AI agents are capable enough to automate real work now. But they keep failing because they don't know how a specific company actually operates. Policies in PDFs. Exceptions in Slack threads. Processes in people's heads. How are teams solving this today? Is anyone beyond dumping docs into RAG pipelines?