Five Minutes on AI Agents in Healthcare: Senate Draft Would Test Medicare Payment for Clinical AI Without Review of Each Output

This week's news about AI agents in care was about people. It was about the patient who links a health record to an AI assistant, the nurse who gets a patient's clinical question from an agent, and whoever answers for an agent's mistake when no clinician checked its work. Here are three stories from the past week and one view that pushes back on how we read them.

1. A Senate draft would test Medicare payment for clinical AI

Sen. Bill Cassidy, chair of the Senate HELP Committee, released a discussion draft on Oct. 29, 2026. A discussion draft is a version shared for public comment. It is not law.¹ The draft would have the CMS Innovation Center evaluate a five-year payment model for AI that can perform clinical functions without a clinician reviewing each output. Comments are due Dec. 1, 2026. DistilINFO notes that the draft does not use the term AI agents.¹

Medical Daily reports that the draft, called the PATIENT AI Act, has not been introduced. It would give patients informed notice and preserve their Medicare appeal rights. It leaves the meaning of "autonomously," the payment rates, and liability undefined or unsettled.²

What it means for you: We take no position on the draft. But the questions it leaves open are ones your organization can answer now for any agent that acts without a clinician checking each output: what counts as acting on its own, and who is responsible when it is wrong.

2. Patients are bringing AI to their own records

athenahealth released a connector that lets U.S. patients link their official health record to Anthropic's Claude (Pro and Max subscribers) and get conversational explanations of their history. Patients choose which categories of data Claude can see, and they can revoke access at any time.³

GovInfoSecurity reports that the records can include labs, medications, and diagnoses. athenahealth says HIPAA, the federal health privacy rules, does not apply to this patient-directed exchange. Attorneys quoted in the story say that answer could change if a provider or a business associate directs the arrangement. A business associate is a company that handles health data on a provider's behalf.⁴

There is a wider view, too. In an opinion piece, a Publicis Health executive argues that AI agents now sit between patients, doctors, payers, and drug brands. The piece cites an AI-written prescription that Nolla Health issued in Utah under the state's regulatory sandbox, and AI refill renewals by Doctronic across about 190 medications.⁵

What it means for you: Some patients can now bring an AI assistant to their own record. Before your organization suggests or sets up these links for patients, ask your counsel where your role begins. The attorneys GovInfoSecurity quoted say that role could change the answer.⁴

3. Agents are taking on patient outreach and intake

Luma Health says its Agent Launchpad lets health systems run agents that call or text patients about preparing for colonoscopy, knee replacement, cataract, and other procedures. The agents send clinical questions to a nurse. The company reports a 95% average response rate and more than 20 minutes of coordination saved per patient at UAMS and Mile Bluff Medical Center.⁶

Assort Health says its agents, which already handle scheduling, referrals, and outreach, will now pre-fill intake forms and take payments. It reports more than 200 healthcare organizations as users.⁷

OhMD says its new product builds agents from a stated goal, such as closing care gaps in annual wellness visits. A staff member takes over when a conversation needs a person. Beta access opens in mid-November 2026 for its existing network.⁸

What it means for you: These figures are reported by the companies themselves. Two of the products describe a point where a person takes over: a nurse for Luma Health's clinical questions and a staff member for OhMD.⁶⁸ Before you adopt any agent that reaches patients, find out who staffs that point and how much time they have to act.

The strongest view against ours

The strongest counterargument is that these safeguards already answer the accountability question. Patients control their data, a nurse gets the clinical questions, and staff step in when needed. That view deserves weight, because these are real design choices.

The American Medical Association adds a caution. It says physician oversight cannot replace adequate evidence, sound system design, or conditions that make clinical review meaningful. It proposes matching liability to each party's ability to prevent or reduce harm. That covers developers, the healthcare organizations that deploy AI, and physicians, for the clinical judgment they can meaningfully exercise. Under the proposal, patients keep paths to remedy.⁹

Our perspective so far, offered as a working hypothesis: a hand-off to a person protects patients only if that person has the time, the information, and the authority to act. We are watching for independent evidence on how these hand-offs work in practice. Results showing that the people receiving hand-offs reliably catch problems would change our view.

One question for this week

If one of your agents made a mistake with a patient tomorrow, could you name the person who was in a position to prevent it?

This is not medical, legal, financial, or investment advice.