Responsible AI
Responsible AI assurance
Ten dimensions evaluated on a rolling weekly cycle across all customer-impacting systems. Any dimension below target opens a finding with a named owner and a remediation deadline.
Weighted mean of ten dimensions. Fairness and contestability are the two dimensions currently below their board-agreed target.
Dimension profile
Score against target
Fairness
84
target 90 · Two lending cohorts outside parity band.
Explainability
91
target 90 · Reason codes on all credit-impacting outputs.
Transparency
93
target 88 · AI disclosure present on customer channels.
Privacy
88
target 92 · Masking gap on two vendor-bound pipelines.
Bias detection
82
target 88 · Weekly cohort scans on 31 of 38 models.
Robustness
86
target 85 · Adversarial suite passing at 4.2σ.
Human oversight
94
target 90 · HITL enforced on all restricted actions.
Contestability
79
target 85 · Appeal path missing for collections outcomes.
Safety
92
target 90 · Toxicity and self-harm filters on all channels.
Security
91
target 92 · Injection screen degraded on one retrieval leg.
Fairness — retail lending approval rates by cohort
Approval rate vs benchmark; a gap above 5 points opens a finding