AWS Machine Learning Blog

How Kabbage improved the PPP lending experience with Amazon Textract

This is a guest post by Anthony Sabelli, Head of Data Science at Kabbage, a data and technology company providing small business cash flow solutions.

Kabbage is a data and technology company providing small business cash flow solutions. One way in which we serve our customers is by providing them access to flexible lines of credit through automation. Small businesses connect their real-time business data to Kabbage to receive a fully-automated funding decision in minutes, and this efficiency has led us to provide over 500,000 small businesses access to more than $16 billion of working capital, including the Paycheck Protection Program (PPP).

At the onset of COVID-19, when the nation was shutting down and small businesses were forced to close their doors, we had to overcome multiple technical challenges while navigating new and ever-changing underwriting criteria for what became the largest federal relief effort in the Small Business Administration’s (SBA) history. Prior to the PPP, Kabbage had never issued an SBA loan before. But in a matter of 2 weeks, the team stood up a fully automated system for any eligible small business—including new customers, regardless of size or stature—to access government funds.

Kabbage has always based its underwriting on the real-time business data and revenue performance of customers, not payroll and tax data, which were the primary criteria for the PPP. Without an established API to the IRS to help automate verification and underwriting, we needed to fundamentally adapt our systems to help small businesses access funding as quickly as possible. Additionally, we were a team of just a few hundred joining the ranks of thousands of seasoned SBA lenders with hundreds of thousands of employees and trillions of dollars in assets at their disposal.

In this post, we share our experience of how Amazon Textract helped support 80% of Kabbage’s PPP applicants to receive a fully automated lending experience and reduced approval times from multiple days to a median speed of 4 hours. By the end of the program, Kabbage became the second largest PPP lender in the nation by application volume, surpassing the major US banks—including Chase, the largest bank in America—serving over 297,000 small businesses, and preserving an estimated 945,000 jobs across America.

Implementing Amazon Textract

As one of the few PPP lenders that accepted applications from new customers, Kabbage saw an increased demand as droves of small businesses unable to apply with their long-standing bank turned to other lenders.

Businesses were required to upload documents from tax filings to proof of business documentation and forms of ID, and initially, all loans were underwritten manually. A human had to review, verify, and input values from various documents to substantiate the prescribed payroll calculation and subsequently submit the application to the SBA on behalf of the customer. However, in a matter of days, Kabbage had tens of thousands of small businesses submitting hundreds to thousands of documents that quickly climbed to millions. The task demanded automation.

We needed to break it down into parts. Our system already excelled at automating the verification processes commonly referred to as Know Your Business (KYB) and Know Your Customers (KYC), which allowed us to let net-new businesses in the door, totaling 97% of Kabbage’s PPP customers. Additionally, we needed to standardize the loan calculation process so we could automate document ingestion, verification, and review to extract only the appropriate values required to underwrite the loan.

To do so, we codified a loan calculation for different business types, including sole proprietors and independent contractors (which totaled 67% of our PPP customer base), around specific values found on various IRS forms. We bootstrapped an initial classifier for key IRS forms within 48 hours. The final hurdle was to accurately extract the values to issue loans compliant to the program. Amazon Textract was instrumental in getting over this final hurdle. We went from POC to full implementation within a week, and to full production within two weeks.

Integrating Amazon Textract into our pipelines was incredibly easy. Specifically, we used StartDocumentAnalysis and GetDocumentAnalysis, which allows us to asynchronously interact with Amazon Textract. We also found that using forms for FeatureTypes was well suited to processing tax documents. In the end, Amazon Textract was accurate, and it scaled to process a substantial backlog. After we finished integrating Amazon Textract, we were able to clear our backlog, and it remained a key step in our PPP flow through the end of the program.

Big impact on small businesses

For perspective, Kabbage customers accessed nearly $3 billion in working capital loans in 2019, driven by almost 60,000 new customers. In just 4 months, we delivered more than double the amount of funding ($7 billion) to roughly five times the number of new customers (297,000). With an average loan size of $23,000 and a median loan size of $12,700, over 90% of all PPP customers have 10 or fewer employees, representing businesses often most vulnerable to crises yet overlooked when seeking financial aid. Kabbage’s platform allowed it to serve the far-reaching and remote areas of the country, delivering loans in all 50 US states and territories, with one third of loans issued to businesses in zip codes with an average household income of less than $50,000.

We’re proud of what our team and technology accomplished, outperforming the nation’s largest banks with a fraction of the resources. For every 790 employees at a major US bank, Kabbage has one employee. Yet, we surpassed their volume of loans, serving nearly 300,000 of the smallest businesses in America for over $7 billion.

The path forward

At Kabbage, we always strive to find new data sources to enhance our cash flow platform to increase access to financial services to small businesses. Amazon Textract allowed us to add a new arrow to our quiver; we had never extracted values from tax filings prior to the PPP. It opens the opportunity for us to make our underwriting models more rich. This adds another viewpoint into the financial health and performance of small businesses when helping our customers access funding, and provides more insights into their cash flow to build a stronger business.

Conclusion

COVID-19 further revealed the financial system in America underserves Main Street business, even though they represent 99% of all companies, half of all jobs, and half of the non-farm GDP. Technology can fix this. It requires creative solutions such as what we built and delivered for the PPP to fundamentally shift how customers expect to access financial services in the future.

Amazon Textract was an important function that allowed us to successfully become the second-largest PPP lender in the nation and fund so many small businesses when they needed it the most. We found the entire process of integrating the APIs into our workflow simple and straightforward, which allowed us to focus more time on ensuring more small businesses—the backbone of our economy—received critical funding when they needed it the most.


About the Author

Anthony Sabelli is the Head of Data Science for Kabbage, a data and technology company providing small businesses cash flow solutions. Anthony holds a Ph.D. from Cornell University and an undergraduate degree from Brown University, both in applied mathematics. At Kabbage, Anthony leads the global data science team, analyzing the more than two million live data connections from its small business customers to improve business performance and underwriting models.