From AI support copilot to agentic customer service

Short summary

A detailed engineering post about evolving an AI support copilot into a full agentic customer service system, revealing three PII leakage boundaries: ingress, return path (tool results), and emission. The author discovered that tokenizing customer input at intake was insufficient because CRM query results returned raw PII directly into model context. The solution involves tokenizing into a vault rather than redacting, plus regex-based detection and checksum confirmation for identifiers.

• •Three PII leakage boundaries exist: ingress, return path (tool results), and emission — only one was originally instrumented
• •CRM lookup results bypass intake tokenization, delivering raw PII into model context
• •Tokenize into a vault rather than redact, since redaction destroys the bridge needed to send answers back
• •GDPR erasure requirements force you to track exactly where personal data went

Generated with AI, which can make mistakes.