CUSTOMER Q&A: Purpose Financial

How Purpose Financial implements continuous decisioning at scale

For Purpose Financial, continuous risk decisioning sits at the intersection of two critical questions: Will this customer repay? And are they who they say they are? Answering those questions quickly and accurately is an essential component of the company’s digital products and online experiences. But as customer expectations around speed continue to rise, traditional approaches to credit and fraud decisioning are becoming harder to rely on. 

In this interview from Plaid Effects, Pedro Sanzovo, Head of Fraud and Identity at Plaid, sat down with Dennis Kramer, Director of Credit Risk and Data Science at Purpose Financial, to discuss how richer consumer-permissioned financial data and identity signals can unlock a continuous decisioning framework in line with the real world.

This conversation has been edited for clarity and length.

Sanzovo: To start, tell us about your role at Purpose. What are the big questions your team is trying to answer every day?

Kramer: My team sits in a really interesting place because we’re focused on two questions that are often treated independently. The first is: Will this customer repay? Do they have the ability and willingness to repay? The second is: Is this customer who they say they are? Across our branches, online channels, and digital products, we have to answer both in real time every day. And we’re often answering them with data that’s historical—sometimes a day old, sometimes six months old.

The thing is, on some level, the math is solvable; I’m not sure that’s the main problem. Fraud and credit risk are inherently behavioral problems. So how do we understand customer behavior, fraudsters’ behavior, and model that in a way that helps us identify fraud rings or systematic over-borrowers while still moving the business forward?

Sanzovo: Has continuous decisioning become more important for Purpose?

Kramer: Very clearly, yes. Two things have happened at the same time. First, customer expectations around speed have collapsed. Five or 10 years ago, instant approval or instant funding really stood out. Now, that’s the expectation. Second, we’re seeing rapid growth in fraud: systematic identity takeovers, coordinated fraud rings, and fraudsters adapting quickly. That forces us to be agile in how we identify and respond to risk. We’re in an environment where batch scoring is no longer the most effective way to do that.

Sanzovo: Let’s talk about data. When you’re trying to make decisions continuously, what actually matters?

Kramer: There’s been a big shift from single points in time to a broader network of information—signals across the network. We used to ask a simple question: Does this identity check out? Yes or no. Now we’re asking: How does this identity act across the network? Are there multiple accounts across devices? Is there underlying behavior we would expect from an authentic customer?

The same is true on the credit side. Historically, bureau data gave us a single point in time. It’s still important, but we think about it as a historical snapshot. What we want to understand now is trajectory. How do we use payment information, transaction insights, bank data, and network insights to tell us not just where the customer has been, but where they’re likely to go? That’s one of the things I’m most proud of in our work at Purpose. It’s the difference between looking at a picture and understanding the movie.

Sanzovo: That moves beyond pure identity data or static credit data. How do teams need to think differently?

Kramer: Within a continuous decisioning framework, the teams involved are really distinct. My fraud team is reactive and investigative. My credit risk team is analytical and policy-driven. The data science team is focused on models and what those models predict. As we think about the next generation of continuous decisioning, we need all three. We need people who can think with an investigative mindset and also understand how to integrate that thinking into models.

I don’t just need someone who can build the next gradient-boosted model or use SQL to monitor something in production. I need someone who can navigate the technical details and talk to senior leadership about why a cut was made, why a policy changed, or why a dynamic decision was made—and put that in context.

Sanzovo: You also think about this as a behavioral science problem. How does that show up in credit and fraud?

Kramer: Customers are often likely to overestimate the immediate relief they expect to receive while underestimating the long-term obligations or impacts. We’re wired so that what’s in front of us matters more than the long-term impact. That’s why we talk about current me helping future me. In our models, we have to think about how to account for behavioral outcomes like that overestimation of benefit.

Sanzovo: A lot of companies struggle with alignment, especially around risk appetite. How do you think about getting that right?

Kramer: I truly believe setting risk appetite is an essential senior leadership activity. Not because it’s about passing the buck, but because centralized decisioning helps data science, credit risk, and fraud teams execute on the agenda. If we don’t have a centralized and documented risk appetite, every implementation decision my team makes becomes a negotiation. It’s no longer execution of an agenda; it’s negotiating how much risk we should take. That creates significant barriers to operating at scale.

I think of risk appetite as the top of the waterfall. It helps us understand how to execute with excellence.

Sanzovo: Risk appetite isn’t one-dimensional, right?

Kramer: Exactly. We may be willing to have higher risk in one segment and pull back risk in another. The overall risk appetite can be relatively similar, but we can differentiate within it.

From a data scientist’s perspective, it’s easy to get eyes down on the model and say, “The model predicts this, the model predicts that.” But when leadership says, “Here’s the strategic agenda,” the conversation about where to pull back or push forward on risk becomes much more dynamic.

Sanzovo: Let’s talk about model governance. Should teams retrain models on a schedule, or should something else trigger it?

Kramer: The default across the industry is calendar-based models. You refit, retrain, or redevelop models every quarter. That can feel disciplined. But fraudsters don’t work in quarters. They’re not saying, “I’m going to implement this new fraud strategy in Q2 so Dennis can fit his model to address it.” Macroeconomic conditions don’t change on a quarterly basis either.

We’ve been thinking about governance from an evidence-based trigger perspective. How do we create triggers based on PSI, approval rate shifts, or loss rates moving outside our tolerance band? The goal is to let signals that are deeply connected to the business tell us where something may not be optimal.

Sanzovo: Some of those signals are early indicators, and early indicators can be noisy. How do you account for that?

Kramer: You have to model the uncertainty. Early indicators like first payment default or payment behavior are certainly more volatile than long-run loss outcomes. But again, the math is solvable. The bigger change is the mentality shift. Calendar-based governance feels disciplined. But getting to a more dynamic place should really be the default, because then action is constantly happening.

Sanzovo: If you’re operating that way, what matters most in the tools you use?

Kramer: Two things matter most to me. First is the speed at which we can make adjustments. Can we fine-tune a feature, change a threshold, or update a model quickly? If we identify a needed change but have to wait three, four, or five sprints for the tool to implement it, we’ve missed the window for impact.

The second is connecting what the model says is true with what is actually happening on the ground. How do we build models that are constantly learning from sources of truth? If a model can’t learn from the actual source of truth, it becomes static. Then you lose the ability to detect the outcomes you care about.

Sanzovo: Where does human judgment fit into that?

Kramer: In fraud, we have investigators who adjudicate customers we flag as potential fraud. They get a deep sense of data, context, and information, and they make high-quality decisions.

The question is: How do we capture that high-quality context and put it into a quantitative model? How do we quantify it in a way that allows us to rely less on investigators over time? That’s something we’re constantly thinking about, and I think AI provides an opportunity to support that.

Sanzovo: One of my takeaways is that continuous decisioning is not just about building a really cool model. It’s an operating model—data, people, processes, and tools all working together. Does that reflect how you think about it?

Kramer: Yes. Building a good model is part of it, but it’s not enough on its own. You need the right data, the right people, clear processes, and tools that allow teams to move quickly. You also need leadership alignment on risk appetite, so teams know where to push forward and where to pull back. The goal is a framework that can keep learning as customer behavior, fraud patterns, and business conditions change.