How to Govern AI End-to-End
This workflow shows how to resolve fictitious customer-service cases while keeping personal data, prompts, model access, and outputs under control. This workflow wraps a single LLM call in governance layers, from masking personal data before it reaches the model to storing only human-approved responses. In particular, the workflow shows how to:
Mask PII. Replace names, IDs, telephone numbers, and email addresses in case text before the AI sees them. The PII Masking component combines Presidio detection with custom masking rules and keeps both mappings for unmasking later.
Enforce prompt policies. Check each masked prompt against company policies for safe, ethical, compliant content free of sensitive data. Only prompts that pass move forward.
Use admin-approved models. Connect to the Knime GenAI Gateway and select from LLMs your Knime admin whitelists and manages centrally.
Scan for AI risk. Scan the used LLM using Giskard framework to evaluate the model's propensity for hallucinations, bias, fairness, security concerns, and other vulnerabilities.
Validate outputs. Apply output guardrails that enforce a consistent response format and flag forbidden words, non-compliant content, time commitments, and automated refunds. Keep only rows with an AI vulnerability score below 80.
Review and approve (by human). Inspect masked responses, approve or reject each row, and export a CSV audit trail that stays masked.
Unmask and store. Restore original PII in approved responses only and write them to Excel.