AI Privacy
Working with AI on sensitive data — the redaction pattern
A redaction pattern for using AI tools on sensitive data: swap secrets for placeholders, send, restore, and know when that is not enough.
When you must use an AI tool on data that has sensitive fragments, redact before sending.
The pattern
- Identify the sensitive fragments: names, account numbers, addresses, dates of birth, case IDs.
- Substitute each with a placeholder:
CUSTOMER_A,ACCOUNT_REF,ADDRESS_1. - Track the mapping in a local file the AI never sees.
- Send the redacted version to the AI.
- Process the response.
- Substitute back using the mapping.
Example
Input: "John Doe called about account 4521-9988 asking why payment from 2026-05-12 hasn't posted."
Redacted: "CUSTOMER_A called about ACCOUNT_REF asking why payment from DATE_1 hasn't posted."
Mapping (local file only): {"CUSTOMER_A": "John Doe", "ACCOUNT_REF": "4521-9988", "DATE_1": "2026-05-12"}
When this is enough vs. not enough
Enough: internal-confidential data that is not regulated.
Not enough: PHI, FedRAMP-scoped CUI, ITAR — those require approved enterprise tools, not redaction.
Full guidance in the [LearnTrainAI Week 2 module](https://learntrainai.com).