Artificial Intelligence in Lending

“Can machines think?” Alan Turing posed this question at the outset of his 1950 paper “Computing Machinery and Intelligence,” a seminal piece of literature in the field of artificial intelligence. Turing wanted to know whether computers would eventually imitate humans’ responses so well that people wouldn’t be able to tell whether they were interacting with a human or a machine.

Decades later, in an era when computers are capable of recognizing and responding to human speech, processing images and even driving cars, the question for data scientists and engineers has become “What else can machines think about?”

In my previous position at OnDeck Capital, I became particularly interested in how AI could be used to rethink the financial industry. OnDeck is a lending company that provides working capital and loans to small businesses. My work there provided my first exposure to computers making decisions that were previously reserved for humans.

Before this begins to sound too much like sci-fi, let me clarify that the work we were doing at the time—and the mission of the banks and fintech companies we work with at Yhat—is still created and maintained by humans. This work is also subject to the same financial regulations as lending decisions made by humans, like the Fair Credit Reporting and Equal Credit Opportunity Acts. The algorithms generated by these computer systems do not replace human logic, ethics or creativity but augment and automate it.

For example, consider a core question in lending: “Will this person be able to repay a loan?” For banks, the due diligence required to answer this question for small businesses looking for loans in the median range of $140,000 is time- and cost-intensive. A local shop that needs enough capital to cover its payroll during the slow season may not receive a decision for weeks or even months.

Thanks to an abundance of novel data sources, new open source technologies and libraries, and a massive decrease in computing costs over the past decade, alternative lenders are able to make faster, more informed decisions than human underwriters. Companies “train” automated lending systems on previous lending decisions and ask those systems to detect and codify a way to “think” about making future lending decisions. Continuous underwriting—that is, evaluating credit risk before approval, during repayment, and after a loan has been paid in full--creates a much wider surface for understanding their customers.

Most of the non-traditional fintech players we work with have invested heavily in continuous underwriting capabilities. Estimating risk throughout the lifecycle of a loan from origination through servicing as opposed to a snapshot approach at the time of the application gives a lender a far stronger posture from which to influence loss outcomes. This type of active underwriting is also incredibly powerful with regard to the targeted and timely mid-term offers and other powerful retention and cross-sell strategies it enables you to do.

Each company’s credit decision algorithm is unique but often includes variables like cash flow metrics and social data that have been linked to borrowers’ ability to repay loans.

Machines have also proven more successful than humans in detecting suspicious applicants in the loan process. Fraud in finance comes in many forms, from “innocent” misinformation (like using my mom’s address instead of mine by mistake) to true felonies like identity theft and money laundering. Computers monitor users’ behavior patterns during the digital application process and detect anomalies with significant correlations to fraud, such as applying for a loan at odd times or inputting multiple social security numbers. This may trigger additional identity checks like a call to a credit bureau dataset, or the user may simply be asked for more information.

AI has already impacted credit decisioning and fraud detection, but I believe that that’s only the tip of the iceberg. In other areas of finance we’ve seen how computers can nudge human behavior. Smart wallet apps like Mint and Digit learn users’ habits and coach them to better manage their money. What if this contextual awareness were integrated with consumer and small business loans to inform and personalize collections practices? I’m also optimistic about how AI can be used to provide access to credit for unbanked and underserved populations, like microfinance organizations like Kiva have set out to do.

One of the most fascinating aspects of AI in lending is just how seamlessly human and machine collaborate to perform a centuries-old task—the movement of money. As Turing predicted, the line between human and computer has blurred so that applicants can’t always tell which one is on the other end of their interaction.

Machine learning and AI techniques allow for immense economies of scale. This makes them appealing among to businesses and is likely why fintech thought leaders have invested heavily in them. AI-based automated underwriting not only reduces the cost of origination for the lender, it also means approval policies can be based on thousands of data points rather than just a few that a human underwriter is capable of considering by hand. This is really important because it means that credit unions can expand the eligibility criteria for their products, resulting in more financial inclusion.

My hope for AI in lending is that our technological advances will both break down extraneous barriers to accessing capital and promote financial and social responsibility.

Austin Ogilvie is CEO and co-founder of Yhat, Inc., a machine learning software company based in Brooklyn. He was previously at OnDeck Capital, the largest online small business lender in the United States.