AI Governance & Trust Overview

Data Privacy & "No-Training" Guarantees

Our policy mandates strict compliance with data privacy laws like GDPR. We achieve this through our choice of infrastructure:

• Reasoning (Google Vertex AI): We use Gemini models via the Google Vertex AI enterprise platform.
• Data Isolation: Unlike consumer AI tools, Vertex AI contractually guarantees that customer data is never used to train global models and remains isolated within our secure instance.
• Audio & Sentiment (Deepgram): Speech-to-text and sentiment analysis are handled by Deepgram’s proprietary models.
• Zero Retention: We have configured these for zero data retention, ensuring voice data is processed in real-time and not stored or used for model training.

The Control Layer (LiteLLM)

To meet our policy requirements for Security and Safe Usage, we use LiteLLM as a centralized AI Gateway.

• Security Guardrails: LiteLLM allows us to manage API keys centrally and enforce rate limits to prevent system abuse.
• Standardization: It ensures that every request follows our internal security protocols before reaching the AI models.

Continuous Audit & Transparency (Langfuse)

Our policy requires Transparency and Explainability in how AI tools work. We utilize Langfuse as our primary observability and audit tool.

• Full Audit Trail: Every AI interaction is "traced," providing a complete record of the prompt, the model's logic, and the final output.
• Quality Monitoring: We use Langfuse to debug and monitor for accuracy, ensuring we meet our commitment to identifying and mitigating bias.

Operational Guardrails & Reliability

Beyond infrastructure, we enforce strict behavioral constraints at the prompt level to ensure the AI remains a specialized tool rather than a general-purpose chatbot.

• Tool-Bound Logic: The AI is strictly limited to an authorized toolkit, such as verified Knowledge Base searches and specific skill transfers.
• Scope Enforcement: The system is explicitly instructed to decline any tasks or questions that fall outside its provided capabilities or documented parameters.
• Hallucination Mitigation: We prohibit the AI from using "external knowledge." It must rely solely on the verified information provided through its tools to ensure accuracy and prevent misinformation.