Masking PII Before a Chatbot Replies: A Local Presidio Test

A developer tested how HoverBot masks personal data in visitor messages before the reply model sees them, using synthetic inputs on November 2. Presidio replaced a sample email address and a test card number with labels, while the surrounding question stayed intact. The approach runs detection in a self-hosted deployment before any external model call, but it only protects the entity types and confidence thresholds chosen. In six probes, a phone-shaped string was missed even though phone numbers were among the selected types, and names, addresses and order numbers were deliberately left unmasked. The author stresses these are too few samples to measure protection rates, and that masking a request does not erase data stored in the application's own records.