When someone pastes a customer email into an AI tool, or a system assembles a prompt from a record, personal data travels to a model, a provider and often a log. Most of that data isn't needed for the task. Summarizing a complaint doesn't require the customer's phone number; classifying an invoice doesn't need a bank account. Data minimization is a privacy principle and also a practical design habit.
Start by mapping what flows
For each AI feature, list the data fields that enter the prompt, where the prompt goes, how long it and the response are kept, and who can read the logs. Most teams find fields included out of convenience: the whole record, not the three fields the task needs.
Three techniques, in order of preference
- Don't send it — build the prompt from only the fields required. This is the strongest control and costs nothing at runtime.
- Redact — detect and remove or mask identifiers (names, emails, phone numbers, account numbers) before the call. Pattern matching catches the structured ones; names and free-text details are harder, so test on real examples.
- Pseudonymize — replace identifiers with consistent placeholders, call the model, then map them back in the response. The model can reason about 'Customer A' without learning who that is, provided the mapping table stays on your side.
Redaction is never perfect. Treat it as a way to reduce exposure, and combine it with the contractual and technical controls on where the remaining data goes and how long it stays.
Know when to stop
- Sensitive categories — health, financial, children's data, and anything where a leak causes serious harm may call for processing inside your own environment or not using a model at all.
- Regulatory duties — consent, purpose limitation, retention and breach rules apply to model inputs as they do elsewhere. Check them with someone qualified for your jurisdiction.
- Logs — prompt and response logs are data stores. Apply retention limits and access controls to them like any other.
Make it testable
Include prompts containing personal data in your test set and check what is sent, stored and returned. Add a check to the release process that flags new fields entering a prompt. A short review at design time is far cheaper than a disclosure after launch.