AI-specific checkpoint
A model is another data destination.
If an AI feature receives customer messages, internal documents or staff input, record exactly what is sent, which provider processes it, whether it is retained or used for training, and what a human can inspect. Retrieval sources should be named. Sensitive information should not be added merely because it improves an output.
- Test a prompt that includes information the model should never receive.
- Test what happens when a source is missing, conflicting or stale.
- Provide a human route for consequential or uncertain output.
- Log enough to investigate errors without creating a second uncontrolled personal-data store.