How the AI Risk Evaluation Engine Works

A comprehensive, transparent guide to input metrics, weighting formulas, and scoring logic behind the SME AI Risk Classification Engine.

Step 1: Input Parameter Selection

When evaluating an operational AI integration, users evaluate five core risk dimensions using weighted percentile sliders (0% to 100%). Each metric isolates a distinct vulnerability channel in modern Generative AI implementations.

Step 2: Weighted Index Calculation Formula

The Composite Risk Index Score ($R_{composite}$) is computed using a weighted linear combination model:

R_{composite} = (P \times 0.25) + (H \times 0.20) + (G \times 0.20) + (C \times 0.15) + (HITL \times 0.20)

Where:

  • P (Privacy Risk): Weighted at 25% due to regulatory penalties (GDPR/CCPA) and NDA breach consequences.
  • H (Hallucination Impact): Weighted at 20% to account for operational downtime and financial inaccuracy costs.
  • G (Governance Authority): Weighted at 20% for organizational accountability and sign-off verification.
  • C (Copyright Exposure): Weighted at 15% covering intellectual property and open-source license risk.
  • HITL (Human-in-the-Loop): Weighted at 20% assessing the degree of human intervention required.

Step 3: Risk Profile Radar Mapping

The engine dynamically maps the five metric values onto a 5-axis SVG radar polygon. This visual profile instantly highlights asymmetric risks—such as a workflow with low copyright risk but critically high data privacy exposure.

Step 4: Contextual Safeguard Rules Engine

Based on metric thresholds, the system evaluates logical safeguard triggers:

  • If Privacy ≥ 60%: Triggers mandatory PII sanitization & Enterprise Zero-Retention API requirements.
  • If Hallucination ≥ 60%: Triggers Retrieval-Augmented Generation (RAG) grounding & SLA verification.
  • If HITL ≥ 60%: Triggers explicit Human Interception approval gate requirements.