Showing your work wasn’t just for seventh-grade math class anymore. When regulators come knocking, they won’t just want the right answer from your credit union’s AI tools—they’ll want to see the exact chain of thought.
In her latest interview for The Credit Union Connection podcast, Sarah Snell Cooke catches up with Titan Chief Customer and Growth Officer Matt Creatore to unpack why general-purpose LLMs are a compliance trap, how vertical-specific banking AI turns manual “stare and compare” busywork into real operating leverage, and why explainability is your ultimate competitive advantage.
Stop waiting on regulatory guidance to build your AI governance framework. Watch or listen to the full conversation on what separates the winners from the losers in the next three years.
NOTE: This transcript may contain minor imperfections courtesy of our AI overlords-in-training. We’re not complaining. We’re definitely not complaining.
Sarah Snell Cooke: Hello, and welcome everyone. I am Sarah Snell Cooke, your host here at The Credit Union Connection. I am joined today by Matt Creatore. Welcome.
Matt Creatore: Thanks for having me on the show, Sarah.
Sarah Snell Cooke: Absolutely. Great to have you, and Matt is the chief customer and growth officer at a company called Titan. Why don’t you tell us a little bit more about yourself and the company?
Matt Creatore: Operating within the financial services space for my entire career, mostly on the SaaS selling side and the go-to-market build side. Most recently at Peach Finance, which is an API-first loan servicing company. And Titan is the AI platform for banking, bringing the models, context, and agents that banks and credit unions need to move faster, smarter, and safer in this new age of AI where all financial institutions are learning how they can adopt and create value out of platforms like ours.
Sarah Snell Cooke: Perfect. Awesome. Thank you so much for that. Right now credit unions are hopefully in the ending phase of the learning part of AI and understanding it better. What is the importance of not waiting for model risk management guidance before building your governance frameworks?
Matt Creatore: The recent guidance from the OCC and the Fed has stayed silent with SR 26-02 on GenAI frameworks because they are in a wait-and-see mode, and the size of the institution may dictate the level of governance expected. What seems to be tried and true across the institutions we engage with is that regulators may ask how AI got to the conclusions that it did. It’s only natural for them if we’re taking humans out of the loop in certain workflows or use cases. We need to understand how the AI reached the response or outcome that it did. What off-the-shelf general purpose AI lacks today is that explainability. When you were in seventh grade and your math teacher asked you to show your work, it’s that level of detail-orientedness that platforms like ours provide with no regrets. In order to build conviction and trust within the institution, and to turn around and produce that as evidence for a regulator on how a loan decision was made, that is something regulators will need regardless of specific guidance passed down. So in our view, adopting AI that is explainable has no regrets. Hopefully that helps.
Sarah Snell Cooke: Yeah, for sure. I always showed my work. Whether the teacher could read it or not was their problem. You mentioned about not having humans in the loop. Where is a case where it’s better or not to have a human in the loop with AI work?
Matt Creatore: The moments at which deep nuance is required. We like to say there are so many workflows credit unions go through on a regular basis. I have dozens of systems. I need to search and retrieve for data across a few or several of those systems, and then I need to stare and compare those data that I pull against my policies and procedures. That takes a lot of manual time for operators within credit unions today. That is something that an agent is very good at facilitating. The moments at which judgment is required, that’s where a human should remain in the loop. Have AI work for you to pull the information, to flag things that require attention, and then have the human recognize—not have to spend the time pulling the data and inferring those insights necessarily—to understand if something the AI is flagging should be adjusted. Let’s just stick with the theme of a loan decision. It’s not just should the applicant be approved. Because we in financial services are in the business of pricing risk, it’s understanding what key insights have been flagged by the AI to recommend an interest rate adjustment or take a specific personalized action for underwriting a specific applicant. Those are the moments where an underwriter is uniquely positioned to be in the loop to make a final decision on what the borrower should get in terms of their terms. That’s one example of what ends up being hundreds of examples where humans are best positioned to be in the loop, when they have to apply their judgment and deep expertise.
