Matthew Pearce, vice president of Fraud Risk Management and Dispute Operations at i2c, says the collision of innovation, tightening oversight and real-time payments is creating a “perfect storm” that few financial institutions are fully prepared to navigate. Read more in this PYMNTS interview.
Transcript
This is What's Next in Payments, a PYMNTS Podcast. Forward-looking insights from industry leaders on the trends and technologies reshaping payments and fintech. In this episode, i2c's Matthew Pearce discusses how it's possible to redefine what fraud resilience means in financial services while delivering performance.
John Gaffney:Hi everyone, and welcome to PYMNTS TV, where we're tracking AI and payments as it moves from promise to performance. Matthew Pearce, Vice President of Fraud Risk Management and Dispute Operations at i2c, is here to share how his company is using Gen AI and predictive AI to transform how financial institutions manage risk, credit, and customer operations. Matt's going to discuss innovations that help clients resolve complex technical questions in real time, reducing friction and operational costs. Behind these advancements at i2c is a disciplined approach to augmented intelligence, where more than 50 AI specialists and QA experts ensure every model's output is explainable, secure, and regulator ready. The result is a blueprint for how banks and fintechs are going to harness AI to compete smarter in the era of intelligent money movement. Matthew Pearce, welcome.
Matthew Pearce:Thank you. It's good to be here.
John Gaffney:Good to see you again. So let's talk about the I I don't know. When I hear the word audit, it kind of strikes fear into my heart personally. But not that we're in trouble there at all. But what do banks need to know about AI and audits, not just A-B tests? And how does i2c define the concept of explainable enough across different use cases?
Matthew Pearce:Yeah, you bet. And you know, uh whenever you talk audits and regulatory ability to track things, uh it becomes a little, everybody gets a little nervous around that. So for us, explainable enough uh depends on the risk of the decision. Uh in credit or fraud, really every outcome has to be traceable from the features and rules behind it to the business impact that it creates. So we build explainability into the model and into the model lifecycle. It's not an afterthought, it's actually part of the process. That means every model has documentation, versioning, and fairness testing baked in. When regulators or clients ask why a decision happened, we can show them the full story. Uh, data, data lineage, rationale, and governance all aligned to tell that full story.
John Gaffney:Interesting. So, my understanding, Matthew, is that training and inference have to those models have to protect issuer and cardholder data. What data sources power your predictive and generative models? And how do you ensure that that data quality doesn't overfit any one client?
Matthew Pearce:Yeah, you bet. We draw insights from a broad mix of transaction data, data dispute dispute outcomes, and behavioral patterns as well, always with strong validation and lineage controls. So each data set goes through schema checks, drift tracking, and challenger testing before a model moves into production. Our approach really is federated by design. Models learn from global trends but adapt locally to each client's environment. And that lets us maintain performance accuracy without overfitting or biasing the model to a single portfolio.
John Gaffney:So, Matthew, I'm interested to know how you can keep customer data out or sensitive data out of a model if you're moving a lot of data into it. Could you help me explain that a little bit?
Matthew Pearce:Yeah, you bet, that's a great question. So privacy protection really begins upstream. So personal identifiable information really never enters into the training pipelines. We tokenize or we use hash identifiers and separate PII from analytical data at the architectural level. So models only ever see attributes relevant to prediction, not the customers actually behind them. And so when we generate explanations, they come from structured metadata and not raw data or personal details. It's about building really privacy into the process so that transparency never becomes the cost or that never becomes at the cost of security, I should say.
John Gaffney:Right. Interesting. So um we we know we're dealing in in a situation in payments. I mean, they're they're won or lost in milliseconds. So the AI has to be accurate and fast. But how do you balance what we're going to call catch rate with customer friction? What metric should matter most here?
Matthew Pearce:Yeah, great, great question. I think everybody's always trying to figure out how do you walk that fine balance between that friction and catching fraud. And it is a balancing act. That success really is a balancing act. The value of catching fraud really is minimized if legitimate customers are caught in the net. So leading institutions measure performance across multiple dimensions, and they tune models then continuously to maintain that equilibrium. Some of those performance metrics include capture rate, false positive rates, approval lift, loss savings, and dispute cycle times. But for me, the critical few that I really look at to balance that success is really my fraud loss ratio, my fraud decline rate, and false positive rate. Those are really the critical few that I look at.
