An inside look at the technology behind Tally fraud, and why proactive, platform-integrated fraud prevention matters for today’s loyalty programs.

As loyalty programs grow in value, they also become increasingly attractive targets for fraud. From manipulating promotional earning opportunities to exploiting redemption processes, fraudulent activity can quickly translate into lost program value. Yet many loyalty programs still heavily rely on manual investigation to identify and address suspicious activity, oftentimes after points have already been extracted.

Phaedon’s new Tally fraud model takes a different approach. Built directly into the Tally loyalty platform, the machine learning-powered module analyzes member activity and transaction history to identify suspicious redemption behavior and prevent fraudulent transactions before points leave the program.

Phaedon’s VP of Data Science & AI, Brady Begeman, discusses the approach behind building Tally’s fraud model, what makes it different, and how brands can use machine learning to make their existing fraud teams more effective.

Q: What was Phaedon’s approach to designing the Tally fraud model and why did you go in this particular direction?

A: The first thing we wanted to solve for was flexibility.

Our clients operate very different businesses, and their loyalty programs are structured differently. A traditional machine learning model typically relies on predefined data structures, meaning we’d have to build different data models to accommodate each client’s unique implementation. Instead, we developed a sequence model using a transformer architecture. This allows us to analyze up to 31 prior events leading up to a redemption, even when those events have different structures or represent different types of activity.

That’s important because fraud isn’t always about a single suspicious transaction. It’s often about the sequence of events leading up to it. How did someone earn their points? What promotions did they participate in? How quickly did they accumulate and attempt to redeem those points?

We also encode the amount of time between events, because timing can be a significant indicator of fraud. For example, someone attempting to extract points through dozens of small, evenly spaced redemptions may exhibit a very different pattern than a legitimate member making an occasional redemption.

Ultimately, this approach gives us more flexibility to accommodate different loyalty program structures while capturing more information about member behavior. It also delivered a substantial improvement in model performance and helped us avoid overfitting compared with our traditional machine learning baseline.

Q: What makes Tally’s fraud model different from other fraud prevention solutions on the market?

A: The biggest differentiator is the ability to proactively interrupt a redemption transaction before points are actually extracted from the system.

Many fraud solutions focus on identifying suspicious activity after it happens or protecting against account takeover. What we’ve built allows the fraud model to evaluate an individual redemption using the member’s activity history and interrupt that transaction if it meets the configured criteria for fraud.

But it’s important to understand that you wouldn’t necessarily start blocking transactions on the first day you turn on the model. Part of the onboarding involves a feedback process where the client’s fraud team reviews transactions the model identifies as potentially fraudulent. They can confirm whether each transaction is fraud or legitimate, and that information becomes part of the data we use to retrain the model.

That feedback is especially important because loyalty programs often have fraud that has never been identified. You might have two transactions that look nearly identical, but one has been investigated and labeled as fraud while the other has gone unnoticed. The model doesn’t inherently know that both are fraudulent until someone investigates and confirms it. Once we have enough validated information and reach the level of precision a client is comfortable with, we can begin automatically blocking transactions based on configurable risk thresholds.

Q: What can an integrated fraud solution do that a standalone fraud tool might struggle with?

A: The biggest advantage is having access to the actual loyalty program event stream.

A standalone fraud solution might focus on account takeover. It can look at things like changes to a member’s email address or password, unfamiliar devices, or suspicious login activity. Those are useful indicators, but they don’t necessarily tell you what’s happening inside the loyalty program itself.

Loyalty fraud can be much more complicated. You need to understand not only how points are redeemed, but also how those points were earned in the first place. For example, someone might use stolen credit cards to make a large number of hotel bookings, collect promotional bonuses, and then attempt to redeem those points before the original transactions are reversed. To identify that type of activity, you need visibility into the events leading up to the redemption. Because Tally’s fraud capability is connected directly to the loyalty platform, it can analyze that history and identify patterns that a solution operating outside the program wouldn’t see.

Q: How does the model adapt as fraud tactics evolve?

A: One of the biggest parts of the value proposition is our automated retraining process.

Even after automatic transaction blocking is enabled, the feedback process doesn’t stop. Fraud teams can continue reviewing transactions that are flagged by the model, including those that score highly for fraud but fall below the threshold for automatic blocking. For example, a brand might choose to block transactions with a 90% fraud probability while still reviewing transactions that score above 80%. Those thresholds don’t have to be the same. As the fraud team identifies new patterns and confirms fraudulent activity, that information feeds back into the model. Over time, the model becomes better at recognizing those behaviors.

We’re not just trying to train the model to recognize a handful of very specific fraud scenarios. We’re trying to help it understand what fraudulent activity generally looks like inside a loyalty program. That gives it a foundation for identifying evolving patterns rather than relying exclusively on previously documented cases.

Q: Where do you think the loyalty industry gets fraud prevention wrong today?

A: For the most part, it’s still a manual process.

Many programs have a fraud desk that investigates suspicious transactions when they’re reported by a call center or flagged by someone managing the loyalty program. The team determines whether fraud occurred, then takes action by clawing back points or suspending an account. The problem is that this doesn’t necessarily prevent the same type of fraud from happening again. There’s no system automatically applying what the team learned to future transactions. They still have to identify and investigate that behavior each time it occurs. That’s where machine learning makes a difference. It allows you to interrupt a transaction based on its risk before the points leave the program.

Banking and financial services have been doing this for years, using risk models to stop suspicious transactions before they’re completed. Loyalty programs have a more complicated set of data to work with, though. They involve promotions, bonuses, different earning structures, and all kinds of member activity beyond simple credits and debits. That’s why loyalty fraud requires a solution designed specifically around the way these programs operate.

Q: What should loyalty leaders understand when considering Tally for fraud prevention?

A: The biggest differentiator is making your fraud team more useful.

Historically, fraud teams spend a lot of time handling activity after the value has already been extracted. They’re investigating what happened, closing accounts, and trying to prevent it from happening again. Tally’s fraud model changes that process. It evaluates the member’s history at the time of redemption, identifies suspicious activity, and can block transactions before points leave the program.

It also helps fraud teams prioritize their work by surfacing transactions with the highest likelihood of being fraudulent. When the team confirms a particular scenario as fraud, that information helps train the model to recognize and automatically prevent similar activity in the future. The goal isn’t to replace the fraud team. It’s to eliminate the repetitive work of identifying the same types of fraud over and over again.

If we can proactively prevent those points from leaving in the first place, it saves a lot of time and money. And that’s really what drives the model forward: people identifying and confirming fraud versus legitimate activity. That’s how the model gets better.

Bringing a More Proactive Approach to Loyalty Fraud

As loyalty programs become increasingly valuable to both brands and members, protecting the points economy requires more than simply reacting to suspicious activity. Tally now brings machine learning directly into the loyalty transaction process, combining visibility into member behavior, proactive transaction intervention, and human-validated feedback to help brands identify and prevent fraud more effectively. By giving fraud teams the tools to act before value leaves the program, Tally turns fraud prevention from a reactive investigation process into a proactive, continuously improving capability.