected to review the relevant work. A successful demonstration with synthetic data is not proof that the process can be trusted without oversight in everyday care.

CVS Health is also using computer-use tools for cross-system administrative work, according to the announcement. OpenAI says CVS Health uses the capability for sending documents, processing referrals and supporting specialist consultations. These are company-reported uses, not independent measurements of error rates or patient outcomes.

Two impressive percentages answer different questions

OpenAI reports that 99% of 325 reviewed clinical-inbox tasks were judged safe under its evaluation criteria. It also reports favorable physician helpfulness ratings in 93% of 646 assessments. Those figures describe reviews of model behavior on selected tasks; they are not a measured population-level reduction in patient harm, nor a certification of clinical accuracy.

For administration, OpenAI says GPT-6 Astra achieved 94.3% of available rubric points on HealthAdminBench. GPT-6 Luna achieved 91.8%, with an estimated model cost of $0.085 per task, or 98% lower cost in the company's comparison. These are company-reported scores on an independently developed benchmark. A rubric-point percentage is particularly easy to misread: it does not necessarily mean the system finished that percentage of entire authorizations correctly.

That caution is supported by the original HealthAdminBench research, developed by Stanford-affiliated investigators in April 2026. Their benchmark covers 135 administrative tasks across an electronic health record, two payer portals and a fax interface. In the earlier study, the best-performing tested agent completed only 36.3% of entire tasks, even while some agents performed much better on individual steps. Those are older models and different evaluations, so the percentage cannot simply be subtracted from OpenAI's new scores to claim a particular improvement.

The last mile determines whether an AI agent saves work

Consider an insurance authorization with ten required steps. An agent could identify the correct policy and fill most fields, but still fail to attach a document or submit the request to the correct destination. Its partial score might look strong even though a human must repeat the process. This is why healthcare buyers need to track finished transactions, rework, exceptions and final system state alongside benchmark results.

There is also an important distinction between preparing material and making a clinical decision. An agent might help gather supporting information, but whether imaging, medication or specialist care is appropriate remains a professional judgment governed by relevant clinical and organizational controls. OpenAI describes administrative controls for agent permissions, review and action records. Regulatory compliance and data protection depend on the specific deployment, contracts, access controls and operations.

These reliability questions echo our earlier examination of AI agent authorization: the fact that an agent can use a tool does not by itself show it was entitled to make the change.

What healthcare organizations should test next

The most useful pilot would measure a representative set of complete workflows from start to finish, including failed documents, missing attachments and human handoffs. Providers should record what information an agent accessed, what it changed, which steps were verified, and when a licensed professional took responsibility. Privacy and contractual controls must cover every system in the chain, not just the AI endpoint.

OpenAI's announcement is a genuine sign that computer-use agents are moving beyond demonstrations into administrative operations. Their economic promise is credible because healthcare teams spend substantial time transferring information. The unanswered question is whether the whole task becomes reliably easier, not whether a model can impressively complete most of its steps.

Sources and measurement notes

Image: Original AI-assisted conceptual illustration by Nextwith.ai, November 18, 2026. No real patient data or product screenshot is depicted.