Seriously, It’s AI Seriously, It’s AIAI News • Tools • Trends Subscribe
Latest
AI News

ChatGPT for Healthcare Gets Agents That Do the Paperwork — 99% Safe, and One Honest Catch

November 17, 2026 By Seriously, It’s AI 0 Comments


How I checked this

I read OpenAI's November 16 announcement, "Advancing computer-use agents for healthcare teams", in full, plus the companion ChatGPT for Healthcare update and the Becker's Hospital Review report. Every number below comes from OpenAI's own evaluation — I have not tested these agents in a hospital myself, and no independent lab has published a rerun.

What the agents actually do

These are computer-use agents, not chatbots with a medical mode. They read the screen, click, type and move between the applications a hospital team already uses — the electronic health record, the insurer's portal, the scheduling tool — and carry out multi-step work across them. OpenAI's example is an MRI request: a patient messages about lower back pain, the agent reads the visit notes, physio records and prior imaging in the EHR, drafts the MRI order, an encounter note and a reply to the patient for the clinician to review, then prepares the prior authorisation in the insurer's portal for a specialist to review. The demo used synthetic patient data. That is the shape of the work: referrals, prior authorisations, discharge planning, outreach between visits — the admin that eats clinicians' days.

The same agent idea is spreading fast. Perplexity's Teammates own ongoing work on your behalf, and ChatGPT Wallet lets an agent pay without seeing your card. The pattern is the same: agents that do things, not just answer questions.

The numbers OpenAI is proud of

OpenAI tested GPT-6 Astra on the tasks clinicians handle in their EHR inboxes — multi-step work shaped by physician input. Reviewers rated 99% of 325 tasks safe, meaning errors were unlikely to plausibly harm patients, and "very helpful" or "extremely helpful" in 93% of 646 assessments.

On the admin side, OpenAI reports HealthAdminBench — an independent benchmark from Stanford researchers — at 94.3% of rubric points for GPT-6 Astra, and 91.8% for the smaller GPT-6 Luna, which costs 8.5 cents per task, about 98% less. That is the number that matters for hospitals: a cheaper model keeping about 97% of the flagship's score on routine admin work.

Alongside the agents, ChatGPT for Healthcare gets better clinical search with citations and more than 20 new plugins — scheduling, eligibility, benefits, records — from companies including Redox, Stedi, Turquoise Health, UpToDate Expert AI, Luma Health and Metriport.

What hospitals are already doing with it

This is not a roadmap slide. CVS Health is running the agents in production: one cleared a months-long document-sending backlog over a weekend, and another handles the system steps for tens of thousands of consults a day. CVS is open-sourcing its Agent Kit, built with OpenAI's Agents SDK. Cigna Healthcare's case managers used to navigate more than 10 separate clinical and admin systems to gather context on a member; computer use now pulls that together in seconds. Ambience built the guardrails: when the record is ambiguous, its agent stops, asks the clinician, and waits for approval before submitting — and health systems get a step-by-step log for audit. Abridge is building agents that resolve discharge barriers, working on medication access and home-health coordination at the same time.

Why this matters

Healthcare's real bottleneck was never medical knowledge — it was the retyping. Systems that do not talk to each other, portals that need the same information entered three times, authorisations that wait days for a human to move a file. An agent that works inside the existing software, instead of demanding a new system, attacks the actual bottleneck. And the economics finally work: at 8.5 cents a task, the cost argument for automating admin is hard to fight. This is the same week healthcare AI went from reading DNA — DeepMind's AlphaGenome 2 finds nearly twice as many disease-linked variants — to doing the paperwork around it.

The catches

First, the safety figures are OpenAI's own. 99% of 325 tasks rated safe is a strong result, but it is a physician review of the model's work, not measured patient outcomes — and 1% of tasks were not rated safe. In healthcare, 1% is not a rounding error. Second, the MRI demo used synthetic data, and the partners named are large, well-resourced organisations; a small clinic will not have CVS's validation checks or Ambience's guardrails. Third, access comes with strings: HIPAA-compliant workflows need a Business Associate Agreement and the right security and retention settings, and workspace admins must enable the plugins and the computer-use features. My verdict: if you run a health system, pilot this on one workflow — prior auth is the obvious one — with human approval gates on and the audit log reviewed. The tech is real and the price is low, but the 1% is why the doctor stays in the loop. Skip it if you cannot staff the review step.

Would you let an AI agent handle your prior authorisation, with a human checking its work? Tell me in the comments.

Sources

OpenAI, "Advancing computer-use agents for healthcare teams", November 16, 2026 — https://openai.com/index/advancing-computer-use-for-healthcare-teams
OpenAI, "More ways to move care forward with ChatGPT for Healthcare", November 16, 2026.
Becker's Hospital Review, "OpenAI builds healthcare AI agents with CVS, Cigna, Abridge, Ambience", November 16, 2026 — https://www.beckershospitalreview.com/healthcare-information-technology/ai/openai-builds-healthcare-ai-agents-with-cvs-cigna-abridge-ambience/

Comments

Browse by Category

Stay Ahead with Seriously, It’s AI

Get the latest AI news, trends and expert insights — explained simply, every day.

Follow Us →

Stay updated with AI

One short email a week. No spam, unsubscribe anytime.