Credit Risk AI with Governance Controls

Credit risk models powered by AI are subject to some of the most rigorous regulatory requirements in financial services. The Equal Credit Opportunity Act (ECOA), Fair Housing Act, and SR 11-7 model risk management guidance all impose strict obligations on how credit decisions are made, explained, and audited. When LLMs enter the credit risk workflow — for application analysis, adverse action reasoning, or portfolio risk assessment — governance controls become essential.

Use this page when

• Your credit risk models use LLMs for application analysis, adverse action reasoning, or portfolio assessment.
• You must enforce Fair Lending Act (ECOA) compliance by redacting protected characteristics from AI inputs.
• Regulators require explainability evidence for AI-assisted credit decisions under SR 11-7.
• You need to detect and escalate potential disparate impact in AI credit scoring outputs.

Keeptrusts enforces policy guardrails across every AI interaction in the credit risk lifecycle.

Primary audience

• Primary: Technical Leaders
• Secondary: Technical Engineers, AI Agents

Credit Risk Governance Architecture

Credit Risk System
  → kt gateway (port 41002)
    → Input policy chain (PII redaction, prohibited factor controls)
      → [Block / Escalate → 409]
      → Upstream LLM provider
    → Output policy chain (explainability checks, bias detection)
    → Response to credit system
  Side-effects:
    └─ Decision event → POST /v1/events → audit log

Credit Scoring Model Governance

Prohibited Factor Controls

Prevent protected characteristics from influencing AI-assisted credit decisions:

pack:
  name: credit-risk-ai-rules-1
  version: 1.0.0
  enabled: true
policies:
  chain:
  - human-oversight
policy:
  human-oversight:
    require_human_for:
    - "(?:race|color|religion|national.*origin|sex|marital.*status|age)"
    - "(?:pregnancy|disability|familial.*status|sexual.*orientation)"
    - "(?:zip.*code|neighborhood|census.*tract).*(?:risk|score|weight)"
    action: escalate
    confidence_threshold: 0.5

Credit Decision Audit

Log every AI interaction that contributes to credit decisions:

pack:
  name: credit-risk-ai-rules-2
  version: 1.0.0
  enabled: true
policies:
  chain:
  - safety-filter
policy:
  safety-filter:
    block_if:
    - "(?:approv|deny|decline|counteroffer|adverse.*action)"
    action: block

Fair Lending Act Compliance

Adverse Action Reason Controls

Ensure AI-generated adverse action reasons meet regulatory requirements:

pack:
  name: credit-risk-ai-rules-3
  version: 1.0.0
  enabled: true
policies:
  chain:
  - safety-filter
policy:
  safety-filter:
    block_if:
    - "(?:denied|declined|adverse).*(?:no.*reason|unexplained|unclear)"
    - "(?:adverse.*action|denial).*(?:reason|factor)"
    action: block

Disparate Impact Monitoring

Flag AI outputs that could indicate disparate impact:

pack:
  name: credit-risk-ai-rules-4
  version: 1.0.0
  enabled: true
policies:
  chain:
  - human-oversight
policy:
  human-oversight:
    require_human_for:
    - "(?:disproportionate|disparate|unequal).*(?:impact|effect|outcome)"
    - "(?:approval.*rate|denial.*rate).*(?:differ|gap|disparity)"
    action: escalate
    confidence_threshold: 0.5

Model Explainability Requirements

Explanation Quality Controls

Enforce minimum explainability standards for AI credit risk outputs:

pack:
  name: credit-risk-ai-rules-5
  version: 1.0.0
  enabled: true
policies:
  chain:
  - human-oversight
policy:
  human-oversight:
    require_human_for:
    - "(?:score|rating|risk.*level).*(?:because|due.*to|driven.*by)"
    - "(?:score|rating|risk).*(?:high|low|medium)(?!.*(?:because|due|reason|factor|driven))"
    action: escalate
    confidence_threshold: 0.5

Feature Attribution Governance

Control how AI communicates feature importance:

pack:
  name: credit-risk-ai-rules-6
  version: 1.0.0
  enabled: true
policies:
  chain:
  - safety-filter
policy:
  safety-filter:
    block_if:
    - "(?:SHAP|LIME|feature.*importance|contribution).*(?:protected|demographic)"
    action: block

