The Case for Physician-Supervised AI in Metabolic Medicine

The metabolic medicine conversation is increasingly dominated by two competing claims: that AI will transform patient care, and that AI cannot be trusted with patient care. Both miss the point. The real question — and the only useful one — is not whether AI should have a role, but what that role should be and who remains in command.

In metabolic medicine, the answer is not complicated. A physician brings irreducibly complex judgment to every encounter: the ability to weigh a lab value against a patient's stated experience, to notice what isn't in the chart, to decide that the evidence-based protocol doesn't apply today because of something the patient mentioned in passing. AI cannot do any of this. What AI can do — and what it can do well — is the preparation: assembling the full picture, identifying who needs attention, surfacing what changed since the last visit.

The risk in clinical AI is not that it does too little. The risk is that it is allowed to do too much — that the physician becomes a reviewer of AI outputs rather than the generator of clinical judgment. This isn't hypothetical. It is the design failure we see most often when AI is introduced into clinical workflows: the AI provides a recommendation, and the physician, over-relied on and time-constrained, checks a box rather than applying judgment.

Physician-supervised AI means something specific: the AI prepares, the physician decides. Every clinical output passes through a physician before it reaches a patient or a clinical record. The system is designed so that this cannot be bypassed — not by a user, not by a workflow shortcut, not by the AI itself. The physician is not a formality in the loop; the physician is the loop.

This is what we built. Not because it was required, but because it is correct. The value of a clinical AI that the physician can actually trust — rather than merely tolerate — compounds over time. Physicians who trust their tools use them fully. Patients whose physicians are properly supported see better outcomes. That is the case for physician-supervised AI: not an ethical constraint, but a design principle that makes the whole system work better.