John Gaffney:And we know, Matthew, that fraudsters are going to fraud, right? So how quickly can your models adapt to new attack patterns, new attack vectors?
Matthew Pearce:Yeah, absolutely. So adaptability matters as much as accuracy. So modern defense blends real-time anomaly detection with controlled retraining cycles. So early warning models surface suspicious patterns in hours, and feature gating and canary deployment really tests the responses safely before broad rollout. So you kind of dip your foot in the pool, if you will, before you dive in. This layered approach really prevents overcorrection and it keeps, you know, keeping the fraud catch rate high while the stability for existing portfolios uh stays stable. Agility without volatility really is the new definition of resilience.
John Gaffney:So i2c has done a great job of getting its it its best practices, its um its innovations out to the marketplace. So let's talk about your secret sauce a little bit, which is the human, the promise of AI plus human in workflow design. What does human in the loop look like for fraud analysts and dispute agents? And where do they hand off? How do they know when to hand off a bet.
Matthew Pearce:Great question. So human in the loop really isn't a it's not a fail-safe for us. It's actually a system design. So low risk or high confidence decisions are automated while uncertain cases route to analysts with full context. So that means key features, historical outcomes, and then policy notes for them to review those edge cases, if you will. Escalation triggers are clearly defined around ambiguity, compliance obligations, or potential systemic risk. And analysts don't just resolve exceptions, they actually generate feedback that sharpens the next model cycle. So the outcome is really a symbiotic workflow where humans provide ethical guardrails and machines really deliver scale.
John Gaffney:So we we we also know, Matthew, that as important as this is, it doesn't come without effort, right? So banks in fintechs need a path from pilot to business impact. They need to be able to execute on this. What does a typical pilot to production journey look like? Um, and how does it integrate with other platforms in the company like CRMs?
Matthew Pearce:Yeah, you bet. So effective AI adoption flows uh really they follow a disciplined 90-day cycle, if you will. Scope and success criteria first, integration and configuration next, and then to follow is really the limited rollout, is kind of that last step. The process, this process really minimizes disruption by connecting through APIs that coexist with legacy cores and CRMs. Client resources stay light. In essence, they have data access, compliance oversight, and a technical liaison, while the provider really shoulders the setup and governance. And really the goal of this uh is not proof of concept, it's really proof of impact. And that's measured in approvals, fraud savings, and faster dispute resolution.
John Gaffney:Wow, proof of impact. I've I haven't heard that before. Is that is that is that a standard metric in your bit in your side of the business?
Matthew Pearce:You know, it's uh it's it's not, but it's it's something where I think we're raising the bar on what proof of concept really means. It's you know, it's really can you can you take it to what the impact of the business is? The concept may be fine, but what's the impact to the business? And that's really for us the key indicator.
John Gaffney:Wow, very well said. Matthew, last question for you. What are the biggest adoption blockers here? What what what what's the block to executing on a structure like the one i2c has?
Matthew Pearce:Yeah, you bet. So the toughest barriers barriers for us are not technical, they're organizational. Governance approvals, you know, data quality and regulatory comfort often slow AI more than coding ever does. The fastest path through it really is transparency. So model documentation aligned with SR117, uh, pre-tested explainability templates, and embedded compliance workflows for Reg EZ or ECOA. And then when governance becomes a design feature instead of an afterthought, adoption accelerates naturally. It really just becomes part of the ecosystem.
John Gaffney:Which is part of the flywheel effect, I would think, from the i2c approach with the human in the loop, correct? That's exactly right. Okay. All right, that's gonna do it for this episode of PYMMTS TV. My guest has been Matthew Pearce. He is vice president of fraud risk management and dispute operations at i2c. Matthew, thanks for joining us. Thank you so much.
Narrator:That's it for this episode of the PYMNTS Podcast, the thinking behind the doing. Conversations with the leaders transforming payments, commerce, and the digital economy. Be sure to follow us on Spotify and Apple Podcasts. You can also catch every episode at pymnts.com/podcasts. Thanks for listening.