When an AI agent decides which emails to delete, you quickly see who really controls the process.
I ran a simple test on my archived mailbox, over 100 000 messages. The goal was modest: trim the store below a target size.
The agent started correctly, pulling the twenty oldest messages as the first batch. Once that batch vanished, the client switched to newest‑first sorting. The AI didn't notice. From then on the AI deleted the most recent emails instead of the old ones.
It wasn’t a catastrophic loss; I knew the risk and used a test environment. It reminded me that tools follow the exact rules we give them, and a slight deviation can change behavior unexpectedly.
We’re far from trusting AI with mission‑critical data without rigorous validation.
What safeguards are you building before you let an algorithm clean up your most valuable information?