HomeTech NewsOpenAI Opens Computer-Use Agents to Healthcare Teams

OpenAI Opens Computer-Use Agents to Healthcare Teams

OpenAI has opened computer-use agents to healthcare teams, saying the tools can click, type and move between the applications those teams already use, including for tasks that application programming interfaces do not fully cover. In a 16 November post, the company said healthcare teams, developers and individual clinicians can use the capabilities today through the OpenAI API, ChatGPT for Healthcare and ChatGPT for Clinicians. A companion post the same day presented the agents together with updates to clinical search and more than 20 new healthcare plugins.

The company illustrated the idea with a demonstration of a patient message about lower back pain that leads to a draft MRI order and a prior-authorisation request. OpenAI said that demo used synthetic patient data. It also said organisations can configure approvals and agent access, and can record audit logs. To support HIPAA-compliant workflows involving protected health information, it said, organisations must use eligible services covered by their Business Associate Agreement with OpenAI, together with the required security and retention settings. The companion post added that those protections should be extended to connected systems. Organisation data, OpenAI said, is not used to train its foundation models.

What OpenAI reports from its own tests

OpenAI said it evaluated GPT-6 Astra on tasks clinicians handle in their electronic health record inboxes. Reviewers, it said, examined the model’s actions alongside the notes, orders and patient messages it produced. Clinical safety was assessed by asking whether inaccuracies, omissions or actions could plausibly lead to patient harm. In OpenAI’s own evaluation, which has not been peer-reviewed, 99% of 325 reviewed tasks were rated safe, and physicians selected “very helpful” or “extremely helpful” in 93% of 646 assessments, on a five-point scale.

Separately, OpenAI reported scores on HealthAdminBench, an independent benchmark from Stanford researchers that measures performance on common administrative workflows such as prior authorisations, appeals and medical-equipment orders. OpenAI said that, at maximum reasoning effort, GPT-6 Astra earned 94.3% of available rubric points and GPT-6 Luna earned 91.8%, “despite costing 98% less at 8.5 cents per task.” Those figures are OpenAI’s results on the Stanford benchmark, not a score published independently by the benchmark’s authors in the announcement.

The companion post also said that, in a blind evaluation of an updated clinical-search experience, physicians rated the new version more highly across all eight dimensions evaluated, in more than 1,300 reviews. That comparison, too, is OpenAI’s own evaluation, and it is not peer-reviewed.

GPT-6 Astra and GPT-6 Luna are the same model family already described on TSN when GPT-6 reached every ChatGPT tier. The new healthcare agents sit alongside earlier OpenAI work on software-using models, including GPT-6.1 Terra, a cheaper model built for computer use, and the Businesses in ChatGPT and agents beta. They are also a long way from TSN’s earlier account of people asking ChatGPT informal health questions: this launch is aimed at teams working inside clinical and administrative software, not at consumer symptom queries.

What the partners say they are doing

OpenAI said it is working with Abridge, Ambience, Cigna Healthcare and CVS Health. The examples below are the partners’ claims, as presented in OpenAI’s post.

CVS Health said that, in one production workflow, an agent “cleared a months-long document-sending backlog over a weekend,” and that another handles the system steps for “tens of thousands of consults each day.” It said it uses validation checks and human oversight, with screenshots and action records, and that production data is not used to train foundation models. CVS has also open-sourced CUA Agent Kit, which it describes as “a starter kit for automating browser and Windows workflows, using AI to discover the steps, then code to repeat them,” built with OpenAI’s Agents SDK. Lagan Khare, executive director of engineering at CVS Health, said: “Less time on administrative work means more time for care. That’s a win for our teams and everyone we serve.”

Cigna Healthcare said its case managers previously had to navigate “more than 10 separate clinical and administrative systems” to gather context before member outreach, and that computer use can now help bring that information together. Madhu Nutakki, the company’s chief information officer, said: “Moving from concept to deployment in under eight weeks shows what’s possible when we bring together AI, our expertise in care management, and the right guardrails for responsible use.”

Abridge described a planned inpatient use case in which clinicians would ask its assistant to help resolve a patient’s discharge barriers, with agents working on steps such as medication access and home-health coordination. That is a planned use, not a claim of a completed deployment. Ambience described a demonstration in which its software prepares a prior-authorisation request and, when the record is ambiguous, stops and asks the clinician before submitting.

ChatGPT for Clinicians, the companion post said, is free for eligible, verified clinicians in the United States. ChatGPT for Healthcare is the governed workspace for organisations. Both, OpenAI said, can be covered by a Business Associate Agreement.

An in-house safety score is not a clinical validation

The 99% and 93% figures are serious as a vendor’s own test, and they are also limited in ways the announcement itself makes clear. They come from OpenAI’s evaluation of selected EHR-inbox tasks, reviewed for whether harm was plausible and for how helpful physicians found the output. They have not been peer-reviewed. They are not a randomised clinical trial, a regulator’s finding, or a measure of what happens when the same agents run unsupervised on live records. A task rated “safe” in that review is not the same as a completed, signed clinical action.

The HealthAdminBench numbers should be kept in the same frame. The benchmark is independent, and comes from Stanford researchers. The 94.3% and 91.8% rubric scores, and the claim that Luna costs 98% less at 8.5 cents per task, are OpenAI’s reported results from running its models on that benchmark at maximum reasoning effort. A high share of rubric points is not a guarantee that a live prior-authorisation or appeals workflow will finish correctly, or that a health system will spend only 8.5 cents once integration, review and exceptions are included.

The same caution applies to the partner examples. CVS’s weekend backlog and “tens of thousands of consults,” Cigna’s “under eight weeks,” Abridge’s discharge-barrier plan and Ambience’s stop-and-ask demo are the organisations’ accounts of their own work. They are not, in the materials published today, independently audited production studies. OpenAI also said the public demonstration used synthetic patient data, and that HIPAA-compliant use of protected health information is available only through eligible, BAA-covered services with the required settings. Until a health system publishes its own measured outcomes, the documented event is a product opening, a vendor evaluation, and a set of partner claims, not a completed proof that computer-use agents are safe or cheap enough for every clinical desk.

Sources

Related stories

Follow @tsnmediahub on X for breaking AI news as it happens.

Follow on X

Share this story

Latest stories

  1. TensorWave and OpenAI Agree a 457MW AMD MI455X Deal
    AI Chips & Compute

More in this category

Latest stories

  1. TensorWave and OpenAI Agree a 457MW AMD MI455X Deal
    AI Chips & Compute

Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.