What “Muse for Healthcare” Is Missing

Today, Bessemer investor Morgan Cheatham posted a short list titled “some ideas for ‘muse for healthcare’”: a morning digest that pre-charts the day’s patients, ambient monitoring of labs and imaging with alerts on actionable findings, in-basket prioritization and triage, and agents that call other clinical sites on your behalf. Within the hour, builders were piling into the replies claiming pieces of it. The ideas are good. They are also the easy part.

The hardest reply in the thread came from a skeptic: “What if it made a mistake and who would be accountable for such a system, healthcare works on trust and no lab out there has proven that they are worthy enough to be recommended to people when it comes to healthcare.” That is the entire problem in two sentences, and none of the four ideas answers it.

Healthcare does not run on accuracy. It runs on accountability. When a person gets it wrong, there is a name attached: a license, a chart note, a malpractice carrier, a human being who can be asked what they were thinking. When an agent gets it wrong, there is a log file. Until that asymmetry is resolved, every one of these agents is a demo, not a deployment.

So the missing layer is not a better model. It is an accountability architecture, something that makes four promises: every AI-generated output is labeled as such; every clinical action is traceable to a licensed clinician who reviewed it; the full reasoning trail is preserved in a form a compliance officer can actually read; and the deployment sits inside the same legal frame as everything else in the clinic, BAA and all. One replier put it more bluntly: “A BAA and this thing wins.”

That is the thesis behind Antiphon. Not a model, a supervised response layer: clinician review before clinical actions, AI labeling on every output, cited sources attached to clinical content, and an audit trail built for the people who actually carry liability. The technology to draft the pre-chart digest already exists. What does not exist yet, at production quality, is the governance that lets a physician sign their name under it.

The winners of “muse for healthcare” will not be the teams with the best models. They will be the teams whose systems a clinician would stand behind, literally, with name and license on the output. Build the agent. Then build the accountability. In that order, the second one is the product.

Responding to Morgan Cheatham, “some ideas for ‘muse for healthcare’” (September 28, 2026).