Fifth Third Bank is using AI to enhance processes spanning underwriting, portfolio management, marketing and customer experience, and Jay Budzik is playing a key role in executing those initiatives.  

Budzik, senior vice president and director of AI/ML development at the $300 billion bank, views AI as a key driver of business growth rather than merely an efficiency or productivity tool, he told FinAi News.  

“A lot of folks talk about AI from the perspective of trying to make the work faster,” he said. “The growth story that AI enables is kind of the unsung benefit.” 

For example, Fifth Third recently deployed AI to enhance credit card underwriting accuracy, translating to billions of dollars more in consumer balances despite cards accounting for a relatively small portion of its operations, Budzik said.  

Budzik, who is scheduled to speak at the FinAi Lending Summit in October, recently sat down with FinAi News to discuss AI integration into Fifth Third’s mobile app, AI’s potential to transform lending processes and the future of customer-facing AI. 

What follows is an edited version of the conversation.  

FinAi News: What recent AI-related developments at Fifth Third Bank are you most excited about?  

Jay Budzik: On June 22 we launched what we call universal search for our chatbot, Genie, in the mobile app. This is a small language model to front-end our search and customer service interactions. We have around 10,000 individual functions in the app, many of which are kind of hidden under the hamburger menu, and you have to navigate that to get to the option that you want. 

The thought was that by using AI, we could get people to their desired action faster, and we’re seeing a takeoff in the usage of this functionality. We had a chatbot on the website that could previously answer around 60% of customers’ questions, without requiring an escalation to a customer service agent. We’re now at around 80%.

We also recently launched our credit card underwriting model that incorporates relationship data, including a customer’s history with us and their use of their deposit account. That might seem obvious, but it’s not something the industry has historically done to a significant degree.

A checking account is the primary operating account for many people’s personal finances. That makes it extremely rich in terms of understanding someone’s financial behavior.

By using AI for credit card underwriting, we’re seeing a 20% increase in the approval rate on organic applications — people who are unsolicited or just walk into a branch — with no increase in bad deals. We’re not losing money. We’re expanding access to people who the model identifies as creditworthy.

We’re also using the model in a process we call instant pre-screen. We determine in advance whether someone is eligible for a credit card but still have to determine which customers should actually be selected for an offer. This model allows us to identify 50% more people who should be selected for a card.  

FinAi: What are your primary considerations when developing the infrastructure for customer-facing AI tools? 

JB: First and foremost, we need to protect our customers’ data. There are a whole set of processes related to information security and data exfiltration risk, especially when you talk about agentic AI. There are a variety of systems working together and being coordinated, usually through an API gateway or a [model context protocol] layer on top of that gateway. So it’s really important to have a thorough review of all the vendors involved and the data moving through those systems. 

That applies not just to information security and the authorization and authentication flows that work through the system, but also to the accuracy and errors that are introduced in a multi-agent system. You might test each individual component and see an acceptable level of error, but when you string them all together, those errors can compound, and you end up with a result that’s not acceptable at the end. 

It really requires a new way of thinking about models and systems. It used to be that the unit of control or the unit of analysis when we thought about model risk was the model. But increasingly, these models are being strung together into compositional processes that require us to think about the system as the unit of control versus the model. 

The other thing is, how easy is it to integrate? How much does it cost? I need to look at the total cost of implementing the AI solution. I need to integrate with my legacy systems. I need it to be easy to manage. I need it to integrate with the vendors. That leads us to look for solutions that are reusable and extensible. 

FinAi: How would you describe Fifth Third’s strategy regarding customer-facing AI? 

JB: We definitely want to meet our customers where they are and enable our capabilities to be embedded into the experiences that people use. 

One example of that is our Newline offering, which is our embedded payments platform. It allows app developers to embed money movement into their applications. Think of a game where you want to be able to buy some virtual coins. Somebody needs to clear that transaction and make it available.  

We’ve now fully MCP-ified all of those APIs so agent developers can use Newline’s infrastructure — decades of high-volume transaction-processing infrastructure, which handles billions of transactions a day — through agentic front ends. 

FinAi: Where does AI deliver the most value in terms of improving loan portfolio health? 

JB: We’re seeing huge lifts in underwriting. AI also helps with portfolio management, which is particularly important for something like a home equity line of credit. 

Almost half of defaults happen after the 36-month mark. That’s where, for these bigger-ticket products, portfolio management matters the most. We’re seeing gains in portfolio management performance by switching from traditional to AI models. We can see their responsible or irresponsible behavior and use that for more accurate decisions.  

We’re also able to use AI models to power line increases. One of the biggest benefits of using checking account and credit data in an AI model is that we can more effectively identify low-risk, new-to-credit individuals — people who might have a small credit card limit or no credit card, but who are responsible users of their checking account. 

We can ramp them up into enough usage that would allow them to qualify for an auto loan or eventually a mortgage. 

FinAi: How else can AI help identify quality borrowers when applied to front-end processes? 

JB: A lot of our underwriting activity is solicited. We send out direct mail, for example, to get someone to apply for products. If you think about that process, there are lots of opportunities for waste. I could mail the whole country, but my response rate would be terrible. I want to effectively reach the people who are in a high-need state, but who will also qualify for the product.  

We can use AI to improve the selection of the population that we would market to or try to get in the door to begin with. There’s also a ton of opportunity to personalize the offers presented to people to make them reflective of their predicted needs. 

You can offer a home equity line of credit, for example, to refinance a high-interest-rate credit card bill or balance that you’ve been carrying. If you’re paying 29% on some past purchases, it might make sense to trade that for something more like 7% using a portion of your home equity. That type of knowledge is attainable by using AI, and then the offers can be more personalized and targeted using AI.  

Register here for the FinAi Lending Summit, set for Oct. 7-8 in Las Vegas. This inaugural event will include speakers from Fifth Third and Capital One as well as a fireside chat with Piermont Bank founder and Chief Executive Wendy Cai-Lee.