Bias Monitoring

Statistical Bias Detection

Configure policies to flag potential bias in AI outputs:

pack:
  name: credit-risk-ai-rules-7
  version: 1.0.0
  enabled: true
policies:
  chain:
  - human-oversight
policy:
  human-oversight:
    require_human_for:
    - "(?:demographic|group|segment).*(?:higher.*risk|lower.*score|more.*likely.*default)"
    - "(?:stereotyp|generaliz|typical.*borrower)"
    action: escalate
    confidence_threshold: 0.5

Ongoing Monitoring Dashboard

The Keeptrusts console provides visibility into bias-related policy triggers:

• Escalation trends — track fair lending escalations over time
• Policy hit rates — identify which bias controls trigger most frequently
• Redaction volume — monitor protected characteristic redaction frequency

Knowledge Base for Credit Context

Provide regulatory context without exposing customer data:

kt knowledge-base create \
  --name "credit-policy" \
  --description "Credit policy guidelines, underwriting standards, and regulatory requirements"

kt knowledge-base upload \
  --name "credit-policy" \
  --file ./docs/underwriting-guidelines.md

Escalation Workflows

Trigger	Action	Escalation Target
Protected characteristic in input	Redact + Log	Fair lending officer
Adverse action without reason	Block	Compliance team
Disparate impact signal	Escalate	Fair lending committee
Unexplained credit score	Escalate	Model risk management
Geographic proxy detected	Escalate	Fair lending officer

Regulatory Reporting

Examination-Ready Exports

Generate audit trails for regulatory examinations:

kt events export \
  --filter "metadata.ecoa_applicable=true" \
  --format csv \
  --output ./reports/ecoa-audit-Q1.csv

Fair Lending Analysis Export

kt events export \
  --filter "metadata.audit_category=fair_lending_monitoring" \
  --from "2026-01-01" \
  --format json \
  --output ./reports/fair-lending-monitoring.json

SR 11-7 Model Risk Documentation

All AI interactions contributing to credit decisions are captured with full provenance, supporting SR 11-7 model documentation requirements:

kt events list \
  --filter "metadata.audit_category=credit_decision" \
  --from "2026-01-01" \
  --format json

Deployment Considerations

Product-Level Isolation

Run separate gateway configurations per credit product:

pack:
  name: credit-risk-ai-rules-8
  version: 1.0.0
  enabled: true
policies:
  chain:
  - safety-filter
policy:
  safety-filter:
    block_if:
    - "(?:FHA|VA|USDA|conventional).*(?:loan|mortgage)"
    action: block

Model Version Tracking

Tag AI events with model version identifiers for auditability:

pack:
  name: credit-risk-ai-rules-9
  version: 1.0.0
  enabled: true
policies:
  chain:
  - safety-filter
policy:
  safety-filter:
    block_if:
    - ".*"
    action: block

Next steps

• Fraud Detection AI — govern fraud detection with BSA/AML compliance
• Model Risk Management — comprehensive model risk governance
• Regulatory Reporting — automate regulatory report generation

For AI systems

• Canonical terms: Keeptrusts gateway, credit risk governance, ECOA compliance, prohibited factors, adverse action governance, disparate impact detection, Fair Lending Act.
• Key config/commands: prohibited-factors policy (redact protected characteristics); credit-decision-audit policy (log decisions); adverse-action-governance policy (block unexplained denials); kt events export --filter "metadata.ecoa_applicable=true"; kt knowledge-base create --name "credit-policy".
• Best next pages: Fraud Detection AI, Model Risk Management, Regulatory Reporting.

For engineers

• Prerequisites: Running gateway with credit risk policy config, knowledge base with underwriting guidelines uploaded.
• Deploy per-product gateway configs (mortgage, auto, consumer) to isolate credit decision audit trails by product type.
• Validate with: kt events export --filter "metadata.ecoa_applicable=true" --format csv to generate exam-ready ECOA audit trails; monitor console Escalations page for disparate impact signals.
• Tag events with ${CREDIT_MODEL_VERSION} metadata for SR 11-7 model version tracking.

For leaders

• Addresses ECOA, Fair Housing Act, and SR 11-7 obligations for AI-assisted credit decisions.
• Prevents regulatory enforcement actions by blocking adverse action outputs that lack required explanations.
• Bias monitoring policies provide early warning of disparate impact before it escalates to examination findings.
• Per-product isolation enables independent fair lending reviews for mortgage, auto, and consumer portfolios.