OpenAI's ChatGPT for Healthcare can now run agents that click and type through your EHR, payer portals and fax queues. Before anyone turns them on, a health system or payer has to decide which applications an agent may reach, which steps a person must approve, and which of the new plugins your BAA actually covers. OpenAI's own score on a Stanford benchmark looks strong, but the published number measures partial credit, and OpenAI disclosed no error rates and no price.
The release landed on November 16. According to AlphaSignal's write-up of the announcement, it adds computer-use agents in ChatGPT Work and Codex, a better clinical search, and more than 20 healthcare plugins. OpenAI's own posts blocked our fetch, so the details below come from trade coverage quoting them, OpenAI's published HIPAA guide, and the benchmark paper itself.
What OpenAI Shipped to Health Systems and Payers
The release puts agents that operate software the way a person does into a workspace many hospitals already license. A computer-use agent is a model that reads the screen, then moves the mouse and types, so it can work in applications that have no usable API. AlphaSignal reports that the agents are aimed at EHRs, payer portals and other software that lacks usable APIs, with workflows that include reviewing records, preparing prescriptions for clinician review, coordinating care across hospital and payer systems, and checking insurance eligibility.
Access comes through four doors:
- Enterprise workspaces: agents in ChatGPT Work and Codex, clinical search and the plugins, all subject to workspace settings and administrator controls.
- Developers: computer use through the OpenAI API.
- Individual clinicians: a free ChatGPT for Clinicians account for verified U.S. clinicians.
- Procurement: OpenAI sales or private offers on AWS Marketplace, per AlphaSignal.
StartupHub's report lists the plugins live now: Aiva, Amion, Availity, Elation Health, IMO Health, Metriport, Oracle HDI, Redox, Stedi, TriZetto, Turquoise and UpToDate Expert AI, with workspace admins required to enable them. Eligible verified clinicians get select plugins through the free accounts.
The named early deployers are CVS Health, Cigna Healthcare, Abridge and Ambience Healthcare. CVS has production agents that send documents and process referrals; StartupHub reports that one agent cleared a months-long document backlog over a weekend and that a separate workflow handles system steps for tens of thousands of consults a day. Cigna's case managers previously worked across more than 10 clinical and administrative systems to prepare member outreach. Ambience showed an agent pausing for clinician input when a patient record was ambiguous.
OpenAI launched ChatGPT for Clinicians in April as a free product for verified U.S. physicians, nurse practitioners, physician assistants and pharmacists, and Google followed with its own healthcare workspace, which we covered on November 10.
What the 94.3% HealthAdminBench Score Measures
The headline score is a share of rubric points, which is closer to partial credit than to finished work. StartupHub reports that on HealthAdminBench, GPT-6 Astra earned 94.3% of available rubric points and GPT-6 Luna earned 91.8% at 8.5 cents per task, which OpenAI says is 98% cheaper. Those scores come from OpenAI's announcement, not from the benchmark's authors.
HealthAdminBench comes from a Stanford-led paper from April 2026 with four simulated environments (an EHR, two payer portals and a fax system) and 135 tasks across prior authorization, appeals and denials, and durable medical equipment orders, broken into 1,698 evaluation points. The paper's central finding is the gap between those two levels. The best subtask score among the seven configurations tested was 82.8% (GPT-5.4 CUA), while the best end-to-end task success was only 36.3% (Claude Opus 4.6 CUA). A task only counts as a success when every subtask is right.
Stanford's Nigam Shah described that gap to Stanford Medicine as "Almost a 50% drop." The same article lists where agents broke:
- Downloading a file from one system and attaching it in another.
- Failing to stop. In one equipment order, the chart held a face-to-face evaluation more than six months old, which invalidates the order; many agents kept going and tried to submit anyway.
- Losing track of facts needed many steps later.
The researchers also simplified logins, CAPTCHAs and session timeouts, so real portals are harder than the benchmark.
Put those facts together and a 94.3% rubric score is compatible with a much lower rate of claims or prior-auth packets that go out fully correct. Neither trade write-up reports an end-to-end task success rate for GPT-6 Astra, and AlphaSignal says OpenAI did not disclose workflow-level error rates or intervention rates. The steel-man for OpenAI: GPT-6 Astra is newer than any model the paper tested, and if rubric points track subtask credit, 94.3% sits 11.5 points above the paper's best. That would be real progress. A rubric score still measures how much of each task the agent got right, while your denial rate depends on how often it got every step right.
