The shadow AI framing is accurate, but the harder problem is that even approved AI tools create data exposure risks. When a team uses a sanctioned LLM via API and pastes in a customer record to "just get a quick summary," that's not shadow AI — it's compliant behavior that still leaks data. The perimeter isn't tool approval anymore, it's prompt-level data handling. Most orgs aren't close to governing at that granularity yet.