How / What / When

How / Answer

How to protect private data in an AI application

Protect private data by sending less of it, limiting who and which models can access it, isolating tenants, logging access, defining retention, and testing the system for accidental disclosure.

For product teams handling customer, employee, or confidential business information. · Reviewed 2026-09-29

Start with data minimization

Do not send an entire customer record when the model needs one field. Remove identifiers, secrets, and unrelated history before the request. Smaller context is also cheaper and easier to inspect.

Controls that belong in the application

  • Separate tenants and projects at the data layer.
  • Use server-side credentials, never client-exposed admin keys.
  • Allow only approved models and tools for sensitive workflows.
  • Redact credentials and unnecessary personal information.
  • Log who accessed what without copying sensitive content into logs.
  • Set deletion and retention rules for prompts, outputs, backups, and vendor systems.

Test the privacy boundary

Try cross-tenant questions, prompt injection inside documents, requests for deleted records, and attempts to reveal system instructions. A privacy policy is necessary, but it is not a test that the implementation actually enforces the policy.

Keep learning

Get one useful AI answer each week.