Agents can read a PDF in a demo. Feeding thousands of heterogeneous documents into real decisions is a different problem, and it's where automation breaks silently. The first part of this talk covers what reliable document processing actually requires: extraction that survives the documents that break generic tools. Then we put the thesis to the test, together with Iberia's tech team: how one of Europe's largest airlines automated document processing that defeat traditional systems, including 50+page docs with tables spanning every page, and what it took to trust that automation in operations. Because an agent that can't prove what it read isn't autonomous. It's risk with an API.