Blueprint: Retail Return Agent

Multimodal PII scrubbing for E-commerce Agents

Agents processing customer returns often receive photos of damaged goods, receipts, or even customer IDs for verification. This multimodal data contains high-density PII.

This blueprint demonstrates how to implement In-Payload PII Redaction for unstructured retail data.

Architecture

1. Master Policy Definition

{
  "name": "retail-privacy-policy",
  "priority": 7500,
  "rules": [
    {
      "id": "mask-retail-pii",
      "effect": "allow",
      "principals": ["agent:return-bot"],
      "actions": ["llm:generate"],
      "resources": ["*"]
    }
  ],
  "content_policy": {
    "enable_pii_detection": true,
    "pii_action": "mask",
    "blocked_patterns": ["\\bORD-\\d{4}-\\d{4}\\b"]
  }
}

2. Implementation

Python Prototype

from openai import OpenAI

client = OpenAI(
    api_key="ignored",
    base_url="http://cz-gateway:8001/v1",
    default_headers={"X-ControlZero-App": "returns"}
)

def process_return_request(message: str):
    # This message contains an Email and a custom Order Pattern (ORD-xxxx-xxxx)
    try:
        response = client.chat.completions.create(
            model="gpt-5.4-mini",
            messages=[{"role": "user", "content": f"Customer Request: {message}"}]
        )
        return response.choices[0].message.content
    except Exception as e:
        return f"Governance Error: {e}"

# Scenario: Message with sensitive identifiers
raw_msg = "My email is user@example.com and my order number is ORD-1234-5678."
result = process_return_request(raw_msg)

# The LLM receives:
# "Customer Request: My email is [EMAIL] and my order number is [REDACTED]."
print(f"Agent Response: {result}")

3. Validation Checklist

• Surgical Accuracy: Verify that only the PII is masked while the intent of the message remains clear for the agent.
• Custom Patterns: Confirm that the ORD- regex pattern correctly triggers the redaction.
• Audit Trail: Check the audit log store to ensure that request_payload_encrypted contains the raw data (for SRE review) while common logs show masked data.