Computer-use agents, clinical search and plugins

OpenAI announced three updates to "ChatGPT for Healthcare," its offering for healthcare organizations, in an official blog post dated Nov. 16. The first is computer-use agents, which carry out multi-step work by reading the screens of software care teams already use, clicking, typing and moving between applications. The second is an improved "trusted clinical search," which finds, evaluates and cites medical evidence. The third is more than 20 new healthcare plugins (add-ons that connect to outside services).

OpenAI said healthcare teams need to piece together information from the EHR, hospital policies, scheduling systems and payer portals to do their work. It said the features are meant to help them spend less time navigating systems and more time caring for patients. For computer-use agents, OpenAI said it is offering more capable, safe and affordable models for clinical and administrative workflows, along with access controls and audit logs for secure deployments. It is working with Abridge, Ambience, Cigna Healthcare, CVS Health and others to process referrals, move documents between systems and bring together the information care teams need.

Other companies are also building screen-operating AI into their products. On Oct. 30, Google made "computer use," the Gemini API's screen-control feature, generally available (ORIENS article). In October, Cursor added a feature that lets its AI agents operate desktop apps (ORIENS article).

GPT-6 Astra rated safe on 99% of tasks in OpenAI's evaluation

OpenAI evaluated "GPT-6 Astra" on tasks clinicians handle in their EHR inboxes. The tasks were shaped by physician input and require both clinical reasoning over a patient's longitudinal history and navigation of complex interfaces. Reviewers examined the model's actions along with the notes, orders and patient messages it produced, and judged whether the work would be helpful and clinically safe.

According to OpenAI, 99% of 325 reviewed tasks were rated safe. Safety was judged by whether inaccuracies, omissions or actions could plausibly lead to patient harm. Physicians rated helpfulness on a five-point scale and selected "very helpful" or "extremely helpful" in 93% of 646 assessments.

For administrative work, OpenAI evaluated tasks such as prior authorizations (approvals an insurer gives before paying for a treatment), appeals and medical equipment orders. On "HealthAdminBench," an independent benchmark from Stanford researchers, GPT-6 Astra earned 94.3% of available rubric points. "GPT-6 Luna" earned 91.8%, at a cost 98% lower than Astra's, or 8.5 cents per task. The blog post does not give Astra's cost per task. Both models were evaluated at maximum reasoning effort. OpenAI ran and reported these results; the benchmark itself is external, but the blog post does not say a third party verified the results.

From a patient's message to an MRI prior authorization

OpenAI published a demo using synthetic patient data. When a patient messages their clinician about lower back pain, the agent reviews visit notes, physical therapy records and prior imaging in the EHR. It then drafts an MRI order, an encounter note and a reply to the patient for the clinician to review. The agent uses the same records to prepare a prior authorization request in the insurer's portal, which an authorization specialist reviews.

Clinical search beats the previous version on all eight dimensions in a blind physician review

The clinical search improvements apply to both ChatGPT for Healthcare and "ChatGPT for Clinicians," the version for clinicians. OpenAI said they help clinicians connect relevant evidence to clinical questions and translate it into practical next steps, with citations they can inspect.

According to OpenAI, physicians reviewed responses to real-world clinical questions across multiple specialties without knowing which version produced them. The evaluation included multi-turn conversations. In more than 1,300 reviews, physicians rated the updated experience more highly on all eight dimensions evaluated, including factual correctness, safety, actionability and source authoritativeness. The blog post does not give scores for each dimension or the size of the differences.

Blog lists 21 healthcare plugins

OpenAI said it added "more than 20" new healthcare plugins, and the list in the blog post names 21. As written in the list, they are Aiva, Amion, Assort Health, Availity Extend, Elation Health, Ferry, IMO Health MCP Server, Infinitus Relay, K Health, Luma Health, Metriport, Notable Integration Specs, Oracle HDI Overview, PenguinAI Medical Coding, Photon, Redox, Stedi, Subsalt, Trizetto, Turquoise and UpToDate Expert AI. All are available the same day in ChatGPT for Healthcare, including through "ChartGPT," which brings ChatGPT into the EHR workflow, OpenAI said.

