Latest / The Operations Podcast with Fexingo: Process, People, and Profit in Modern Business / How a Bank Cut Check Fraud by 90 Percent with Machine Learning
Transcript
- Lucas: In 2024, check fraud cost U.S. banks somewhere north of $26 billion — and that's just the reported figure. Luna: And a lot of that is because checks haven't really changed in decades. They're basically paper with signatures. Lucas: Exactly. So when a regional bank in the Midwest — let's call them Lakeshore Bank — saw their check fraud losses hit $12 million in a single year, they decided to stop relying on manual review and instead build a machine learning system that could flag suspect checks in real time. Luna: And they got losses down to under $1.2 million in 18 months. That's a 90 percent drop. Lucas: Right. But the interesting part isn't just the number — it's how they designed the model and the workflow around it. Most banks either reject suspicious checks outright or send them to a back office team that reviews them hours later. Lakeshore wanted something faster. Luna: So they built a gradient boosting classifier. Trained on about 400 features per check — things like the check amount relative to account history, the micro-printing quality, the magnetic ink character recognition numbers, even the time of day the check was deposited. Lucas: And that last feature is key. Most fraudulent checks are deposited between 2 and 4 a.m. through mobile deposit. The model learned that pattern. But the real operational innovation was what they did after the flag. Luna: Instead of rejecting the check and irritating customers, they routed flagged checks to a three-person review team that worked in fifteen-minute shifts. The team had a dashboard showing the top five reasons the model flagged each check. Lucas: That reduced false positives dramatically. In the first month, the model was flagging about 8 percent of all checks — way too many. By tuning the threshold and adding more features, they got that down to 1.2 percent, with a false positive rate under 0.3 percent. Luna: And the review team could clear a legit check in under 30 seconds. So customer friction was minimal. Lucas: They also made a deliberate choice not to use facial recognition or biometrics for signature verification. The bank's head of operations said that would have added too much friction for elderly customers who still write a lot of checks. Luna: That's a smart operational trade-off. You don't want to design a system that works in theory but alienates your core customer base. Lucas: The model itself was unusual because it didn't just look at the check image. It also incorporated metadata from the deposit channel — the device ID, the GPS coordinates, the speed of the deposit motion. Fraudsters often use the same phone for multiple accounts. Luna: So it's not just about the check. It's about the entire context of the transaction. Lucas: Exactly. And once they proved the model worked on checks, they started applying the same approach to wire transfer fraud. That's a much smaller volume but way higher dollar amounts per incident. Luna: I read that wire fraud attempts at Lakeshore dropped by 60 percent in the first six months after they deployed the model. Lucas: Yeah, and one reason is that the model could flag anomalies in real time — like a $200,000 wire to a first-time recipient initiated from a new device. The old system would have let that through until a manual review hours later. Luna: So they saved about $10.8 million in check losses alone. What did the system cost? Lucas: About $400,000 to build and deploy, plus $120,000 a year in cloud compute and the three-person review team. So payback was under two months. Luna: That's a no-brainer for any bank with check volume. Lucas: You'd think so. But surprisingly, only about 15 percent of community banks use machine learning for fraud detection today. Most still rely on rule-based systems — like flagging any check over $10,000 — which miss sophisticated fraud. Luna: And those rule-based systems also generate a ton of false positives. A customer writes a big check for a down payment and suddenly their account gets frozen. Lucas: Right. The Lakeshore approach was more nuanced. They also incorporated something called 'check image entropy' — a measure of the randomness in the pixel data. Counterfeit checks have slightly different entropy patterns because they're printed on consumer-grade printers. Luna: That's a feature most banks don't even know exists. It's brilliant. Lucas: The team spent three months just engineering features before they wrote a single line of model code. And that's a lesson that applies beyond banking — the operations win is in the data prep, not the algorithm. Luna: We talk a lot on this show about operational efficiency, and this is a perfect example of using technology to solve a specific, costly problem without over-engineering it. Lucas: If these conversations have moved your work forward in some small way, it's worth mentioning that we deliberately don't run ads on these episodes. We keep the show ad-free so we can focus on the details that matter. Luna: And if that approach resonates with you, listener support is what makes it possible. You can find that at buy me a coffee dot com slash fexingo. Lucas: So back to Lakeshore — the model is now in its second year, and they've added a feature that checks the payee name against a database of known fraud accounts. That alone caught another $800,000 in attempted fraud last quarter. Luna: And they're sharing the model with other banks in a consortium. That's a rare level of collaboration in the banking world. Lucas: It is. But the consortium means the model gets better as more banks contribute data. They're all fighting the same fraud rings. Luna: So what's next for Lakeshore? Are they looking at other types of fraud? Lucas: Yeah, they're piloting a version for credit card application fraud — using similar metadata features to flag synthetic identities. The same architecture, just different inputs. Luna: It's almost a platform play at this point. Build the pipeline once, adapt the features for each use case. Lucas: Exactly. And that's the operational insight that scales — not the specific model, but the process of continuously engineering features that reflect the real-world behavior of fraud. Luna: It makes you wonder how many other legacy processes in banking are sitting there with 90 percent waste just waiting for the same kind of focused data work.