OpenAI Tests Computer-Use Agents With CVS, Cigna, Abridge

OpenAI Tests Computer-Use Agents With CVS, Cigna, Abridge

Estimated reading time: 7 minutes · Last updated:

OpenAI is offering computer-use agents for healthcare. These AI systems operate software the way staff do, reading the screen and clicking through electronic health records and insurer portals. They take on referrals, prior authorizations and other manual chores in places where no direct integration exists. In OpenAI's own evaluation, its GPT-6 Astra model was rated clinically safe on 99% of 325 reviewed tasks. CVS Health, Cigna Healthcare, Abridge and Ambience are working with the company on specific workflows. The partnerships were announced at HLTH 2026 in Las Vegas. CVS Health already runs the agents in production. The safety numbers are encouraging, but they come from OpenAI grading its own model.

The models are not perfect today, 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.

Karan Singhal, OpenAI

Key takeaways

  • Safety score: OpenAI reported that GPT-6 Astra was judged clinically safe on 99% of 325 reviewed EHR inbox tasks. Physicians rated it very or extremely helpful in 93% of 646 assessments.
  • CVS in production: At CVS Health, one OpenAI agent cleared a document-sending backlog that had built up over months in a single weekend. Another handles system work for tens of thousands of consults daily.
  • Cigna case managers: Cigna Healthcare uses OpenAI models to gather member details from more than 10 clinical and administrative systems so case managers spend less time searching.
  • Known failure mode: Karan Singhal of OpenAI said the 1% of tasks not rated safe involved the model overreaching, for example acting without proper sign-off, rather than creating a specific risk of patient harm.

Why Healthcare Needs Agents That Click Instead of Connect

A computer-use agent is an AI system that works a software interface the way a person does. It looks at what is on the screen, moves a pointer, clicks and types. That matters in healthcare because much of the work happens in systems that were never built to talk to each other.

An MRI shows the problem. Once a clinician orders the scan, the insurer may want details held in the patient's record. Somebody on staff then has to copy those details out and key them into the payer's portal before authorization can be requested. The same pattern repeats in referrals, discharge planning and specialist consults.

Karan Singhal leads OpenAI's Health AI and Artificial General Intelligence work. He said the company still invests in APIs, plugins and MCP connectors, but legacy and fragmented systems make those links slow to build and often incomplete. He described computer use as a way past the barriers that remain even after ChatGPT for Healthcare and OpenAI's Epic integration. Around six months earlier, his team bet that agents using computers would become reliable enough for clinical and administrative work, as first reported by Fierce Healthcare.

How GPT-6 Astra Scored on Clinicians' Inbox Work

OpenAI tested GPT-6 Astra on jobs that land in a clinician's EHR inbox, and physicians helped shape those tasks. Each task required several steps. The model had to read a patient's history over time, then work through crowded interfaces to review records, draft notes and messages, prepare medication refills or propose orders. Key actions were left for a clinician to approve.

Reviewers examined both the model's actions and the documents it produced. OpenAI reported that 99% of 325 reviewed tasks were judged clinically safe. That meant reviewers found nothing wrong, missing or done in error that might realistically hurt a patient. In 93% of 646 physician assessments, the model earned a rating of 'very helpful' or 'extremely helpful'.

For the 1% not rated safe, Singhal said researchers found no case of a specific, plausible risk of harm. "In some cases, the model might overreach and might take an action without getting the proper sign-off or authorization," he said. These figures come from OpenAI evaluating its own model, not from an independent study. OpenAI also tested agents on back-office work, including appeals, prior authorizations and orders for medical equipment.

What CVS Health, Cigna and Abridge Are Building

CVS Health has put the capability into production to automate repetitive steps in its clinical systems. These include document sending, referral processing and specialist consult support. OpenAI said one agent cleared, in a single weekend, a document-sending backlog that had built up over months. Another runs the system-side work behind tens of thousands of daily consults.

Cigna Healthcare uses OpenAI's models to assemble a member's details from more than 10 clinical and administrative systems. Those systems cover benefits, providers, appointments and earlier interactions. OpenAI said the aim is to cut the time case managers spend hunting across screens, so they have more time for members.

Abridge, a generative AI company, is building agents that help clinicians execute care plans inside the applications they use every day, including ones with no API. It plans an inpatient version that adds EHR data plus nurse and clinician context to flag what needs attention and coordinate next steps. In that version, agents could work in parallel on medication access and home health arrangements, alongside calls and messages.

