How to Build a Reversible PII Redaction Layer to Protect Data Sent to LLM APIs

As AI agents increasingly process emails, tickets, and databases, sensitive personal data such as ID numbers and phone numbers routinely gets embedded in prompts and sent to third-party LLM APIs without deliberate intent. A developer reviewed multiple Vietnamese codebases and found a recurring pattern where raw ticket content containing customer PII was passed directly to API calls. To address this, they built a reversible redaction layer using Python and Microsoft Presidio, an open-source PII detection library, extended with custom recognizers for Vietnamese national ID numbers and mobile phone formats. The system replaces sensitive values with placeholder tokens before sending data to the LLM, stores the mappings in a local session vault, and restores real values only after the response is returned. The approach is framed as a compliance necessity given Vietnam's Decree 13/2023 and the Personal Data Protection Law taking effect on January 1, 2026, which impose real legal risk on uncontrolled cross-border data transfers.