Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Use AI for Credit Card Transaction Categorization
Transcript
- Lucas: So you open your banking app, and under 'groceries' it shows you spent four hundred and fifty dollars this month on food. Have you ever wondered how the app knows that a charge at 'Fresh Mart' is groceries but a charge at 'Target' is split into multiple categories? Luna: I have, actually. I mean, sometimes I see a transaction labeled as 'entertainment' that was clearly a coffee shop, and I think — how does this work under the hood? Lucas: Yeah, and that's exactly what we're digging into today: the AI behind credit card transaction categorization. It's a deceptively complex problem — one that powers everything from Mint to YNAB to your bank's own spending insights. Luna: It's also one of those features that, when it works well, feels almost invisible. You just assume the app knows your spending. But when it's wrong, it's frustrating. Lucas: And if today actually helps you understand why your app thinks that coffee was entertainment, then you might find value in knowing these systems stay ad-free through listener support. If that's useful, a quick way to keep this going is buy me a coffee dot com slash fexingo. Luna: Yeah, it's a small gesture that helps us keep digging into questions like this without any sponsors. Appreciate anyone who does. Lucas: So let's get into the mechanics. At its core, transaction categorization is a classification problem. You have a stream of raw transaction data: merchant name, amount, date, maybe a merchant category code — MCC — if you're lucky. Luna: And an MCC is that four-digit number assigned by card networks that roughly describes the business type — like 5411 for grocery stores, 5812 for restaurants. Lucas: Exactly. MCCs are the backbone of early categorization systems. But they're coarse. A supermarket with a pharmacy might have an MCC of 5411, so the entire transaction gets tagged as 'groceries' even if half your spend was on aspirin. Luna: Right, so the AI needs to go beyond MCC. It's looking at the transaction description, maybe the merchant's name, and sometimes even the location to get more granular. Lucas: That's where natural language processing — NLP — comes in. Modern fintechs use large language models or more lightweight NLP models to parse the merchant name and description. For example, 'AMZN MKTPLACE' is likely 'Amazon Marketplace' and that could be general merchandise or books or electronics. Luna: But how does it decide? Does it just look at keywords? Lucas: Keywords are part of it, but it's more sophisticated. These models are trained on millions of labeled transactions — where a human has already categorized a sample — so the model learns patterns. 'UBER' plus a certain time pattern might be 'transportation', but 'UBER EATS' is 'food delivery'. Luna: So the model is essentially learning that the word 'EATS' changes the category. That makes sense. Lucas: Exactly. But there's another layer: using the transaction amount and frequency. A five-dollar transaction at Starbucks is probably coffee, not office supplies, even if Starbucks also sells mugs. The model learns probability distributions. Luna: What about the really tricky cases? Like Walmart — you can buy groceries, electronics, clothes, even gas. How does an app know what you actually bought? Lucas: That's the holy grail. Many apps just default to 'general merchandise' or 'shopping' for Walmart because without item-level data, it's a guess. But some fintechs are using receipt scanning or linking to loyalty programs to get line-item detail. Luna: So there's a trade-off between accuracy and privacy. If you want more accurate categorization, you might need to share more data. Lucas: Right, and that's a key tension in the industry. The most accurate systems are the ones that can read your receipt or tap into your purchase history. But that raises privacy concerns — consumers may not want their bank to know they bought diapers and beer at three AM. Luna: And regulators are starting to pay attention. The CFPB has been looking into data aggregation practices, especially around open banking. Lucas: That's a good point. Under Section 1033 of the Dodd-Frank Act, consumers have the right to access their financial data. But how fintechs use that data for categorization is still a bit of a gray area. Luna: So let's talk about accuracy. What's the state of the art? Are we at 99% correct categorization? Lucas: Not quite. In production, most fintechs report around 85 to 95 percent accuracy for major categories like groceries, dining, and transportation. For more granular subcategories — like 'fast food' vs 'fine dining' — it drops to maybe 70 to 80 percent. Luna: That sounds about right from my own experience. I notice misclassifications happen most often at big-box retailers or gas stations