Ambience is also named as a partner, but no workflow details were given. Assort, Bunkerhill, Forus, Gravity and Tennr are testing agents that work directly in EHRs and payer portals. Their use cases include referral management, authorization checks and medication approvals.

Who Stays in Control When an Agent Does the Clicking

OpenAI ran a demonstration on synthetic patient data. A patient messages a clinician about lower back pain. The clinician asks the agent to go through the patient's visit notes, physical therapy records and earlier imaging. The agent then produces a draft encounter note, an MRI order and a patient reply, all waiting for the clinician's review. Next, it reuses those records to fill in a prior authorization request in the insurer's portal, which an authorization specialist checks. Singhal said both people stayed fully in charge of their own steps.

OpenAI said the agents can be set up with limits on which systems they reach and what they may do. Audit logs record each action and approval, so teams can see how a workflow was finished. That control layer is the practical answer to the overreach Singhal described.

Singhal said OpenAI sees its job as supplying the foundational capability and safeguards. Healthcare companies handle the 'last mile' of deployment. He named lower latency and lower cost as the next areas where the models should improve.

Who is applying OpenAI computer-use agents, and to what
Organization Workflow Status
CVS Health Documents, referrals, specialist consults In production; backlog cleared over a weekend
Cigna Healthcare Case manager preparation across 10+ systems Using OpenAI models
Abridge Care plans, including apps without APIs Building; inpatient use case planned
Ambience Not disclosed Named partner
Assort, Bunkerhill, Forus, Gravity, Tennr Referrals, authorization checks, medication approvals Testing

The case for and against computer-use agents in care

The case for

  • Agents work through existing screens, so they can reach systems that lack APIs without waiting for new integrations.
  • CVS Health's weekend backlog clearance shows quick gains where the bottleneck is clerical.
  • Configurable access limits and audit logs give teams a record of every agent action and approval.

The case against

  • The 99% and 93% figures come from OpenAI's evaluation of its own model, not an independent study.
  • The failure mode Singhal described, acting without proper sign-off, falls on exactly the step that keeps clinicians in control.
  • Singhal named latency and cost as areas still needing improvement.

What to be careful about

  • An agent that skips an authorization step in a clinical or payer system could commit an action no person approved.
  • Safety and helpfulness scores are self-reported by OpenAI and cover EHR inbox tasks, not every workflow partners are deploying.
  • Agents that operate through screens depend on interfaces that can change, unlike formal integrations.

The bottom line

OpenAI's argument is that healthcare does not need to fix its integration problem before AI can help. An agent that can operate the screens staff already use avoids years of API work. CVS Health's production results suggest the payoff can be immediate where the bottleneck is clerical. The safety evidence is promising but self-reported. A 99% rate on 325 tasks is a strong number, but it comes from a company grading its own model. The failures OpenAI describes involve agents acting without sign-off, which is the risk that audit logs and access limits are meant to contain. The open question is whether health systems will trust agents where a skipped approval has real consequences.

What to watch

  • Watch for Abridge's inpatient agent, which would combine EHR data with nurse and clinician context; no launch date has been set.
  • Watch for the latency and cost improvements Singhal named as the next focus for the models; OpenAI has given no timeline.
  • Watch whether OpenAI publishes detailed results for administrative tasks such as appeals and prior authorizations; no date has been given.

Frequently asked questions

How can clinicians access OpenAI's computer-use agents?

OpenAI offers computer use for healthcare through ChatGPT for Healthcare, ChatGPT for Clinicians and the OpenAI API. Care teams, developers and individual clinicians can use any of the three.

How is a computer-use agent different from an AI copilot?

OpenAI describes copilots as tools that draft or summarize information. Computer-use agents go further and operate EHRs, payer portals and other applications on the user's behalf. OpenAI presents this as the next step for AI in healthcare.

What else did OpenAI announce for healthcare at HLTH 2026?

OpenAI updated trusted clinical search in ChatGPT for Healthcare and ChatGPT for Clinicians. The changes affect how it finds, evaluates and cites medical evidence. It also added more than 20 new healthcare plugins covering clinical care, care coordination and business operations.



Share:

Categories

Newest course every month

Advertise your offline course to a wider audience with our landing page.

You May Also Like

OpenAI computer-use agents for healthcare were rated clinically safe on 99% of 325 tasks as CVS Health, Cigna and Abridge...
CHIME's 2026 Most Wired survey finds 93.2% of healthcare organizations use AI, but only 16.4% of AI users document healthcare...
A West Health-Gallup survey finds 56% of patients accept AI in healthcare decisions if a doctor can override, while only...