AI triage tools promise faster rare-disease variant review, but the ground truth is messy. Retrospective results look great on slides but often shrink in prospective use where case mix and pressure are different. Domain shift NICU vs adult, panel vs WES/WGS, plus CNV/SV/mito/splicing regularly breaks apparent gains. And even when the model helps, the hidden operational load (integration, SOP updates, compliance) can eat the efficiency. That’s the gap: big promises, thin proof in real workflows. What’s your experience?