Automating Business Lending and Underwriting - Powered by LLMs (2024)

The Evolution of Automatic Underwriting: Leveraging LLMs for Better Decision-Making

As the financial landscape rapidly evolves, automated underwriting systems powered by cutting-edge Large Language Models (LLMs), are at the forefront of this transformation. These innovations are not only reshaping how lenders and underwriting technology providers operate but also significantly boosting productivity and accuracy in decision-making.

Automated Underwriting Systems—Explanation & Industry Adoption

A report by McKinsey found that nearly 60% of leading financial institutions turn to advanced Artificial Intelligence (AI) technologies to automate processes—including underwriting. More recently, NVIDIA’s State of AI in Financial Services Report from 2024 found that an overwhelming 91% of financial services companies are either assessing AI or already using it in production. 

Automated Underwriting Systems (AUS) automate the risk evaluation and underwriting of borrowers’ loan or mortgage applications. 

Historically, the underwriting process for businesses and other legal entities has been highly manual. AUSs leverage algorithms and AI to automate various aspects of the process.

Here is how it works:

1. Input: The underwriter feeds the key data points and documents into the AUS software.

2. Analysis: The AUS applies algorithms and AI to analyze the data, assessing whether the loan meets the lender’s criteria and determining the associated risks (including the appropriate spread).

3. Output: The AUS returns the loan eligibility, the prospective loan offer, and might request additional documents from the borrower if needed. The data returned by the AUS can be used to make a decision automatically or augment the decision process of a human that makes the final decision. 

What are the Benefits of Automated Underwriting Systems?

Besides the automation of manual human labor and the faster processing of data, automated underwriting provides many benefits:

Better Accuracy and Outcome Prediction

AI can analyze millions of data points, which results in more accurate outcome predictions. Additionally, humans make errors when typing numbers into spreadsheets or analyzing an application against hundreds of relevant guidelines. If these errors result in the incorrect denial of the loan application, they can lead to a significant loss in revenue.

As we will learn in the next section, LLM-powered tools like Trellis constitute a new level of accuracy. They can reliably extract unstructured and complex data points across millions of documents. 

Reduced Bias and Subjectivity 

Leveraging algorithms and AI minimizes human bias and subjectivity, promoting consistency and fairness in lending decisions. This reduction in bias enhances compliance with fair lending laws and regulations, ensuring equitable treatment of all loan applicants.

Extended Addressable Market

Thanks to reduced time and cost constraints, lenders can now cater to a broader customer base, including markets previously deemed less economically viable. This expansion diversifies their portfolio and opens up new revenue streams by reaching customers who were once overlooked.

Automating Underwriting System Technology

Simple Automation

Knockout questions are a simple form of automation that allows borrowers to triage and filter loan applications based on their custom criteria. For example, underwriters might auto-reject specific applications deemed too risky based on particular application fields. This could include the prospective borrower’s industry or the loan size they are applying for. However, knockout questions require the applicant to invest time in filling out questionnaires. They are also less reliable in assessing the true financial health of the applicant. 

Data Extraction with OCR

Optical Character Recognition (OCR) and its slightly improved version, Intelligent Character Recognition (ICR), are techniques used to digitize text. They do this by identifying individual characters in scanned images.

We can use OCR and other techniques to extract specific data from the prospective borrower's documents. Documents including:

• Bank Statements: Shows cash flow stability, average balances, and financial management.

• Tax Returns: Provides insights into the business's earnings and tax compliance over several years.

• Credit Bureau Checks: Reflects the business's credit history and financial reliability.

• Firmographic Information: Includes crucial data such as industry type, size of the company, years in operation, and ownership structure.

• Know-Your-Business Compliance Checks: Necessary for verifying the legitimacy of the business and for anti-money laundering purposes.

• Account Verification: Confirms the authenticity of the business's financial accounts.

