OpenAI’s Computer-Use Agents Are Now Working Inside Hospitals — Clicking Through EHRs and Insurer Portals

Hospital paperwork has always been the quiet tax on healthcare. Clinicians and administrators spend hours copying information between systems that were never designed to talk to each other — the electronic health record on one screen, the insurer’s portal on another, and a human being in the middle doing the retyping. On November 16, 2026, OpenAI announced at the HLTH26 conference that its computer-use agents are now built for exactly that gap: AI that reads screens, clicks, types and moves between the applications hospitals already run.

The launch pairs two product updates — computer-use agents for healthcare teams, powered by the new GPT-6 Astra model and its cheaper sibling GPT-6 Luna, and a ChatGPT for Healthcare update with stronger clinical search and more than 20 new healthcare plugins. The agents are available now through the OpenAI API, ChatGPT for Healthcare and ChatGPT for Clinicians, and in ChatGPT Work and Codex subject to workspace settings.

What OpenAI announced at HLTH26

A computer-use agent is different from a chatbot. Instead of answering a question in a chat window, it operates software the way a person would: viewing the screen, moving the cursor, clicking buttons and typing into fields. In a hospital, that means an agent can open a patient’s chart, pull the relevant history, and then switch to an insurer’s portal to prepare the paperwork — without anyone building a custom integration between the two systems.

That matters because healthcare software is famously fragmented. As Fierce Healthcare reported, OpenAI’s head of Health AI, Karan Singhal, framed computer use as the way past a long-standing barrier: OpenAI already works with health systems through ChatGPT for Healthcare and its Epic integration, which arrived on September 1, 2026, but agents could only be so useful while they could not actually operate the software care teams use every day. Computer use, in his words, is “a way of resolving a lot of these barriers that exist between AI capabilities and healthcare software so that care teams can deliver better care, more effectively and more efficiently.”

The announcement is part of a much bigger week for agentic AI. At Microsoft Ignite 2026, which opened the same day, Microsoft is pushing agentic operations across Azure and Copilot — a sign that every major platform is now racing to put agents to work inside real enterprise software, not just in demos.

How a computer-use agent works inside a hospital

OpenAI’s demonstration, built on synthetic patient data rather than a real case, walks through the pattern clearly. A patient messages their clinician about lower back pain. The agent reviews the visit notes, physical therapy records and prior imaging in the electronic health record, then drafts an MRI order, an encounter note and a reply to the patient — all for the clinician to review. Then it uses the same records to prepare the prior authorization request in the insurer’s portal, for an authorization specialist to check.

Hands typing at a keyboard beside a monitor showing an abstract electronic health record dashboard

Nothing in that flow is autonomous in the way people fear. The clinician and the authorization specialist stay in control of their workflows; the agent does the gathering, drafting and cross-system legwork, and a human signs off. The target work is deliberately administrative rather than diagnostic: referrals, prior authorizations, discharge planning, specialist consults and outreach between visits. That focus is what makes the launch plausible as a near-term product rather than a research promise.

It also reflects how fast the underlying models are moving. The same week, Google’s Gemini 4 Hydrogen showed how far agentic coding models have come, and NVIDIA’s Nemotron for Science demonstrated open models built for real laboratory workflows — the whole field is shifting from models that answer questions to models that do multi-step work inside real software.

The safety numbers: 99% of 325 EHR tasks rated clinically safe

OpenAI’s headline figure comes from its own evaluation of GPT-6 Astra on the multi-step tasks clinicians handle in their EHR inboxes — work shaped with physician input, testing clinical reasoning over a patient’s longitudinal history and navigation of complex interfaces. Reviewers examined the model’s actions alongside the notes, orders and patient messages it produced, judging whether inaccuracies, omissions or actions could plausibly lead to patient harm.

The results: 99% of 325 reviewed tasks were rated clinically safe, and physicians rated the model’s work “very helpful” or “extremely helpful” in 93% of 646 assessments. HIT Consultant’s breakdown of the launch notes the same figures, and adds that the agents also arrive with improved trusted clinical search — how ChatGPT finds, evaluates and cites medical evidence — inside ChatGPT for Healthcare.

The 1% deserves as much attention as the 99%. Singhal told Fierce Healthcare that in the tasks not rated safe, researchers did not identify any case where the model created a specific, plausible risk of patient harm — but in some cases “the model might overreach and might take an action without getting the proper sign-off or authorization.” That is exactly the failure mode the human-review controls are designed to catch, and it is why OpenAI is careful to call these research results rather than proof. “The models are not perfect today,” Singhal said, “and what we’re sharing is an update on research results where we’re seeing this critical point where the models seem to be useful in practice.”

The admin benchmark: HealthAdminBench and GPT-6 Luna’s 8.5 cents per task

The clinical evaluation was paired with a separate test of the administrative side. HealthAdminBench, an independent benchmark from Stanford researchers, covers prior authorizations, appeals and medical-equipment orders. On it, GPT-6 Astra earned 94.3% of available rubric points. GPT-6 Luna, the cheaper sibling built for high-volume administrative work, earned 91.8% while costing about 8.5 cents per task — roughly 98% less than Astra — both at maximum reasoning effort.

