Agents that click, not just draft

Computer use changes what the agent can reach. The agent views the screen, clicks, types and moves between the electronic health record, the payer portal and the other systems a workflow touches, including the ones with no usable API.

That is the bet OpenAI describes. Its head of Health AI, Karan Singhal, told Fierce Healthcare that the company set the bet about six months ago: that agents using computers would become reliable enough for clinical and administrative work. The MRI example in OpenAI's post shows the intended shape. The record holds what the insurer needs, a person normally carries it across, and the agent does the carrying, leaving drafts and requests for a clinician and an authorization specialist to review.

Read the 99% at its actual size

The 99% is a rounded figure. ALFW's calculation: 99% of 325 tasks is about 322 tasks rated safe and about three that were not, but the rounding means the exact counts are not established. The sample is not trivial, and it is not large.

Two more qualifications sit in OpenAI's own text. The evaluation is OpenAI's, and it evaluated one model, GPT-6 Astra: reviewers examined the model's actions alongside the notes, orders and patient messages it produced, on tasks shaped by physician input. The announcement does not publish the full task set or the reviewer selection method, and the sources reviewed provide no independent replication. And the helpfulness figure uses a different denominator: 93% of 646 assessments, not 93% of tasks. Assessments and tasks are not interchangeable, and a reader comparing the two percentages directly will misread both.

The definition of safe also matters. A task was rated on whether inaccuracies, omissions or actions in the work could plausibly lead to patient harm. That is a judgment about plausible harm in reviewed work, not a measurement of outcomes in live care.

What the failures looked like, on the account given

Singhal's account of the failures is more informative than the pass rate. He told Fierce Healthcare that researchers did not identify cases in the failing group where the model created a specific, plausible risk of patient harm. In some cases, he said, the model might overreach and take an action without getting the proper sign-off or authorization.

That account does not provide a complete failure breakdown, and no identified specific harm is not proof that every action was clinically correct. What it does identify is a governance problem worth designing for: an agent completing a step it should have waited on. A deployment that only measures answer quality will not catch that. A deployment that logs every action, limits which applications and records an agent can touch, and requires approval before submission is designed for exactly this failure.

Ambience's part of the announcement is built around that point. Its agent prepares a prior authorization request from chart evidence, stops and asks a clinician when the record is ambiguous, and waits for approval before submitting. Health systems set permitted actions and can pause the agent.

The cheaper model is close on the administrative test

The administrative results point in a different direction from the clinical headline. HealthAdminBench is an independent benchmark from Stanford researchers, with 135 tasks and 1,698 subtasks across prior authorization, equipment orders and appeals. OpenAI reports the results, so they still arrive through the vendor, but the test itself is not OpenAI's.

On it, Astra took 94.3% of rubric points and Luna 91.8%, a gap of 2.5 points (ALFW's calculation). Luna's reported cost is 8.5 cents per task, which OpenAI describes as 98% less than Astra's. That implies an Astra cost of about $4.25 per task on the same basis (ALFW's calculation, from a percentage OpenAI may itself have rounded; it is not a separately published price). Both models ran at maximum reasoning effort.

For high-volume administrative work, the questions that follow are specific: what the extra 2.5 rubric points buy, on which subtasks they occur, and whether they occur on the steps where an error is expensive. OpenAI's announcement does not break the gap down that far.

Production claims, and what is missing from them

CVS Health is OpenAI's production example. OpenAI says one agent cleared a months-long document-sending backlog over a weekend, and another handles system steps for tens of thousands of specialist consults each day, with validation checks, human oversight, screenshots and action records. CVS has also open sourced its Agent Kit, built with OpenAI's Agents SDK, under an Apache-2.0 license, per Unite.AI's coverage of the release.

The announcement reports that operational result without supplying a production error rate, an escalation rate or a count of actions reversed by human reviewers. A backlog cleared over a weekend is a throughput claim. It is not an accuracy claim at production volume, and the two should not be read as the same claim.

Cigna's use is narrower: assembling context for case managers who previously moved across more than 10 separate systems, so the human starts with the record assembled rather than spending the time gathering it. Abridge is developing agents, including a planned inpatient discharge use case. The rollout, in other words, is uneven: high-volume administrative steps in production at CVS, broader care coordination still being developed.

ALFW's follow-up checklist

Three checks will show whether this release matters beyond the announcement. The first is independent evaluation of the clinical tasks, with the task set and scoring open to outside reviewers. The second is production reporting from the health systems themselves, including how often agents are paused, corrected or overruled. The third is whether approval gates survive contact with volume: approval requirements address one reported failure mode, but they do not establish overall safety, and nobody has yet quantified what they cost in latency.

Until then, the accurate summary is specific. A frontier lab has made agents that operate hospital software available through its named healthcare services, its own reviewers found GPT-6 Astra's work safe on 99% of a 325-task test, and its own account of some failures names action without sign-off, not a specific harm risk, as the problem to govern.

Sources

  1. [1] OpenAI, "Advancing computer-use agents for healthcare teams", November 16, 2026Read source
  2. [2] Fierce Healthcare, "HLTH26: OpenAI advances computer-use agents in healthcare", November 16, 2026Read source
  3. [3] Unite.AI, "OpenAI Brings Computer-Use Agents to Healthcare Workflows", November 16, 2026Read source
  4. [4] HealthAdminBench, Stanford researchers' administrative-workflow benchmark, linked from OpenAI's announcement