OpenAI pushes computer-use agents into the messy middle of healthcare workflows

OpenAI is moving beyond text-generating copilots and into software操作 that looks far closer to the way a human assistant actually works at a desk. The company is expanding computer-use agents for healthcare, giving AI systems the ability to look at a screen, move through an electronic health record, click through payer portals and complete repetitive steps that still slow clinical operations across the care continuum.

The strategic bet is straightforward: healthcare software remains fragmented enough that many critical tasks still depend on people copying information from one system into another. In day-to-day practice, that means prior authorizations, referrals, discharge planning, specialist consults and case management can require repeated logging in, record hunting and portal navigation before anyone can move a patient forward. OpenAI is positioning computer use as a way to work across those broken seams when application programming interfaces do not exist or do not cover the full workflow.

A weekend backlog clearance at CVS Health captured the operational appeal in one blunt example. In one production workflow, an agent cleared a months-long document-sending backlog over a weekend, a result that speaks less to novelty than to a familiar bottleneck in modern care delivery: tedious administrative work can accumulate until it begins to block treatment decisions.

The appeal lies in execution, not drafting

OpenAI’s health AI lead, Karan Singhal, described computer use as the next step after copilots that summarize or draft text. The aim is not merely to generate a note, but to carry out multi-step work inside the software clinicians already use, while leaving clinicians and staff in control of the final action. In practice, that matters because a draft message or a summarized chart does not by itself move an MRI request, a referral or a medication authorization through a payer workflow.

A research demonstration using synthetic patient data illustrated the concept in a concrete way. A patient with lower back pain messages a clinician, who then asks the agent to review visit notes, physical therapy records and prior imaging inside the EHR. The agent drafts an MRI order, an encounter note and a patient reply for review, then uses the same records to prepare a prior authorization request in the insurer’s portal for an authorization specialist to review. The workflow is not fully autonomous, and it is not intended to be. It is designed to absorb the repetitive interface work that often consumes the time of clinicians and administrative teams.

A single agent handling the routine clicks does not replace clinical judgment, but it can collapse the lag between a decision and the paperwork required to make that decision real.

OpenAI evaluated GPT-6 Astra on EHR inbox tasks shaped by physician input, focusing on multi-step work that required both clinical reasoning over longitudinal context and navigation of complex interfaces. Reviewers assessed the model’s actions along with the notes, orders and patient messages it produced, judging whether the work would be helpful and clinically safe in the context of the task. In that evaluation, 99% of 325 reviewed tasks were rated clinically safe, and physicians rated the model “very helpful” or “extremely helpful” in 93% of 646 assessments.

Clinical safety was judged by whether inaccuracies, omissions or model actions could plausibly cause patient harm. Singhal said the 1% of tasks rated not clinically safe did not yield any specific, plausible risk of patient harm in the review, though the model could overreach or proceed without proper sign-off in some cases. That is a central constraint in healthcare deployment: the risk is often not dramatic error so much as a workflow violation, a missed approval step or an action taken before the right human has reviewed it.

CVS Health, Cigna and Abridge are testing the same idea in different settings

CVS Health is using OpenAI’s computer-use capabilities to automate repetitive work across clinical systems, including document sending, referral processing and specialist consult support. The scale described in the briefing suggests these agents are being aimed at high-volume, low-ceremony work where latency and manual re-entry are especially costly. Another agent reportedly handles the system steps for tens of thousands of consults each day, which signals an operational use case rather than a pilot confined to a small test queue.

Cigna Healthcare is using OpenAI models to pull together member information from more than 10 clinical and administrative systems, including benefits, providers, appointments and prior interactions. The purpose is to reduce the time case managers spend searching across disconnected platforms and to free up more time for direct member support. In a fragmented payer environment, the practical burden is often not the lack of information but the number of places it must be checked before a human can act.

Abridge is building computer-use agents for clinicians who need care plans to follow patients across multiple applications, including systems without APIs. The company also plans an inpatient use case that would combine EHR data with nurse and clinician context to identify what needs attention and coordinate next steps. In the inpatient setting, where medication access, discharge planning and home health coordination can overlap, the ability to keep multiple threads moving inside different software environments could shorten delays that often arise after the clinical decision has already been made.

Safety controls and workflow limits will decide adoption

OpenAI says the agents can be configured with limits on system access and actions, and audit logs are intended to document what the agent did and who approved it. Those controls are not a side issue; they are the difference between an assistive system and a dangerous black box. In healthcare, any agent that touches referrals, authorizations or medication-related work must preserve traceability, because every automated step eventually lands in a chart, a portal or a claims workflow that affects care.

The company is also still working on core performance issues. Singhal pointed to improved latency and costs as important areas for future model progress. Even a useful workflow agent will face resistance if it is too slow, too expensive or too brittle when asked to move across legacy systems that were never designed for machine navigation.

OpenAI’s role, as described in the announcement, is to supply the underlying capability and safeguards while healthcare companies handle the last mile of implementation. That division of labor is sensible, but it also means the clinical value will depend on local workflow design, supervision and the willingness of organizations to trust agents with real operational steps. The unanswered operational question is not whether the software can click through a portal; it is whether those clicks consistently save time without introducing new failure points into already crowded workflows.


Reporting by Dr. A. Nikhil Krishnan, MBBS for Medicine World Foundation, with additional data provided by Fierce Healthcare.