that also have convenience stores. Lucas: And those edge cases are where human correction still plays a role. Most apps let you recategorize a transaction, and that feedback loop is crucial. When you mark something as 'groceries' instead of 'shopping', the model can learn from that correction. Luna: So it's a continuous learning system. The more people use it and correct it, the smarter it gets. Lucas: Exactly. And that's why apps like Mint and YNAB have been around for so long — they have huge datasets of user corrections that improve their models over time. Luna: What about newer entrants? I know some neobanks are building categorization directly into their apps now. Lucas: Yeah, companies like Chime and Current are investing in this. But interestingly, some are taking a different approach: instead of trying to guess the category, they're using merchant partnerships to get more structured data. For example, if you buy from a partner merchant, they send the category along with the transaction. Luna: That's a clever workaround. But it only works for a subset of merchants. Lucas: Right, and it creates a two-tier system where categories are accurate for partner merchants but not for others. So nlp based models are still the standard. Luna: Let's zoom out. Why does this matter beyond just convenience? I mean, knowing I spent two hundred dollars on dining out is nice, but what's the real business value? Lucas: Great question. Accurate categorization is the foundation for any kind of personal financial management — budgeting, spending analysis, even credit scoring. If you can't trust that your 'entertainment' category is accurate, your budget is meaningless. Luna: And for banks, it drives engagement. If a customer opens their app and sees a clear breakdown of where their money went, they're more likely to stick around and use that bank as their primary account. Lucas: Exactly. JPMorgan Chase found that customers who use their spending insights feature are significantly more engaged and less likely to churn. So it's not just a nice to have — it's a retention tool. Luna: And it also opens up opportunities for personalized offers. If the bank knows you spend a lot on dining, they can offer you a restaurant cashback deal. Lucas: That's a fine line, though. Personalized offers based on transaction data can feel creepy if not done transparently. But when done right, it's a win-win. Luna: So where is this technology headed? Are we going to see real-time categorization as you swipe? Lucas: Some companies are already doing that. With instant payment systems and push notifications, you can get a categorization alert within seconds of a transaction. But it's tough because the model has less context — it's making a decision with just the merchant name and amount. Luna: And I imagine latency adds cost. Running a model on every transaction in real-time is computationally expensive. Lucas: Exactly. So many fintechs batch categorize overnight and update the app the next morning. That's why sometimes you see yesterday's spending categorized today. Luna: But with edge computing and on-device AI, we might see more real-time processing without sending data to the cloud. Lucas: That's the next frontier — running a lightweight model on your phone that categorizes transactions without ever leaving your device. Apple's approach to on-device intelligence could be a model for that. Luna: It also solves the privacy issue we talked about earlier. The data never leaves your phone. Lucas: Right. And I think that's the direction the industry is heading — more accurate, more private, and more real-time. But we're not there yet for most consumers. Luna: What about small businesses? Do they use similar categorization for their expenses? Lucas: They do, but it's even more complex. A small business might have dozens of categories for tax purposes — office supplies, advertising, travel. And the transaction descriptions are often less standardized. Luna: So the models need to be trained on business-specific data. That's a harder problem with smaller datasets. Lucas: Exactly. But companies like QuickBooks and Xero have been working on it for years, and their accuracy is improving. They use a combination of MCC, merchant name, and even invoice data when available. Luna: So to wrap up, transaction categorization is one of those AI applications that we all rely on daily, but the technology behind it is constantly evolving. Lucas: Yeah, from simple mcc based rules to sophisticated NLP models and even on-device AI. And the end goal is the same: to give people a clear, accurate picture of their spending so they can make better financial decisions. Luna: And maybe, one day, no more misclassified coffee shop transactions. Lucas: That's the dream. Thanks for listening, and if you want to hear more deep dives like this, you know where to find us.