Logo

AI Insights Module · Governance

Bias & Fairness Monitor

AI Generated Fairness Summary

Confidence 94.1% Generated 12 min ago

AI analyzed outcome distributions across 247 production models spanning 18.4 million predictions today. Churn Prediction v3.2 shows a 6.2 point risk-score disparity between SMB and Enterprise accounts — within acceptable variance and explained by contract tenure. Fraud Detection v4.1 was flagged for review after its false-positive rate diverged 11.4 points for accounts in the APAC region versus the global baseline, with no legitimate risk factor identified yet. Recommendation: recalibrate Fraud Detection v4.1's regional threshold and re-run parity validation within 5 business days.

Business impact: $340K potential exposure if Fraud Detection v4.1 disparity goes unresolved

Fairness Score by Model

Composite parity score (0–100) across all monitored business-segment outcome checks

Churn Prediction v3.2

91

Fair

Revenue Forecasting v2.7

88

Fair

Fraud Detection v4.1

61

Flagged

Customer Segmentation v1.9

76

Needs Review

Demand Planning v2.3

93

Fair

Sentiment Analysis v1.4

85

Fair
Models scored across 247 production deployments 1 flagged · 1 needs review

Parity Metric Comparison

Higher = more consistent

Five parity dimensions scored 0–100 across three production models

Segment Parity Breakdown

Outcome metrics by business segment — Fraud Detection v4.1

Business Segment Accounts Avg Risk Score Approval Rate False Positive Rate Parity Delta
SMB 14,820 42.1 91.4% 3.8% Baseline
Mid-Market 6,340 44.6 89.9% 4.5% +0.7 pt
Enterprise 2,110 46.9 88.2% 7.2% +3.4 pt
APAC region 5,280 51.3 84.6% 15.2% +11.4 pt
28,550 accounts scored across 4 business segments

What's Driving the Fraud Detection v4.1 Disparity

The 11.4 point false-positive rate gap for APAC-region accounts is not explained by transaction volume, account age, or declared risk category — the three legitimate factors the model is permitted to weight. Cross-referencing device-fingerprint diversity and payment-method mix (both correlated with but not equivalent to region) accounts for roughly 4 of the 11.4 points. The remaining 7.4 points show no confirmed legitimate driver and are consistent with a regional data-drift issue introduced when the APAC transaction gateway was migrated eight weeks ago. Recommendation: hold the APAC threshold at the pre-migration calibration until a re-fit completes.

Root cause confidence 87% · APAC gateway migration, 8 weeks ago

Active Fairness Reviews

Fraud Detection v4.1 — APAC false-positive rate disparity (+11.4 pt)

Assigned to Marcus Lee · Opened 2 days ago

Customer Segmentation v1.9 — tenure-band recommendation skew

Assigned to Priya Patel · Opened 5 days ago

Revenue Forecasting v2.7 — industry-vertical variance under monitoring

Assigned to Sarah Chen · Opened 1 day ago

3 reviews open · 1 critical Avg resolution time: 4.2 days

Fairness thresholds and business-segment definitions are reviewed quarterly by the AI Governance Council.