OpenAI announced healthcare computer-use agents and work with CVS Health, Cigna Healthcare, Abridge and Ambience on November 16, 2026. The examples span production workflows, a deployment, development and demonstrations; the reported evaluation scores measure task performance, not patient outcomes. The agents navigate existing software by reading screens, clicking and typing, including workflow steps that APIs may not cover. OpenAI’s announcement describes the projects and results. The announcement builds on OpenAI’s computer-use work with GPT-6 Astra, which NeoTeo covered in an earlier article on GPT-6 Astra’s computer-use launch.

Treat each example according to its stage: production use, deployment, development and demonstration are not interchangeable.

Where the partner workflows stand

OpenAI describes different levels of activity among the four organizations. CVS Health reports production workflows, Cigna Healthcare describes a deployment, Abridge is developing agents, and Ambience’s cited authorization workflow is a demonstration.

OrganizationStage describedWorkflow and human role
CVS HealthProduction useAgents send documents, process referrals and support specialist consults. Validation checks and human oversight use screenshots and action records to verify actions and application state.
Cigna HealthcareDeploymentComputer use gathers member context across more than 10 clinical and administrative systems for case managers.
AbridgeDevelopment; inpatient use case plannedAgents are being developed to help carry out care plans across applications. The planned inpatient workflow would use EHR, nurse and clinician context to identify needs and coordinate next steps; clinicians remain involved in decisions requiring clinical judgment.
AmbienceDemonstrationAn agent prepares a prior-authorization request from chart information, pauses for clinician input if the record is ambiguous, and waits for approval before submission.

OpenAI says a CVS workflow cleared a months-long document backlog over a weekend, while another handles system steps for tens of thousands of specialist consults each day. Cigna Healthcare CIO Madhu Nutakki said the move from concept to deployment took under eight weeks, according to OpenAI’s announcement.

OpenAI also described an MRI workflow using synthetic patient data. The agent drafted an MRI order, encounter note and patient reply for clinician review, then prepared an insurer authorization request for an authorization specialist. The example shows how computer use can connect administrative steps while leaving those reviews to people.

What the reported evaluations measure

OpenAI reported that 99% of 325 reviewed GPT-6 Astra EHR tasks were rated safe. Reviewers assessed the agent’s actions alongside generated notes, orders and patient messages, considering whether inaccuracies, omissions or actions could plausibly lead to patient harm. In a separate set of 646 physician assessments, 93% rated the model’s work very or extremely helpful. These are task and assessment ratings, not measures of patient outcomes.

OpenAI also reported results on HealthAdminBench, which it describes as an independent benchmark developed by Stanford researchers for administrative workflows. The scores represent shares of available rubric points, and both models were evaluated at maximum reasoning effort.

ModelShare of available rubric pointsEvaluation condition
GPT-6 Astra94.3%Maximum reasoning effort
GPT-6 Luna91.8%Maximum reasoning effort

The HealthAdminBench scores measure administrative tasks; they are separate from the clinical EHR task ratings. OpenAI’s figures describe these specific evaluations rather than results for patients across healthcare settings.

Data protections and access

OpenAI says protected health information workflows require eligible services covered by an OpenAI Business Associate Agreement, required security and retention settings, and protections for connected systems. Using an eligible OpenAI service alone does not make every connected system or workflow HIPAA-compliant. In its statements about data use, OpenAI says organization data in ChatGPT for Healthcare is not used to train its foundation models. Separately, it says CVS production data remains within CVS-controlled enterprise infrastructure and is not used to train foundation models. OpenAI’s healthcare information describes the stated conditions.

OpenAI lists ChatGPT for Healthcare, ChatGPT for Clinicians and the OpenAI API as access routes for computer use. It says individual clinicians in the United States can sign up for free; organizations and developers have separate access routes.