Sarah Snell Cooke: Makes sense. General purpose AI wasn’t built for banking governance specific to a highly regulated environment like banks or credit unions. What is the advantage of having one built specifically for banking versus general purpose AI?
Matt Creatore: It’s a great question. I would say three different nuances. The first is what I already mentioned, which is very important from a regulatory perspective. Large language models are upwards of hundreds of billions or a trillion parameters. There is so much complex math involved in the inference of these models, essentially what outputs the models produce, that makes them very difficult to reverse engineer for the explainability I was talking about earlier. To actually show the work and the chain of thought. Banking AI is not just about the responses; it’s about the ability to come back to an output a year or two later and have reliably the actual explainability of that response because model weights are changing constantly in large language models. The objective for these general purpose large language models is artificial general intelligence. They are trained on an internet’s worth of data, which means it becomes more difficult when your user experience is to produce an answer as quickly as possible while traversing that amount of data to actually have acute attention to detail on the things that actually matter in financial services. For example, for a credit union working with AI to determine compliant marketing copy for a specific type of product, you or any of our clients may find that the response we provide with our small language banking models is 10 words different than the large language models. However, it’s those 10 words that keep you out of trouble with the regulators in terms of really understanding deeply where the lines of compliance and the letter of the law are. Hopefully that helps answer the question, which is getting very focused on what actually matters in this vertical versus applying a general model to something that is highly regulated where you must provide the right responses, not just generally accurate responses.
Sarah Snell Cooke: Makes perfect sense. Regulators can be a little nitpicky about following the rules. Why is explainability becoming a competitive advantage for financial institutions?
Matt Creatore: Competitive advantage because they can innovate with speed and confidence. If I know that the AI platform I’ve chosen as a financial institution creates this level of explainability, then regardless of the way the regulatory environment evolves—because guidance will come from regulators that is more specific around what governance must be put in place depending on institution size, and what they must see from the adopted AI—if I know I’m choosing a platform that can help me produce the things a regulator is looking for, then I can start worrying less about that and worrying more about how fast I am building things that create really high business value. It can save resource time and allow people to do the things they’re best positioned to do. That’s how I would summarize it.
Sarah Snell Cooke: One of the things I think is really important, as you mentioned, is time savings. While having a chatbot is fun and easy for members and available 24/7, back-office work done with AI is infinitely more valuable. Can you talk a little about which way credit unions should go first if they are about to pull the trigger? What should they be working on first?
Matt Creatore: There are so many use cases that are hair on fire at the financial institutions we engage with, which is great because there are a lot of areas of opportunity. There are three primary functions where first use cases tend to emerge: growth, back-office operations, and compliance operations. How do I create operating leverage with the credit and underwriting team I have in order to produce, underwrite, and originate more loans and deposits? The second is back-office operations like QA and QC. The third is compliance operations, for example, BSA/AML and third-party risk management. If you tie it back to a theme we discussed earlier, if you look under the hood at what these use cases are doing in third-party risk, I have an encryption type stated in the software development life cycle of a vendor, but it’s not checked in the SOC 2. To connect those two dots takes someone on a vendor management team hours potentially to come to that conclusion, and that’s something AI can pull forward very quickly so the vendor manager can actually understand if this is a concern based on a policy or procedure that the institution has. Those are several of the use cases where we tend to see the highest readiness and value at an institution. That last example applies to all use cases, which is search, retrieve, stare, and compare work. This is manual work people are spending time on now that they don’t have to.
Sarah Snell Cooke: I love the stare and compare phrase. It’s exactly right. A human wants to check if this field is filled and if it matched this document they provided. There’s also human error that AI eventually will not have. I assume a lot of it’s very accurate, especially, as you said, trained specifically for banking. That said, humans are functioning around this and making things happen. How do you bring your team along to become comfortable with using AI, to be comfortable that their job isn’t being taken over? How do we deal with the humans in this equation?
