Credit scoring is one of the most regulated processes in banking. Every adverse credit decision — a denial, a less favorable rate, or a reduced credit line — triggers regulatory requirements under the Equal Credit Opportunity Act (ECOA) and Regulation B. Lenders must provide adverse action notices that explain the specific reasons for the decision. For large lending operations processing thousands of applications daily, generating these narratives manually is expensive and slow.
The Challenge of Credit Decision Narratives
A typical bank underwriter spends 15-30 minutes per application writing the credit decision narrative. This includes summarizing the applicant's financial profile, explaining the risk assessment, documenting the decision rationale, and drafting the adverse action notice when applicable. For a bank processing 500 applications per day, that's 125-250 hours of narrative writing daily — time that could be spent on more complex credit decisions.
Fine-tuned LLMs address this challenge by automating the narrative generation process. The model takes structured data from the credit application — credit score, income, debt-to-income ratio, loan-to-value ratio, and internal risk ratings — and generates a complete, compliant credit decision narrative in seconds. The underwriter reviews and approves the narrative rather than writing it from scratch.
How the Fine-Tuning Works
The fine-tuning process starts with historical credit memos from the institution's own underwriting system. These documents contain the complete record of credit decisions: the applicant data, the risk assessment, the decision, and the narrative explanation. Each memo is structured as an instruction-output pair, where the input is the structured application data and the output is the completed narrative.
Using LoRA fine-tuning on a Llama 3.3 70B base model, banks can achieve high-quality narrative generation with 500-1000 training examples. The model learns the institution's specific narrative style, credit terminology, and regulatory language. It also learns to identify which factors in the application data are most relevant for the narrative, ensuring that each generated narrative includes the specific reasons required by ECOA.
The fine-tuned model is deployed on-premise via Ollama, keeping all application data within the bank's security perimeter. This is essential for compliance with data privacy regulations and for maintaining the confidentiality of credit decision methodologies.
Regulatory Compliance
The ECOA adverse action notice requirement is the primary regulatory driver for this use case. Under Regulation B, lenders must provide applicants with specific reasons for adverse credit decisions — not just a general statement. The fine-tuned model generates narratives that include the specific factors that led to the decision, such as "credit score below minimum threshold" or "debt-to-income ratio exceeds policy limits."
To maintain compliance, the system operates on a human-in-the-loop model. The fine-tuned model generates a draft narrative, which is reviewed and approved by a qualified underwriter before being sent to the applicant. The system logs all generated narratives and their final approved versions, creating an audit trail for regulatory review.
Results and ROI
Banks implementing fine-tuned credit narrative generation report 70-80% reduction in narrative writing time, with underwriters spending an average of 3-5 minutes reviewing and approving AI-generated narratives instead of 15-30 minutes writing them from scratch. Narrative quality improves due to consistency and adherence to regulatory requirements. The reduction in repetitive writing work also improves underwriter satisfaction and retention.
The total cost of implementing a fine-tuned credit narrative system is $20,000-$50,000 for model development and infrastructure, with ongoing costs of $2,000-$5,000 per month for inference and maintenance. Most banks achieve positive ROI within 3-6 months based on underwriter productivity gains alone.