The single most important realization I've had working with connectors is around how comprehensive AI's "reading" of external data tends to be. What an LLM does instead is "strategically sample" and selectively read only the parts it thinks are relevant, just as a human would.
I ran into this exact trap using MCP connectors in Claude — assuming full ingestion when it had really only grazed a thin slice. Knowing it's strategic sampling, not comprehensive reading, changes what I can actually trust it to know.