Matt Creatore: It’s a great question. It’s something we deal with at nearly every institution we engage with. It’s a matter of enablement, helping them see that what AI actually will do is not replace humans. It will create operating leverage. It will allow you as a human being to apply your unique judgment and skill set with a higher volume because I’m not having to do as much manual work like checking if a vendor software development life cycle encryption type is checked in the SOC. I don’t have to do those things anymore, or in a decreasing amount as you’re rolling out more agents. As the market in general and an employee base at a credit union see that and trust that is the direction AI is taking us all in, that will start to provide a tailwind for people’s propensity to adopt and lean in. The other thing that can’t be understated is to hear from a leader of the institution that message: “We are using AI to upskill you, the people that work at the institution, not replace you.”
Sarah Snell Cooke: I think just saying those things out loud can really help the organization gain the level of conviction it needs to accelerate the adoption process.
Matt Creatore: That’s how I would maybe answer the question.
Sarah Snell Cooke: Yeah, it makes the humans more valuable.
Matt Creatore: 100%.
Sarah Snell Cooke: What about the board members? Credit union board members tend to be a little older, and sometimes maybe not up to speed—and I’m making generalities to my favorite credit union people, not talking about everyone. Having them get up to speed, get comfortable with AI, and then set governance policies… What should a board be asking? What should they know?
Matt Creatore: I think it’s a good question. I’m not sure that I have the perfect answer for it, but the more people live in the day-to-day of using AI, the more they will appreciate the thought processes or patterns that start to get unlocked once you use it. Adopting AI is not just a SaaS platform that you as an institution are starting to use; it’s a transformation, an operational change in everyone’s day-to-day. They have to think differently. For example, speaking from my own experience, in the past I would have to go to a Teams thread, a Word document, or an email where all the context lives spread out over an institution to form in my own brain how I’m going to respond to something or make a decision. If you now have the approach or standard of just going to those sources and feeding it to AI or integrating that context into an AI platform, now you can start to think through: “Why would I go source that information? All I have to do is just ask the AI for that specific information or paste links into the AI.” It sounds so simple in theory, but going through that change in your mind of “It doesn’t have to work the way it has worked for me for the last 10 or 20 years,” there’s so much time and repetition required to get your brain to think like that. The reason it’s relevant to the board question is because the more board members themselves are leveraging this technology, the more I think they’ll be in a better position to advise if needed, and to appreciate what good industry-specific AI looks like versus AI that won’t fit a highly regulated space.
Sarah Snell Cooke: I would think too, given that data is contained inside the institution as I’m understanding, or it’s banking specific, the accuracy, security, and all that kind of information seems to be more accurate and secure. Did I make sense there? I don’t know.
Matt Creatore: You did. The comment I’ll make on it is when you have secure AI—and this is a point I didn’t hit on earlier—when you have secure AI that gives the institution conviction that no sensitive data or PII is ever getting out to any large language model, so there’s no risk there, what it unlocks is a lot of trust that it doesn’t matter what data I’m feeding it, it’s protected and going to stay safe. That gives people more momentum behind: “Oh, I could use it for this, this, and this.” That’s where change management really starts to accelerate, when you have people saying, “I used it for this one use case, but actually these other 25 use cases are the same thing, just in a different part of my workflow.”
Sarah Snell Cooke: Absolutely, yes. I always allow my guests to have final thoughts, and we are at time here. What would you like to leave our credit union audience with today?
Matt Creatore: Choosing the right partner in this journey is going to be critical because of the level of nuance required from good AI. So it’s not just models built for a specific vertical that are being considered; it’s also making sure all of your institutional knowledge and context is considered or invoked before the models actually do the thinking, depending on what you’re prompting the AI for. To create that whole AI prompt-to-response flow in the right way, financial institutions are landing on the fact that it’s not their core competency. To choose the right partner that can help with all of the considerations financial institutions have, that is going to separate over the next, let’s say, three years, the winners and the losers.
Sarah Snell Cooke: Absolutely. Love that. Thank you so much for your time and your expertise today, Matt.
Matt Creatore: Likewise. Thanks for having me, Sarah.