Classify data before choosing a tool
Separate public, internal, confidential and regulated data. The appropriate AI setup depends on the sensitivity and the consequences of disclosure, not only on whether a feature is convenient.
Minimize what you send
Remove names, identifiers and irrelevant fields when the task does not require them. Redaction and pseudonymization reduce exposure, but they are not a substitute for understanding provider terms and retention.
Check retention and training terms
Provider behavior can vary by product, plan, API and region. Verify current official documentation before promising that data is not retained or used for training.
Control connected tools
An agent with access to email, drives or business systems may expose more than the prompt itself. Apply least privilege, explicit scopes and approval before sensitive actions.
Design deletion and incident response
Know where prompts, logs, exports and local state live. A production AI system should have a way to remove data, revoke credentials and investigate unexpected disclosure.