How Adarsh Naidu is trying to stop your card from being wrongly declined

Opinions expressed by Digital Journal contributors are their own.

You are at a checkout or watching a payment spin on your phone, and the transaction fails. Nothing is wrong with your account, and the purchase is ordinary. Somewhere in a fraud system, though, your card was flagged as a risk, and the payment was stopped. For the customer, the moment is small and infuriating. For the bank that declined a paying customer, the cost is real, too, even if it never shows up on a statement.

Adarsh Naidu has spent much of his career on the machinery behind that moment. A Senior Solutions Architect in the Enterprise Banking group at Amazon Web Services, he worked for more than two decades across financial services firms, including American Express and Assurant, before joining AWS. Much of that time has gone into a problem people notice only when it inconveniences them: fraud systems that stop the wrong payments.

The false alarm nobody sees

Fraud detection is often described as catching criminals. In practice, a large part of the daily work is the opposite, deciding which of millions of normal purchases to let through. A system tuned to catch every possible fraud will also block a lot of legitimate spending. A system tuned to wave everything through will miss the real thing. Every bank lives somewhere on that line, and where it sits decides how often ordinary customers get told no.

The wrongly declined transaction is what the industry calls a false positive, and it is expensive in a way that is easy to underestimate. The bank loses the sale, the customer feels distrusted, and the relationship takes a small hit that adds up over time. A shopper who declined once at a bad moment may reach for a different card next time.

“A false decline is invisible on a balance sheet, but the customer feels it instantly. Getting fraud detection right is really about not punishing the honest transaction,” says Naidu.

The difficulty is that the events banks most need to catch are rare, and the data describing them is scarce and tightly guarded. A model trained mostly on normal behavior can look convincing in testing and still stumble when a genuinely new fraud pattern arrives. That gap between the tidy test and the messy real world is where honest customers get caught.

Teaching systems on data that isn’t real

Naidu’s answer, developed over years of work, leans on a technique that sounds counterintuitive: train fraud systems on data that was never real to begin with. At American Express, he says he worked on applying Generative Adversarial Networks, a type of AI more commonly associated with generating images, to fraud and dispute handling. The idea is to learn the shape of genuine transaction patterns and then produce synthetic examples that carry the useful structure without exposing any real customers’ records.

That matters for the false-decline problem in two ways. It lets a fraud system rehearse against a wider range of scenarios, including the rare cases ordinary data barely contains, so it is less likely to panic at something merely unfamiliar. And it does this without pulling live customer information into every experiment, which keeps privacy intact. A system that has already seen many lifelike variations of fraud and legitimate spending is better positioned to tell them apart when it counts.

Naidu applied the same approach to dispute handling, setting it out in 2024 in a reference piece for the AWS for Industries blog on dispute management in banking. His peer-reviewed research runs along the same lines, covering privacy-preserving synthetic data and fraud feature engineering, the work of deciding which signals actually separate a real problem from a harmless anomaly.

Building for the customer, not just the model

None of this removes the tension at the heart of fraud detection. Banks will always weigh the cost of missing fraud against the cost of blocking good customers, and no model can entirely eliminate that trade-off. What better training data offers is a way to move the line, catching more genuine threats while letting more honest transactions through.

There is a regulatory current pushing in the same direction. Bank oversight in the United States has been moving toward a clear expectation that automated decisions affecting customers be explainable and leave a record, raising the bar for anything put into live use. Naidu has responded by authoring a Responsible AI framework for AWS that builds governance into each stage of system development rather than adding it at the end.

For the person at the checkout, the argument is simple. The goal is a system that stops fraud while still allowing legitimate payments through. Naidu has spent the better part of two decades trying to close the gap between those outcomes, and the work is far from finished.