Who decides when an AI agent decides?

, AI and governance

In the US the debate on agentic AI in healthcare has moved past “if”. One large insurer now runs tens of thousands of prior authorization decisions a day, most of them handled autonomously.

That number is impressive and uncomfortable at the same time. When an agent can act across an EMR, a claims platform and a scheduling system without a human in every step, the real design problem is no longer accuracy. It is authority: which decisions we delegate, and which stay with a person.

The FDA drew part of that line in January, pulling autonomous and heavily generative decision-support software under device regulation. Useful, but the harder line runs inside each organization, not in the guidance.

My read from Europe: the winners will not be the teams with the best model. They will be the ones who design the boundary well between what the machine does and what the clinician owns, and who keep that boundary visible to patients and payers alike.