Tools

AI-generated text

OpenAI Tests Computer-Use Agents for Healthcare Workflows

OpenAI is developing computer-use agents to help healthcare teams move information between clinical and administrative systems.

OpenAI Tests Computer-Use Agents for Healthcare Workflows

OpenAI is developing computer-use AI agents that can perform multistep tasks by clicking, typing and moving between applications already used by healthcare workers. The company has published safety, performance and cost results from clinical and administrative workflows and described its work with healthcare organizations to apply the technology.

Healthcare information often sits in separate systems. When a doctor recommends an MRI, for example, the information needed for an insurer’s authorization request may already be in the patient’s record, but a staff member must transfer it to a separate portal. Agents can help with such repetitive steps, including in applications that cannot be fully connected through APIs.

Clinical and administrative tests

In an OpenAI demonstration using synthetic patient data, a patient contacts a clinician about lower back pain. The agent reviews visit notes, physical therapy records and prior imaging in the electronic health record, then prepares an MRI order, an encounter note and a reply to the patient for clinical review. It also uses the same information to prepare a prior-authorization request in the insurer’s portal for an authorization specialist to check.

The company also evaluated GPT-6 Astra on tasks physicians encounter when handling messages in electronic health records. Developed with physician input, the tasks tested both clinical reasoning based on a patient’s history and navigation of complex interfaces. Reviewers examined the model’s actions as well as the notes, orders and patient messages it produced.

Clinical safety was assessed by considering whether inaccuracies, omissions or actions could plausibly harm a patient. Of 325 reviewed tasks, 99% were rated safe. In 93% of 646 assessments, physicians rated the work “very helpful” or “extremely helpful” on a five-point scale.

Administrative tests covered prior authorizations, appeals and medical equipment orders. On HealthAdminBench, an independent benchmark developed by Stanford researchers, GPT-6 Astra earned 94.3% of available rubric points and GPT-6 Luna earned 91.8%. OpenAI said Luna cost 8.5 cents per task, 98% less; both models were tested at maximum reasoning effort.

Healthcare partners’ applications

CVS Health is using OpenAI’s computer-use capabilities to automate repetitive steps across its clinical systems, including document delivery, referral processing and support for specialist consultations. In one production workflow, an agent cleared a months-long document-sending backlog over a weekend; another handles system steps for tens of thousands of consultations each day. The company uses validation checks and human oversight, with screenshots and action logs to verify the work. It is also open-sourcing its Agent Kit, built with the OpenAI Agents SDK. Data remains within CVS Health’s enterprise-controlled infrastructure under its AI governance team and policies; production data is not used to train foundation models.

Abridge is developing agents to help clinicians carry out care plans in the applications they use every day. Its planned inpatient application would combine the clinical conversation with health-record information and context from nurses and clinicians. In the example presented, the Abridge Assistant could work concurrently on medication access and home-health coordination, using computer interaction, calls and messages. Clinicians would still be brought in for decisions requiring clinical judgment.

As part of a broader collaboration between The Cigna Group and OpenAI, Cigna Healthcare is using OpenAI’s advanced models alongside its own clinical expertise to bring together information from different systems. That can include benefits, coverage, providers, appointments and previous interactions. The company said case managers previously had to search more than ten clinical and administrative systems; computer use can help gather relevant information in seconds.

Ambience has built additional controls to help health systems oversee agents’ actions. In its demonstration, an agent uses evidence from a patient’s record to prepare a prior-authorization request on an insurer’s website. If the record is ambiguous, it stops and asks the clinician, and it waits for approval before submitting. Health systems can define permitted actions, pause the agent at any time and receive a step-by-step audit log.

Access and data protection

OpenAI said healthcare applications require not only clinical safety but also protection of patient information and limits on what agents can do without human review. ChatGPT for Healthcare provides role-based access controls and audit logs. Organizations can add review and approval steps and restrict the applications, environments and records agents can access. OpenAI said it does not use organization data to train its foundation models.

For HIPAA-compliant workflows involving protected health information, organizations must use eligible services covered by their business associate agreement with OpenAI and apply the required security and retention settings.

Computer-use capabilities are available through ChatGPT for Healthcare, ChatGPT for Clinicians and the OpenAI API. Existing organizational customers can ask their workspace administrator to enable the feature in ChatGPT Work and Codex. Clinicians in the United States can create a free ChatGPT for Clinicians account, while developers can build applications using the OpenAI API.