Proposal for Responsible AI-Powered Credit Deicisions

| AI Governance · Academic Papers

Note: This is a project I completed for Purdue’s AI Management and Policy Master’s Program . The data and business scenarios are part of a simulated case study and do not represent the actual operations of any real-world entity.

Executive Summary

Ficus Bank (Ficus) faces a pivotal choice: continue with manual underwriting processes that are increasingly slow and exclusionary or transition to an AI-driven model. While AI offers a path to rapid decisioning and broader financial inclusion, it introduces risks regarding algorithmic bias and transparency. This proposal recommends integrating the NIST AI Risk Management Framework (AI RMF) to ensure that Ficus’s modernization is both compliant with emerging state laws (like the Colorado AI Act) and aligned with the bank’s core values of equity and security.

Introduction

Ficus Bank has long been an economic anchor for its community. However, a recent surge in small business loan applications has strained manual underwriting, creating bottlenecks and frustrating customers. Furthermore, traditional credit metrics often exclude creditworthy “nontraditional” businesses.

To stay competitive, Ficus plans to replace manual workflows with machine-learning-based underwriting. By considering alternative data—such as rental histories, utility payments, and employment records—the bank can reduce decision times from days to minutes. However, moving from human logic to “black box” algorithms requires a fundamental shift in risk management.

The Risks of AI Credit Decisioning

1. The Alignment Problem

Ficus must ensure the AI system balances efficiency with ethics. The project must:

2. Algorithmic Bias & Proxy Variables

Unlike traditional software, AI is trained, not programmed. It can internalize historical socio-economic disparities. Even without explicit “race” or “gender” data, the model may use proxy variables (e.g., zip codes or specific shopping habits) to infer protected characteristics.

Data Insight: Existing Disparities in Lending Current data from the Federal Reserve’s Small Business Credit Survey (2024/2025) highlights why manual processes often fail:

  • Black-owned firms are approved for the full amount of credit they seek only 28% of the time.
  • White-owned firms see a much higher full-approval rate of 54%.
  • Hispanic-owned firms fall in between at approximately 31%. AI aims to bridge this gap, but without rigorous oversight, it risks automating and scaling these very inequities.

3. The “Black Box” Problem

“Deep” machine learning models encode millions of relationships, often making it impossible for humans to see why a specific decision was made. This opacity makes it difficult to provide the “adverse action” reasons required by the Equal Credit Opportunity Act (ECOA).

Compliance in an Evolving Policy Landscape

Ficus operates in a shifting regulatory environment where federal and state requirements are increasingly divergent.

Policy Options

Option Strategy Advantages Tradeoffs
1 Avoid the Risk No AI-specific compliance costs. High long-term strategic risk; loss of market share to Fintechs.
2 Targeted Compliance Low upfront cost; uses existing ERM. “Tick-box” approach; high risk of reputational damage.
3 Integrate NIST AI RMF Strong compliance; future-proofed. High upfront investment in governance.

Recommendations: Integrating the NIST AI RMF

We recommend Option 3. To implement this, Ficus should act across the four NIST pillars:

  1. Govern: Establish a cross-functional AI Governance Committee (Legal, Risk, IT) to align AI use with bank values.
  2. Map: Conduct a Governance Gap Analysis to identify where legacy risk structures fail to address data provenance and model lineage.
  3. Measure: Perform a Pre-deployment Impact Assessment to document training data and identify proxy discrimination before it hits the market.
  4. Manage: Define Explainability Solutions, using interpretable machine learning tools to ensure Ficus can always provide specific reasons for loan denials.

Conclusion

AI underwriting is an operational necessity, but it cannot be deployed through legacy governance. By adopting the NIST AI Risk Management Framework, Ficus Bank can solve its operational bottlenecks and expand financial inclusion while securing its reputation as a leader in responsible, secure innovation.

References