A computer workstation on a desk in an empty clinic office

An agent clicks through the clinic software now. The review step is yours to set

OpenAI’s launch post of November 16, 2026 puts a number first: reviewers rated 99% of 325 EHR inbox tasks safe. The product behind the number is a computer-use agent. It reads a screen, clicks, types and moves between applications, doing the multi-step work that lives between an electronic health record and an insurer portal. The capability is available through the OpenAI API, ChatGPT for Healthcare and ChatGPT for Clinicians. The figure is the vendor’s own, and it comes with a condition the buyer has to read before anything else in this piece matters.

The direct answer: The agents handle authorization requests, referrals, discharge planning and outreach between visits, inside the software a clinic already runs. The safety figure rests on a defined test: whether inaccuracies, omissions or actions in the model’s work could plausibly lead to patient harm. A Canadian clinic can use the shape of this launch today, and it has to rebuild the legal frame first, because the launch runs on US rules: HIPAA settings and a Business Associate Agreement. Ontario work answers to PHIPA, and work in most other provinces answers to PIPEDA and its provincial counterparts.

  • What the launch actually ships
  • What the 99% figure does and does not establish
  • The Canadian frame and the first test

What the launch actually ships

The agent works across systems that were never built to connect. A referral starts in the record, moves to a portal, and ends in a confirmation a person has to chase. OpenAI is working with Abridge, Ambience, Cigna Healthcare and CVS Health on that class of work. CVS Health is the production case in the launch record: document sending and referral steps running at volume, with other partners at earlier stages. ChatGPT for Healthcare also added more than 20 plugins in the same launch, which widens the set of systems the agent can reach from 1 screen.

Advertisement

The pattern matches a launch this site covered in a different building. Claude for Manufacturing drafts the quote. Your estimator still sets the price described a tool that does the drafting while a named person keeps the decision. The clinic version has higher stakes and the same shape: the agent moves paper, and a clinician or an authorization specialist reviews the step that commits the clinic.

What the 99% figure does and does not establish

The evaluation deserves a careful read. Physicians shaped the inbox tasks. Reviewers then judged the model’s finished work against the harm standard above, and rated 99% of 325 tasks safe. In a second read, 93% of 646 assessments rated the work very helpful or extremely helpful. On HealthAdminBench, an administrative benchmark covering work such as prior authorizations, GPT-6 Astra earned 94.3% of available rubric points. GPT-6 Luna earned 91.8% at 8.5 cents a task, which is 12 cents CAD at the Bank of Canada rate for November 16, 2026.

Those are screening results on defined tasks, reported by the vendor. They are not measured patient outcomes, and they are not a regulator’s clearance. The 1% of tasks outside the safe rating is the figure a buyer should ask about in writing: which tasks, which failure shape, and which review step would have caught each one. A vendor that can answer that question precisely has run its evaluation like an audit. A vendor that waves at the 99% has run it like an advertisement.

Access control is the other half of the safety story. The launch lets an organization set which applications an agent can open and which steps require human review, with audit logs for the deployment. That setting work is the buyer’s job, and it starts before the first task runs. Give your AI agent an 8-line job description before access gives the short form: named tasks, named systems, named reviewer.

The Canadian frame and the first test

A Canadian clinic reads this launch through different law. The US frame is HIPAA, with a Business Associate Agreement covering the vendor’s handling of health information. Ontario clinics answer to PHIPA, which sets its own rules for personal health information and for the agents and service providers who handle it. Clinics in other provinces work under PIPEDA or a provincial law of similar shape. The questions travel well even where the statutes differ: where does the data sit, who can see the audit log, how long is a record kept, and which named person answers when an agent files the wrong referral.

The free tier underlines the border. ChatGPT for Clinicians is free for eligible verified US clinicians. A Canadian clinician starts at the paid and contractual end of this launch, with a privacy review in front of any pilot. That is a slower start, and it is the correct one for software that touches a chart.

Production discipline from a neighbouring field helps here. The AI on-call agent investigates in minutes. Nobody lets it fix production alone found the same settlement in engineering: the agent investigates and drafts, and a person approves the step that changes a live system. A clinic can hold the same line. And the sign-off habit in Review AI-generated work with one signed line before your name goes on it fits the referral queue exactly: 1 named reviewer, 1 visible approval, no silent submissions.

The recap is plain. The launch is real, the figures are the vendor’s, and the review settings are the product a Canadian buyer is actually purchasing.

Run this test before a pilot: take 20 closed referrals from last month, have the agent rebuild each packet in a test setting, and have your own reviewer score every packet against the filed original. Ask the vendor, in writing, which steps the agent may complete alone and which steps wait for your reviewer. This lane ran no hands-on test of the healthcare agents, so that question, answered in writing, is where a Canadian clinic should start.

Total
0
Shares
Prev
Your first always-on agent looks great on day 9. Judge it on day 90
A manager reviewing a paper checklist beside a laptop on an office desk

Your first always-on agent looks great on day 9. Judge it on day 90

An always-on agent demos well because a demo is short

Next
Claude moved to AMD chips in 3 days. The footnote is the story
Rows of server racks with small status lights inside a data centre

Claude moved to AMD chips in 3 days. The footnote is the story

A chip switch used to cost Anthropic dozens of engineers for 6 to 12 months

You May Also Like