It cited uses such as reviewing patient records and hospital protocols, preparing prescriptions for clinician review, coordinating patient outreach, checking insurance eligibility and analyzing healthcare costs and operational performance. In one example, a user finds tomorrow's appointments that still need confirmation with Luma Health, checks those patients' insurance eligibility with Trizetto and builds a list of what to confirm before each visit. In another, for a patient hospitalized elsewhere for heart failure, a user finds the discharge summary with Metriport and looks up the follow-up monitoring UpToDate recommends for the changed medications.

CVS Health clears a months-long document-sending backlog over a weekend

Deployment examples appear in a related blog post, "Advancing computer-use agents for healthcare teams." CVS Health is using OpenAI's computer-use capabilities to automate repetitive work across clinical systems. In work such as sending documents, processing referrals and supporting specialist consults, agents operate existing applications to complete steps that APIs do not fully support.

According to the blog post, in one production workflow an agent cleared a months-long document-sending backlog over a weekend. Another handles the system steps for tens of thousands of consults each day. CVS Health combines validation checks with human oversight and uses screenshots and action records to verify actions and application state. The company is open sourcing its "Agent Kit," built with OpenAI's "Agents SDK." The blog post describes it as packaging this approach as a repeatable process. "Less time on administrative work means more time for care," said Lagan Khare, executive director of engineering at CVS Health. Data stays within infrastructure controlled by CVS Health, and production data is not used to train foundation models, the post says.

Supplement: According to the README of "cua-agent-kit," the GitHub repository CVS Health published, the kit first has an OpenAI computer-use agent follow a workflow written in plain language while recording screenshots and actions. After successful runs, a Codex skill uses those records to build code for a fixed procedure that runs without a large language model. The README says that asking a model to reason through the same steps on every run, once a workflow is understood, adds inference cost and variability. Because instructions including login credentials and screenshots are sent to the model endpoint during discovery, it asks users to use an endpoint approved for the data involved.

Cigna Healthcare brings together information from more than 10 systems in seconds

Cigna Healthcare is using OpenAI's frontier models as part of The Cigna Group's broader collaboration with OpenAI. Its case managers, who support members' care, previously had to navigate more than 10 separate clinical and administrative systems to gather information such as benefits, providers, appointments and prior interactions. According to the blog post, computer use can now help navigate these systems and bring that information together in seconds. Madhu Nutakki, Cigna Healthcare's chief information officer, described "Moving from concept to deployment in under eight weeks."

Abridge targets discharge barriers, Ambience builds in stops for review

Abridge is developing agents to help clinicians carry out care plans in the applications they use daily, including those without APIs. In a planned inpatient use case, a clinician can ask the Abridge Assistant to help resolve a patient's discharge barriers. Agents work concurrently on medication access and home-health coordination, and the clinician is brought in for decisions requiring clinical judgment.

Ambience built additional controls to make computer use controllable and auditable. In its demo, the agent uses evidence from the chart to prepare a prior authorization request on an insurer's website. When the record is ambiguous, the agent stops and asks the clinician, then waits for approval before submitting. Health systems set which actions the agent can take, can pause it at any point and get a step-by-step log for audit.

According to the blog post, Assort uses computer use in Codex to complete a referral step in the Athena EHR interface that its API integration cannot close.

Access controls and audit logs; HIPAA workflows require BAA-covered services

ChatGPT for Healthcare provides role-based access controls and audit logs. Organizations can build review and approval steps into their workflows and configure the applications and environments agents use to limit access to the records and actions needed for each task. Organization data is not used to train OpenAI's foundation models, the company said.

To support HIPAA-compliant workflows involving protected health information under the U.S. health privacy law, organizations must use eligible services covered by their Business Associate Agreement (BAA) with OpenAI and apply the required security and retention settings, OpenAI said. They must also extend protections to connected systems, it said. Both ChatGPT for Healthcare and ChatGPT for Clinicians offer customers a BAA.

Existing customers and U.S. clinicians get access the same day; no pricing given

ChatGPT for Healthcare customers can use computer-use agents in ChatGPT Work and Codex and the improved clinical search the same day, subject to workspace settings. New plugins are enabled by workspace administrators.

ChatGPT for Clinicians is free for eligible, verified clinicians in the U.S. Individual clinicians can use computer-use agents in ChatGPT Work and Codex and the clinical search with a free account the same day, and select plugins are also available. Developers can use the computer-use capability in the OpenAI API, ChatGPT Work and Codex.

ChatGPT for Healthcare can be purchased directly through OpenAI's sales team or through an AWS Marketplace private offer. Neither blog post gives pricing.