Artificial Intelligence and the Lending Lifecycle
Algorithms and AI Contribute to Consumer Equal Access to Credit
by: Joe DeCosmo, Chief Analytics and Technology Officer
Technological change accelerated during the pandemic, leading many people to adopt new ways to complete everyday tasks. Online tools and mobile applications have exploded for everything from shopping, food delivery, and even financial services.
Fintechs like Enova have led the way in providing working people with online access to financial services regardless of where they live, what they look like, or whether they have an imperfect credit history. Doing so requires technical innovation. For example, most banks and non-bank financial services companies are looking to a full range of automation technologies, from Robotic Process Automation (RPA) and Machine Learning (ML) to Artificial Intelligence (AI) approaches that combine human judgment and AI intelligence.
A significant focus of lenders and industry stakeholders, from analysts to regulators, has been concentrated on underwriting—the use of automation to support or fully replace credit decision-making. While this application has helped improve access to credit for millions of people, full implementation of ML and AI goes beyond that.
From Artificial to Augmented Intelligence
Enova has been looking at AI more broadly.
First, we believe that RPA, ML, and AI come together as a holistic set of support mechanisms for our internal staff and customers. The main difference between ML and AI is the difference between learning and decision-making. We take an even broader approach, viewing automation as “Augmented Intelligence” that increases the efficiency and effectiveness of our team while giving our customers a better experience. ML and AI let lenders look beyond a credit score and see data patterns that help establish the character and capability to repay that are the foundation of lending.
Second, Enova deploys AI throughout the customer lifecycle. While credit approvals and product offers undoubtedly benefit from AI, it can also support everything from the application process to providing proactive customer support that helps borrowers repay on time and in full. As a result, our customer service agents can spend their time on higher value-added activities. Doing so is critical when working with people who have less than perfect credit. The extra time a service representative (or a smart chat tool) spends with an applicant working through their questions or information can help them access the services they need.
In other words, we see these technologies as a way to address unnecessary friction, and to empower our team to work better and smarter rather than a convenient way to reduce headcount.
What About Bias?
The idea of AI bias is one of the most common objections to using traditional ML and AI solutions in lending. The concern often centers on whether algorithms embed historical discriminatory lending practices into automating credit decisions.
For example, many regulators and industry watchers are paying close attention to the risks that automation could turn out to be the latest iteration of practices such as redlining. They suspect that algorithms might intentionally or inadvertently exclude potential borrowers by race, ethnicity or other demographic data. In theory, algorithms could use data that correlates to racial or ethnic identity, among other traits, just as underwriters used neighborhood locations to exclude borrowers when redlining was an accepted practice.
De-Biasing with Augmented Intelligence
At Enova, we believe that bias in AI is an important concern for businesses to address, but also that it is not inevitable. In fact, our models increase access to credit for a broader range of individuals who might be turned down by a human or by a bank or other traditional credit institution. We have built technical and operational controls into our approach to AI throughout the customer lifecycle, from application to repayment. Our goal is to drive both accuracy and consistent, fair application of rules.
Governance and monitoring play a critical role in de-biasing our use of AI. We regularly review the inputs, behaviors, and outcomes of all our models. We have also built audit trails to ensure that models do not discriminate against any groups or protected classes. This scrutiny allows us to identify when models have started to optimize themselves for undesirable discriminatory outcomes.
Developing our own approach to Augmented Intelligence also means we have control over it. It is not a “black box” whose outputs cannot be traced back to rules and inputs. Nor is it automated to the point where optimizations can run amok in ways that lead to bias by using factors such as residence stability, employment history, and others as correlations to credit decisions. We can manage all factors, including how it scores and optimizes and the frequency of model updates.
Model update frequency plays a significant role in this process because it allows us to see and intervene in any issues before deploying changes. Although high-frequency fraud model updates make sense given the relentless evolution of scams and bad actors, factors such as the ability and likelihood to repay a loan have a longer lifecycle. As a result, we can update those models more slowly and in a more controlled fashion, balancing risk management and fairness.
A Broader Set of Use Cases
We also believe that using AI across the entire customer lifecycle can further open up credit to assist underserved populations, including those banks often reject. Especially with our large customer base of subprime and near-prime borrowers, cumbersome aspects of the process can discourage, and therefore exclude, potential borrowers.
Inexperienced borrowers who do not have a robust credit history provide a good case in point. Their credit scores do not necessarily reflect their likelihood of repaying a loan, while looking at their complete tradeline data offers more detail about the timeliness of their payments and other behaviors that may qualify them for certain lending products.
Similarly, inexperienced borrowers may not have the same degree of documentation needed to verify their assets. While a human reviewer could make biased decisions based on documentation quality, banks, transaction sources, sources of cash deposits, and more, an AI can explicitly avoid such biases, doing so consistently and fairly. As a result, it makes more equitable decisions on approval, loan products, and interest rates.
The Full Lending Lifecycle Approach
When used correctly, Augmented Intelligence can achieve two aims: lower friction at every point in the lending lifecycle and, as a result, more successful outcomes for borrowers (i.e., full repayment within their loan terms). We have learned several important things in our AI journey.
Most notably, sub-prime customers are unique, and we can tailor AI to their needs. AI augments our capacity to identify risk profiles and assess the ability to pay outside of the typical target credit score. It also allows us to do so at scale, reaching the nearly one-third of consumers in the U.S. with subprime credit (per Experian). With these consumers, a seamless process can make the difference between these individuals having their financial needs met or overdrawing a bank account. Our experience has allowed us to evolve our AI and our entire organization to meet the needs of this customer base.
In addition, we have learned how to tailor the support we offer to increase success rates. Time-consuming, complicated customer support can discourage borrowers to the point of causing them to miss payments and then struggle to catch up. But AI applied to customer support can anticipate their needs and predictively help them receive an answer quickly. As a result, we lower their barrier to success. We can even apply it proactively to identify potential issues, such as the success or failure of their next-debited payment.
When AI augments rather than replaces good lending practices, it supports fairer outcomes and improves financial wellbeing. We see our business ultimately as helping the customer be successful. AI provides us with the means to achieve that end.