Where Your BAA Stops and the Plugins Start
Your Business Associate Agreement with OpenAI does not automatically extend to every plugin an admin switches on. A BAA is the HIPAA contract that makes a vendor responsible for protecting the patient data you send it. OpenAI's HIPAA guide carves out features that "have not been evaluated by OpenAI for HIPAA compliance" and warns that apps and connectors "can involve transmitting information to third parties." It also bars customers from using the eligible services "to maintain PHI as a part of a designated record set," which matters once an agent starts writing into charts and claims systems.
The plugin vendors say the same thing from their side. Stedi's page for its ChatGPT plugin says it requires a ChatGPT Enterprise plan, that connecting is free and Stedi bills only for production eligibility checks, and that organizations "need a Business Associate Agreement" with OpenAI in most cases. So each plugin carries its own contract, its own meter and its own data path, and the admin toggle is where all three get accepted.
StartupHub reports that HIPAA workflows require eligible services under a BAA plus required security and retention settings, and that OpenAI says content shared with ChatGPT for Healthcare is not used to train models.
The free clinician tier is the part your security team will miss. ChatGPT for Clinicians offers HIPAA support through a BAA for eligible accounts, and those accounts now get select plugins. A physician on your medical staff can sign up individually, and that account sits outside the workspace where you set the plugin allowlist and review audit logs. We saw the same pattern with the ChatGPT Chrome extension acting inside logged-in tabs.
Why Payers Should Compare Agents With the 2027 API Mandate
For prior authorization with the payers covered by CMS's interoperability rule, a portal-clicking agent may be a stopgap. Under CMS-0057-F, Medicare Advantage organizations, state Medicaid and CHIP programs, Medicaid managed care plans, CHIP managed care entities and QHP issuers on the federal exchanges must implement a Prior Authorization API beginning January 1, 2027, and (QHP issuers excepted) have had to decide expedited requests within 72 hours and standard ones within seven calendar days since January 1, 2026.
A provider's agent clicking through a Medicare Advantage prior-auth portal in 2027 would be doing by screen what the mandated API is built to do. Commercial plans outside the rule keep their portals, so the case for computer use there holds. The decision is per payer, and your revenue cycle team already knows which of your top payers fall under the rule.
If you are weighing agents against the scripts you already run, our comparison of Browserbase, Playwright and computer use found a deterministic script still wins on stable portals, and the UiPath alternatives piece covers what that does to an RPA renewal.
What to Do Before the First Agent Touches Production
This Week:
- Pull the list of plugins enabled in your ChatGPT for Healthcare workspace and confirm each one is off until someone has signed for it. Write down, for each, whether your BAA with OpenAI covers it and whether the plugin vendor needs its own BAA.
- Ask your medical staff office how many clinicians have created free ChatGPT for Clinicians accounts. Put the answer in front of your privacy officer with the question of whether PHI is allowed there.
- Ask your OpenAI account team three questions in writing: the end-to-end task success rate on HealthAdminBench, the price of computer-use agents in your workspace, and which agent actions land in the audit log.
This Month:
- Write an application allowlist for agents: which systems they may open, read-only versus write, and which accounts they log in as. Give every agent its own service identity so the EHR audit trail shows the agent, not the nurse who launched it.
- Define human approval points at submission, not at the start. Stanford's failure list says agents go wrong at the attach-and-submit step and at the point where they should have stopped.
- Pick one workflow with a measurable backlog, such as referral document routing (where CVS reported its weekend result), and run it in a sandbox against your own payer mix. As we reported, fewer than half of hospitals have a sandbox for vendor AI; build one first if you lack it.
Before Contract Signature:
- Get the per-task or per-seat price in writing. The only published cost figure, 8.5 cents a task for GPT-6 Luna, is a benchmark number with no workspace price attached.
- Map your top ten payers against CMS-0057-F and keep agent work on portals that will not have a Prior Authorization API in 2027.
The Bottom Line
Revenue cycle teams have bought screen-clicking automation before, in the RPA bots that log into payer portals from billing offices. Those bots follow a recorded script; an agent reads the screen and decides, which is why it can pause on an ambiguous record the way Ambience demonstrated and also why it can keep going on a stale document the way Stanford's agents did. The number that decides whether this pays is the one Stanford measured: how often the whole packet goes out right. Get that rate from OpenAI for your workflows before you scale past one queue, and keep a person at the submit button until you have it.
Continue Reading
- Keep PHI Out of Gemini for Healthcare Until Google Confirms BAA
- OpenAI's Hosted Browser Agent Asks Once Per Site, Not Per Purchase
- Sierra Acquires Anon to Let AI Agents Log In to Websites
- Best Ambient AI Scribes Save 16 Minutes a Day, Not an Hour
- Blue Cross Blames $942M on AI Coding That Treatment Never Followed
- Hospitals Test Vendor AI. Fewer Than Half Have a Sandbox.