That cost figure is the one hospital finance teams will notice. Administrative work is high-volume by definition: thousands of authorizations, referrals and orders every day. A model that is nearly as capable at a fraction of the cost changes which workflows are worth automating. ExplainX’s deep dive into the launch points out that the two models are positioned as a pair for exactly this reason — Astra for the hardest clinical-inbox work, Luna for the repetitive administrative volume.

The economics echo a broader trend across the AI industry. Tencent’s Hy4 Flash recently showed how cheap capable models have become for everyday work, and Alibaba’s Qwen4-Max is competing on the same price-performance frontier. In healthcare, though, the calculation is different: the cost of an error is measured in patient safety, not just dollars, which is why the safety evaluation matters more than the price tag.

Who is already using it: CVS Health, Cigna, Abridge and Ambience

The launch is not purely theoretical. CVS Health is using the computer-use capabilities in production for repetitive work across clinical systems — sending documents, processing referrals and supporting specialist consults. In one production workflow, an agent cleared a months-long document-sending backlog over a single weekend. Another handles the system steps behind tens of thousands of specialist consults a day, with validation checks and human oversight, and CVS has published the work as an open Agent Kit.

Cigna Healthcare is using OpenAI’s models to pull together member information from more than 10 clinical and administrative systems — benefits, providers, appointments and previous interactions — so case managers spend less time searching across systems and more time supporting members. The deployment moved from concept to production in under eight weeks, according to OpenAI’s account.

Abridge is building computer-use agents to help clinicians carry out care plans across the applications they use every day, including those without APIs, and plans an inpatient use case that adds EHR information and nurse and clinician context to flag what needs attention. Ambience is working with OpenAI on similar workflows. Smaller companies including Assort, Bunkerhill, Forus, Gravity and Tennr are testing agents that interact directly with EHRs and payer portals for referral management, authorization checks and medication approvals.

More than 20 new plugins and better clinical search

Alongside the agents, OpenAI expanded the ChatGPT for Healthcare plugin ecosystem with more than 20 new healthcare plugins, available inside ChatGPT for Healthcare and directly within EHR workflows through ChartGPT, which brings ChatGPT into the record itself. The lineup includes Amion, Metriport, Stedi, UpToDate Expert AI, Elation Health, K Health, Luma Health, Notable, Redox, TriZetto and Turquoise, spanning clinical care, care coordination and business operations.

The improved trusted clinical search is the quieter half of the update, but it may matter just as much. When a clinician asks ChatGPT for evidence, the quality of the answer depends on how it finds, evaluates and cites medical literature. OpenAI says the update improves all three, with inspectable citations — a direct response to the trust problem that has limited clinical AI adoption.

The guardrails — and what the numbers don’t prove

OpenAI has built the healthcare agents around explicit controls. Organizations can set which applications agents can access and establish clear controls for which steps require human review. ChatGPT for Healthcare provides role-based access controls and audit logs, so every agent action and approval leaves a record. For workflows involving protected health information, organizations must use eligible services covered by their Business Associate Agreement with OpenAI, apply the required security and retention settings, and extend protections to connected systems. OpenAI says content shared with ChatGPT for Healthcare is not used to train its models.

Doctor and nurse reviewing AI-drafted notes on a tablet in a bright hospital office

Even so, the honest caveats matter. The 99% safety rating comes from physician review of 325 tasks, and the HealthAdminBench scores are evaluations — not measured patient outcomes from live clinical care. The numbers are OpenAI-reported, and the benchmark, while independently developed by Stanford researchers, was run on OpenAI’s account of its own models. OpenAI’s product posts do not assign legal liability if an agent submits a wrong referral or order, and as of Tuesday morning there had been no public response to the launch from the FDA or the Department of Health and Human Services. The FDA’s January 2026 Clinical Decision Support Software guidance explains which functions the agency considers excluded from the device definition; OpenAI has not addressed the agents’ device status.

That context does not diminish the launch — it defines how to read it. A 99% safety rating on reviewed tasks is a strong signal that the technology has crossed from interesting to usable, with a human in the loop. It is not a guarantee, and no hospital should treat it as one.

What this means for healthcare AI

The most important shift here is not a benchmark score. It is the move from copilots that draft and summarize to agents that do the work — reading one system, typing into another, and handing a finished draft to a human for review. Healthcare has resisted that shift longer than most industries because the stakes are higher and the software is older. Computer use sidesteps the integration problem entirely: if a human can click through it, an agent can too.

That also raises the bar for everyone else. Claude Mythos ported itself to AMD GPUs in three days earlier this week, showing how quickly agents can now operate real systems. OpenAI is betting the same capability, pointed at hospital administration, can quietly remove one of the biggest sources of burnout in medicine: the paperwork that keeps clinicians away from patients.

If CVS Health’s production deployments hold up — a backlog cleared in a weekend, tens of thousands of consults a day handled with human oversight — the question for other health systems will not be whether to try computer-use agents, but how quickly they can put the guardrails in place to do it safely. The technology has arrived in the hospital. The humans are still in charge — and that, more than any benchmark, is the point.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top