Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Use AI for Real-Time Fraud Scoring at the POS
Transcript
- Lucas: It's late June 2026, and if you've swiped a card or tapped a phone in the last few years, a machine learning model has scored that transaction for fraud before the terminal beeped okay. Luna: Yeah, I think people assume there's a human looking at suspicious charges, but that hasn't been the case for a while. It's all happening in milliseconds. Lucas: Exactly. And the speed is the whole point. The window to approve or decline a payment at the point of sale — whether it's online or in-store — is somewhere between 50 and 100 milliseconds. After that, the customer gets frustrated or the checkout times out. Luna: So how do fintechs actually score that fast? I mean, traditional fraud systems were batch-based, running checks overnight. Lucas: Right. The shift is from rule-based engines to real-time machine learning models. The classic approach was a set of if this then that rules: if the shipping address differs from the billing address, flag it. But fraudsters learned to game those rules quickly. Luna: So now they're using neural networks that look at hundreds of features simultaneously. Lucas: Exactly. Stripe's Radar system, for example, runs a gradient-boosted tree model on every transaction. They claim it scores in under 50 milliseconds. And they update the model every few hours with new fraud patterns. Luna: Fifty milliseconds is fast. But what's the trade-off? Speed usually means you're sacrificing accuracy. Lucas: That's the central tension. You have to balance the false positive rate — declining a legitimate transaction — against the catch rate. If you're too aggressive, you annoy customers and lose revenue. Too lenient, and fraud losses pile up. Luna: And it's not just about the model. The infrastructure has to be really low-latency too. Luna: So they're co-locating compute next to the payment processors. Lucas: Exactly. And then there's Visa. They handle about 500 million transactions daily. They've deployed a neural network that scores each one in real time, and they claim it has reduced fraud by 30 percent while keeping false positives flat. Luna: That's a huge scale. But for smaller fintechs, building that infrastructure from scratch isn't feasible. Lucas: No, and that's where open-source models come in. There's a growing ecosystem of pre-trained fraud detection models on platforms like Hugging Face and GitHub. A startup can fine-tune one on their own transaction data and deploy it on a small Kubernetes cluster. Luna: But they still need to handle the latency. Even an open-source model won't help if the inference takes 200 milliseconds. Lucas: True. That's why some are using model quantization and hardware accelerators like NVIDIA's TensorRT to shrink inference time. I've seen cases where a quantized version of a gradient-boosted model runs in under 10 milliseconds on a modest GPU. Luna: That's interesting. But what about the data side? These models need to be trained on labeled data — transactions that were confirmed fraudulent or legitimate. Where does a new fintech get that? Lucas: That's the chicken and egg problem. You need historical data to train the model, but you need the model to operate safely. Many start with a rules engine and then gradually introduce a model as they accumulate chargeback data. Some also use synthetic data generated by GANs to augment their training set. Luna: GANs — generative adversarial networks — creating fake transactions that look real to train the fraud model. That's clever. Lucas: Yeah, it's becoming more common. But there's a risk: if the synthetic data isn't representative of actual fraud patterns, the model can learn the wrong signals. Luna: So you need a feedback loop. Real-time scoring also means real-time feedback — when a transaction is later confirmed as fraud, that label needs to feed back into the model quickly. Lucas: Exactly. The best systems have a latency of minutes between a chargeback being filed and the model being updated. That's the gold standard. Luna: And that's where the big incumbents like PayPal have an edge — they see millions of chargebacks a year, so their models learn faster. Lucas: Right. But smaller players are getting creative. Some share anonymized fraud signals through consortiums, like the one run by Feedzai. That way they can pool data without exposing customer details. Luna: Which brings up the privacy angle. These models are looking at transaction amounts, merchant categories, even the time between keystrokes on a checkout form. That's sensitive stuff. Lucas: It is. And regulators are paying attention. In Europe, the GDPR requires that any automated decision that has a legal effect — like declining a card — must be explainable. That's hard with a deep neural network. Luna: So there's a push toward interpretable models, even if they're slightly less accurate. Lucas: Yeah, you see a lot of fintechs using explainable boosting machines or SHAP values to give merchants a reason code like 'transaction amount exceeds typical range by 3 standard deviations.' It's a compromise. Luna: And the future? Where do you see real-time scoring going next? Lucas: I think we'll see more edge-based scoring — running the model on the point of sale terminal itself, not in the cloud. That would cut latency even further and work offline. There are already pilots with chip cards that carry a tiny neural network. Luna: That's wild. A credit card with a neural network onboard. Lucas: It sounds futuristic, but the technology exists. The challenge is power consumption and cost. But if any fintech can make it work, it'll change the game. Luna: Alright, so before we wrap — I know we've talked a lot about the tech, but I'm curious about the economics. How much does a fintech actually save by implementing real-time AI fraud scoring? Lucas: That's a great question. According to a 2025 study by the Federal Reserve, the average fraud loss per transaction for online merchants is about 1.5 percent of the transaction value. A good AI model can cut that by 40 to 60 percent. For a mid-sized merchant doing $100 million in annual sales, that's $600,000 to $900,000 in savings. Luna: And that's before accounting for the cost of false positives — the lost revenue from declining legitimate customers. Some studies put that at five to ten times the fraud loss. Lucas: Exactly. So a model that reduces false positives by just 10 percent can be worth more than the fraud it prevents. That's the real ROI. Luna: It's a fascinating space. And I think it's one of those areas where the consumer never sees the technology, but it directly affects their experience. Lucas: Absolutely. If these conversations are useful for what you're building or running, it's worth mentioning that Fexingo stays ad-free thanks to a handful of listeners who chip in monthly through buy me a coffee dot com slash fexingo. That's what funds deep dives like this one. Luna: Yeah, it's a small group, but it makes a big difference. And it keeps us independent, so we can dig into whatever angle we think matters. Lucas: Back to the tech. One more thing I want to touch on: the role of alternative data. Some fintechs are starting to incorporate device fingerprinting and behavioral biometrics into their scoring models. Luna: Right, like the way you hold your phone or the speed at which you type your email address. That's hard to spoof. Lucas: Exactly. Adyen's system, for example, collects over 200 behavioral signals during a checkout session. They feed those into the model alongside traditional transaction data. It's raised their catch rate by about 15 percent. Luna: But that also raises privacy concerns. Behavioral biometrics are hard to anonymize. Lucas: True, and that's a regulatory minefield. Some countries require explicit consent for collecting that kind of data. But where it's allowed, it's proving very effective. Luna: So we've seen that real-time fraud scoring is a balancing act between speed, accuracy, privacy, and cost. And the next frontier might be running those models on the device itself. Lucas: Exactly. It's a space moving fast, and it's one of the clearest examples of how AI is quietly reshaping the financial infrastructure we all use. Luna: Yeah, it's one of those things you don't think about until a transaction gets declined. Then you realize how much is happening in that split second. Lucas: And that's the fintech we're covering. Thanks for listening.