AI agents learn delivery behaviour from operator corrections, but cannot fix bad source data. Understand what feedback loops can and cannot generalise.