Human approval is a design decision, not a disclaimer
2026-10-06 · 6 minute read
Putting a person in the loop only works when you decide what they see, what they can change, and which actions they must authorize.
"Human in the loop" is easy to put on a slide and hard to operate. A checkbox under a wall of extracted text is not a review. A coordinator who is measured on speed will approve whatever is in front of them. The checkpoint has to be designed or it will become theater.
Decide the boundary first
Some steps are administrative and repetitive: move a confirmed field, propose a slot that matches rules the team already uses, flag a missing page. Some steps change the clinical record, the assignment of a clinician, or a status other teams will trust. Those are different designs.
For the repetitive step, the person should see what will change and the source it came from. For the consequential step, the automation prepares and a named role authorizes. The system should not write until that authorization exists.
Design the queue
Review needs an owner, a sort order, and a definition of done. Low-confidence extractions and rule conflicts belong at the top. A clean, high-confidence proposal can be quicker to confirm, and it should still be confirmable. If everything looks equally urgent, reviewers will rubber-stamp.
Track overrides. They are not a failure of the staff. They are how you learn that a rule is missing. An approval step that nobody ever changes is worth inspecting.
Keep clinical judgment with clinicians
Workflow automation can prepare, route, check for completeness, and update a system after a person says yes. It should not pretend to decide what care a patient needs. That boundary is an implementation choice, made with the clinical and operational leaders who will live with it.
