AI Lending: Should you build or buy?

by Dan Chiazza, Chief Revenue Officer, Zest AI

The adoption of AI in credit risk underwriting is at full tilt, accelerating as lenders see it as key to building competitive advantages in efficiency and portfolio growth and potentially solve significant industry challenges around inclusive lending.

AI and machine learning-based models have proven to be more accurate at predicting default risk and credit eligibility by generating a more holistic applicant view. That leads to more approvals and, with more confidence about whom to say yes to, higher levels of automated decisioning. A recent McKinsey & Company report echoed the upside of adopting ML in underwriting: "By using machine learning, the new-entrant lenders will be able to automate as much as 95% of underwriting processes while also making more accurate credit decisions."

With the business case justified, leaders are starting to focus on the age-old dilemma of build vs. buy. For ML underwriting solutions, the decision is particularly nuanced. ML models are almost always more powerful, but they can get out of whack unless you firmly manage the build and validation process and monitor them in production. Most importantly, says Rushabh Seth, Director of Credit Risk Strategy & Modeling at BECU, “The organization has to get buy-in every step of the way right up to [regulatory review]. The best way to get buy-in is to make sure that the AI is understandable.” Fortunately, ML models are now fully explainable to compliance teams. Just make sure you’re fair lending testing is as robust as before.

Sins of omission, commission, and employee churn contribute to Gartner’s all-too-high 53% failure rate for AI projects overall. It doesn’t have to be that way. With a better understanding of each approach’s pros and cons, you can make the best decision for your organization.

Before we unpack the approaches, you must start thinking through these essential questions:

• What problem are you trying to solve?
• How critical is a faster time-to-market?
• What are the costs of getting it wrong and the implications of getting it right?
• What are your data sources? Are they accessible and available?
• Do you have the talent and the time?

Answering these questions will help shape your decision criteria. Let’s dig into the framework.

Building An AI Solution

The DIY approach can give you complete control over the model development process and build in-house capability if you plan to develop additional models down the road. But the DIY approach requires collecting the right talent, technology resources, and data to build and deploy a series of compliant ML models. A lot of financial institutions stop short for lack of the right people. “We haven't been able to find enough talent. The talent has to know how to do development, but they also have to understand the business, and that's the problem,” says Roderick Powell, Head of Model Risk Management at Ameris Bank.

It takes specific expertise to get a model through compliance and regulatory review, integrate it with an existing LOS, monitor its health and performance in production, and plan for continuous development to keep up with the pace of innovation. Even large financial institutions with big data science teams struggle to solve these problems, and if they have the staff to do it the way they want, it can take far longer at first and cost more.

Buying An AI Solution

Buying generally gets you to market faster at a lower operational cost. The big “a-ha” for financial institutions comes when they realize that “buying” can mean two things. You have software vendors that offer fixed, pre-built models, and you have software “partners” that offer customizable, off-the-shelf models and help you through the model lifecycle.

A good industry partner should bring the needed domain expertise around model validation and risk management to address key stakeholders’ needs in your organization. Through collaboration with a partner, you end up with better in-house knowledge and more access to industry-first features.

Pre-built models are an excellent one-and-done option. They’re usually cheaper and quicker to deploy than tailored models, but you have to consider the performance gap. A generic model ignores the specifics of your organization's customer base, economy, and products. As a result, they typically underperform tailored models. For example, one tailored Zest-built model helped a lender reduce risk by 25% compared with their pre-built models, resulting in economic savings of 37 million dollars a year.

For many leaders, especially CFOs, the opportunity to make an immediate ROI impact with a new generic or tailored ML model outweighs the benefits of the in-house approach. As Jenny Vipperman, Chief Lending Officer at Vystar Credit Union, said about their switch to ML lending, “We wanted the fastest path to increased financial inclusion and data-driven lending innovation without the high costs or risks. For us, it was about finding the right partner who could provide a tailored model along with the experience and expertise to make our project a success.”

The Costs of Getting It Wrong

As mentioned earlier, ML underwriting demands more rigor than a typical AI software build vs. buy decision. Unlike a DIY AI chatbot that can go on the fritz without seriously damaging the business, an underwriting model is the organization’s heart. If not set up correctly, a flawed model could tank a quarter or two of originations and cause unexpected disparate impact in your community. The CFPB is poised to renew stricter industry oversight, placing more pressure on banks and credit unions to ensure their models meet fair lending standards.  Without the expertise in-house to analyze and validate ML models, your DIY operation could severely impact the long-term success—or even viability—of the business.

The Benefits of Getting It Right

The lenders we talk to want three things: more approvals, lower loss rates, reduced operating expenses. ML underwriting typically increases approvals by 15%, cuts losses by 30%, and doubles automation rates. A faster time-to-market amplifies these benefits.

How fast do you move? According to Cornerstone Advisors’ research report Credit Modeling and the Need for Speed, 21% of financial institutions take more than four months to build, validate and deploy a new model. Lenders taking the DIY approach should factor in additional time for project setbacks. Count on a minimum of 6 months but prepare for ten. With the right vendor and a tailored model, you can get through to deployment in under four months.

How much are lenders leaving on the table by delaying ML deployment timelines? The analysis below reflects actual Zest AI customer numbers.

Conclusion

Developing ML models in-house is a great way to address business challenges with specificity and control. Still, it often requires a disproportionate allocation of resources to create, implement, and maintain.

In most situations, working with a reputable and trusted partner to deploy a proven ML model will allow your organization to capture the ROI more quickly and generally ensures that the model will have more comprehensive support throughout its lifecycle.