Physician-in-the-Loop
The conditions under which clinical AI oversight is real
What this book is
Every health system says the same sentence about its clinical AI: there is always a human in the loop. In most deployments that sentence is not true, and the people repeating it have not checked.
A clinician who confirms a recommendation they cannot evaluate, under time pressure, with the default pre-selected and the override three screens deep, is not providing oversight. They are a place for responsibility to come to rest — a signature that transfers accountability without transferring understanding.
Who it is for
CMIOs, chief quality officers, compliance officers and medical directors who have signed a governance document saying a human reviews every output.
The argument
Oversight is not a checkbox in a workflow. It is a set of conditions, and if you have not built the conditions you do not have oversight, whatever your governance document says.
The book diagnoses how human review degrades in real deployments — automation bias, the arithmetic of alert burden, review time that was never put in a schedule, the override that exists and cannot be used, the log nobody reads. It then specifies nine conditions under which a physician genuinely is in the loop, each with a test you can run this week. And it turns those conditions into something a health system can install, audit and defend: at procurement, in the workflow, in the override review, in governance, and in the record.
It also says the thing governance frameworks avoid. Sometimes the conditions cannot be met, and “we will put a clinician on it” is not a mitigation.
The frameworks it introduces
- The nine conditions
- Loop specification
- Override review
- Loop measurement set
Where it sits in the series
The governance capstone covers the operating system from committee charter to certification readiness, at the level of the organization and its artifacts. This volume works one level down, at the point where a clinician meets a machine output, and treats the oversight step itself as the object of design.
Print specification
272 pages, 6 × 9 in paperback. Spine 0.6125 in.
Published under the Healthcare AI Institute editorial standard.
Written and reviewed against the standard by a physician-executive whose career spans three national healthcare systems. Last reviewed on 2026-08-04.
Written from operational experience. No statistic appears that could not be sourced accurately; where a number would have helped, the mechanism is argued and the limit of the evidence is stated in the text. Scenarios are labelled as constructed.
The Institute accepts no vendor sponsorship, holds no vendor equity and takes no referral fees.