This technology can be effective for structured data, like financial tables, but has four significant limitations:

• Accuracy issues: OCR struggles with text clarity due to low-quality scans or unusual fonts. In underwriting, errors in digitizing key financial figures or contract terms can lead to wrong risk assessments.

• Rigidity in document handling: OCR’s dependence on predefined rules for data interpretation is a significant drawback in underwriting. Financial statements and loan agreements often vary in format. This causes OCR to misclassify or overlook critical data.

• Limited contextual understanding: OCR can’t comprehend the context or significance behind a text. This is crucial in underwriting. For instance, it fails to interpret the nuanced content in risk assessments or insurance documents.

• Complex data challenges: OCR may misinterpret key financial ratios in business plans or risk factors in insurance underwriting documents if presented in unconventional formats or alignments.

Deep Learning and Large Language Models in Underwriting

Large Language Models (LLMs), like those used in tools such as Trellis, advance beyond traditional OCR by using deep learning to enhance data extraction from unstructured documents in business lending:

• Flexibility: LLMs handle a variety of document formats and complex layouts common in business lending without predefined rules. These include diverse financial statements and intricate loan agreements.

• Contextual Understanding: LLMs grasp the context of unstructured text. They can interpret complex narratives, which are essential in underwriting. This capability enables them to analyze risk factors, operational strategies, and compliance issues in depth.

Here are some examples of data extraction for underwriting use cases LLMs enable:

1. Revenue Recognition Policies: “Extract the revenue recognition policy for software sales from the latest annual report.”

2. Customer Concentration Risks: “Identify the top 3 customers by revenue and their respective percentages of total revenue for the latest fiscal year.”

3. Loan Covenant Compliance: “Verify if the company complies with the debt-to-EBITDA covenant from the latest loan agreement.”

4. Litigation Risks: “Summarize any ongoing or potential material litigation disclosed in the risk factors section.”

5. Supply Chain Disruptions: “Rate the risk of supply chain disruption for the company's operations or key suppliers on a scale of 1 to 10.”

6. Shares Outstanding: “Extract basic shares outstanding from the Consolidated Statements of Income from the latest fiscal year-end date.”

These advanced language comprehension capabilities enable LLMs to extract any kind of data points from diverse and long documents. This enables a more accurate and comprehensive risk assessment in underwriting.

The Future of Underwriting with LLMs

Trellis' platform enables lenders and underwriting technology providers to leverage the latest LLM advancements for accurate and efficient data extraction from unstructured sources. Here's how it works:

1. Upload Documents: Upload the borrower's documents to the Trellis platform. These can include financial statements, loan agreements, and business plans.

2. Define Transformations: Next, configure your Transformations and specify the data you want to extract, classify, or generate. For example, you could create a Transformation to extract key financial ratios, identify risk factors, or summarize the company's business strategy.

3. Run Queries: Once you define the transformations, Trellis applies them to all the uploaded documents and extracts the relevant data into a SQL-queryable format.

4. Analyze and Integrate: The extracted data can be seamlessly integrated into your existing underwriting systems or analyzed further using Trellis' built-in analytics tools.

You can also add additional Transformations later. Previously, adding a new data point to your underwriting pipeline might have required expensive engineering work to build a new custom extractor. Now, you can add it in a few clicks on a one-line change in your Transformation code using our developer-friendly API.

This approach streamlines the data extraction process, reducing the need for manual review. Lastly, it enables lenders to make more informed and accurate underwriting decisions based on comprehensive data insights.

Conclusion

Traditional data extraction methods falter when faced with the diverse and unstructured documents involved in underwriting. Trellis is the first truly LLM-first platform that transforms complex unstructured data into SQL-queryable formats, enabling effortless integration into existing underwriting systems.

By leveraging Trellis' AI-powered transformations through their developer-friendly APIs, lenders can seamlessly replace convoluted bespoke solutions and extract deep, actionable insights from financial statements, loan agreements, business plans, and more. This powerful capability facilitates comprehensive risk assessments and informed lending decisions at scale.