Skill: Evaluate AI Model Governance

Domain

technology

Description

Assesses AI/ML model deployments against enterprise governance frameworks, regulatory requirements, and responsible AI principles. Evaluates model risk, explainability, bias, data lineage, and operational controls to ensure compliant and trustworthy AI systems.

Business Rules

This skill implements comprehensive AI governance evaluation based on industry standards (NIST AI RMF, EU AI Act, SR 11-7):

•Model Risk Classification: Tier 1 (Critical), Tier 2 (High), Tier 3 (Medium), Tier 4 (Low) based on business impact and autonomy level
•Explainability Requirement: Tier 1-2 models require full explainability; black-box models prohibited for high-stakes decisions
•Bias Testing Mandate: All customer-facing models must pass demographic parity and equalized odds tests
•Data Lineage: Complete data provenance required for regulated industries (financial services, healthcare)
•Model Monitoring: Real-time drift detection required for Tier 1 models; monthly for Tier 2
•Human Override: Tier 1-2 models must have human-in-the-loop controls with <5 minute escalation SLA

Input Parameters

•model_name (string): Name/ID of the AI model
•model_type (string): "classification", "regression", "nlp", "computer_vision", "generative"
•use_case (string): Business use case description
•decision_autonomy (string): "fully_autonomous", "human_assisted", "human_final_decision"
•customer_facing (bool): Whether model outputs directly affect customers
•regulated_industry (bool): Whether deployed in regulated industry
•explainability_method (string): "none", "shap", "lime", "attention", "inherent"
•bias_testing_performed (bool): Whether bias testing was conducted
•bias_metrics (dict, optional): Results of bias testing if performed
•data_lineage_documented (bool): Whether data lineage is fully documented
•drift_monitoring_enabled (bool): Whether model drift monitoring is active
•human_override_available (bool): Whether human override mechanism exists
•last_validation_date (string): Date of last model validation (ISO format)

Output

Returns a governance assessment with:

•compliant (bool): Overall compliance status
•risk_tier (string): Model risk classification (Tier 1-4)
•governance_score (float): Composite governance score (0-100)
•violations (list): List of governance violations with severity
•required_controls (list): Controls that must be implemented
•regulatory_flags (list): Specific regulatory concerns (EU AI Act, SR 11-7, etc.)
•remediation_priority (string): "critical", "high", "medium", "low"
•next_validation_due (string): Required next validation date

Usage Example

python

from ai_governance import evaluate_model_governance

result = evaluate_model_governance(
    model_name="credit-decision-v2",
    model_type="classification",
    use_case="Automated credit approval decisions",
    decision_autonomy="fully_autonomous",
    customer_facing=True,
    regulated_industry=True,
    explainability_method="shap",
    bias_testing_performed=True,
    bias_metrics={"demographic_parity": 0.92, "equalized_odds": 0.88},
    data_lineage_documented=True,
    drift_monitoring_enabled=True,
    human_override_available=False,
    last_validation_date="2025-06-15"
)

Implementation

The governance logic is implemented in ai_governance.py and references:

•risk_tier_matrix.csv - Risk classification criteria
•regulatory_requirements.csv - Regulatory framework requirements
•control_catalog.csv - Required controls by risk tier
•bias_thresholds.csv - Acceptable bias metric thresholds

Evaluate Ai Model Governance