Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Use AI for Real-Time Cross-Border Payment Routing
Transcript
- Lucas: If these conversations are useful for what you're building or running, you're in the right place. I'm Lucas. Luna: And I'm Luna. Today we're looking at something that touches almost every fintech out there — moving money across borders in real time. Lucas: Right. Cross-border payments are still surprisingly fragmented. SWIFT handles something like 15,000 messages per second on peak days, but those messages don't mean the money moves instantly. For smaller fintechs, the challenge is picking the right rail for each payment — and that's where AI comes in. Luna: So we're talking about routing decisions made in milliseconds, not hours. Which company best exemplifies this? Lucas: Wise is probably the canonical example. Formerly TransferWise. They built a proprietary routing engine that uses machine learning to predict the fastest and cheapest path for any given currency pair. They process over ten billion dollars in monthly volume. Their model considers dozens of factors — liquidity in each corridor, time of day, recent FX volatility, even pending transactions in the queue. Luna: Ten billion a month. That's a lot of data to feed an algorithm. How does the actual routing work under the hood? Lucas: The classic analogy is 'matching local bank accounts'. If I send money from London to Berlin, Wise doesn't actually push the wire across borders in the traditional sense. Instead, they take my pounds and deposit them into a local UK account they control. Then, at the same time, their local German account sends euros to the recipient. The AI's job is to figure out the optimal pair of local transfers that clears the fastest with the least FX cost. Luna: So the AI is constantly searching for the best 'match' between incoming funds in one currency and outgoing funds in another. It's almost like a real-time marketplace for liquidity. Lucas: Exactly. And that marketplace is dynamic. A corridor that was cheap at 9 AM might have dried up by noon because a big corporate payment went through. The model has to adapt. Wise publishes annual reports that break down their cost of transfer — it's been dropping steadily, down to around 0.6 percent on average for personal transfers. A lot of that improvement comes from smarter routing. Luna: What about the bigger banks? They have their own networks. Are they using similar AI, or are they still reliant on correspondent banking and SWIFT? Lucas: Banks are catching up, but their legacy systems make it harder. JPMorgan has their Interbank Information Network, which uses blockchain-like technology, but it's more about data sharing than dynamic routing. The real competition for Wise is less the big banks and more other fintechs — companies like Airwallex and Currencycloud. Airwallex, for instance, has its own AI routing engine that optimizes for small and medium businesses doing frequent international payments. They claim they can process payments in under 24 hours to over 130 countries. Luna: But speed and cost aren't the only factors. What about compliance? Know Your Customer, anti-money laundering checks — those add friction. How does AI route around that? Lucas: That's the tricky part. Some corridors have very high compliance requirements — say, sending money to Iran or Syria is basically off-limits. But even between friendly countries, there can be delays. Wise embeds compliance checks into the routing model itself. The AI learns which transactions are likely to trigger manual review based on historical patterns, and it can either queue them differently or suggest an alternative corridor that might have a lower false-positive rate. Currencycloud, which was acquired by Visa a few years ago, takes a different approach — they offer a modular platform where the fintech client can plug in their own compliance rules, but the routing engine still optimizes within those constraints. Luna: So the AI isn't just optimizing for speed and cost — it's also optimizing for likelihood of clearing compliance. That's fascinating. Is there a risk that the model becomes too aggressive and tries to route around legitimate controls? Lucas: It's a real concern. Regulators in Europe and the US have started scrutinizing ai driven payment routing more closely, especially if it seems to be designed to minimize scrutiny. The European Banking Authority published guidelines last year on the use of AI for AML — they want explainability. If a transaction gets routed a certain way, the fintech has to be able to explain why, and not just say 'the model decided it'. Luna: That puts pressure on fintechs to use models that are interpretable, not just black boxes. Which is a challenge because some of the most accurate models — deep neural networks — are notoriously hard to interpret. Lucas: Right. So some companies are using ensemble methods that are more transparent, like gradient-boosted trees, which can still capture non-linear relationships but allow you to see which features mattered most. Wise, for example, has published that they use a form of random forest for some of their routing decisions. That gives them feature importance scores they can show auditors. Luna: Let's talk about the smaller players. A startup entering this space — how do they compete with the Wises and Airwallexes of the world? Is the AI moat that deep? Lucas: The moat is partly data, partly network effects. Wise has over a decade of transaction data — billions of data points — which makes their routing model hard to replicate. But smaller fintechs can partner with infrastructure providers like Currencycloud, or use APIs from companies like Railsbank or Thunes, which already have built-in routing intelligence. The key is to start with a specific niche — say, payments between freelancers in India and clients in the US — and optimize the hell out of that single corridor before expanding. Luna: So the AI itself might be off-the-shelf, but the data and the domain expertise give the edge. That makes sense. Lucas: Exactly. And one emerging trend is using AI for 'backup routing' — if the primary route fails or gets too expensive, the system instantly switches to a secondary or tertiary option. That requires real-time monitoring of success rates and latency across multiple providers. Some fintechs are now building multi-rail engines that can switch between SWIFT, local ACH, real-time payment systems like FedNow in the US, and stablecoin networks in the same transaction. Luna: Stablecoins add another layer. Using USDC or USDT on blockchain rails can settle in seconds, but the FX conversion and off-ramping still involve traditional banks. How does AI handle that hybrid environment? Lucas: It's becoming more common. A company called Circle, which issues USDC, has been working with fintechs to automate when to use blockchain rails and when to stick with traditional ones. The AI decides based on the corridor, the amount, and the counterparty's capabilities. For example, if both sender and receiver have USDC wallets, the AI might route it entirely on-chain. If not, it converts to fiat. That decision has to happen in under a second to feel seamless to the user. Luna: It's a complex optimization problem. And we haven't even touched on settlement risk. If the AI routes through a less established provider, what happens if that provider fails to settle? Lucas: That's where risk scoring comes in. Fintechs assign a credit rating to each routing partner based on their settlement history, regulatory capital, and even real-time news sentiment. Some models scrape news feeds and social media to detect negative signals about a partner bank. It's still early, but the idea is to incorporate reputation into the routing decision alongside cost and speed. Luna: So in a few years, every cross-border payment could be routed by AI that considers hundreds of variables in real time. That's a massive shift from the batch processing of the SWIFT era. Lucas: Absolutely. And if this kind of deep dive into fintech infrastructure is useful for what you're building or running, that's exactly why we keep these shows ad-free. The way that stays sustainable is listener support — buy me a coffee dot com slash fexingo. Just a small way to keep these conversations going without any sponsor pressure. Luna: Yeah, it really helps. And we appreciate everyone who's already chipped in. Lucas: Back to the routing story — one thing we haven't mentioned is the impact of real-time payment systems like India's UPI or Brazil's Pix. Those are domestic systems, but fintechs are starting to use them as 'on-ramps' for cross-border. For example, a US fintech might use Pix to instantly credit a Brazilian receiver, while the sender's dollars are being converted in a separate leg. The AI has to coordinate both legs to ensure the total time stays under a few minutes. Luna: So the routing is actually orchestrating multiple payment systems across different countries. That's a lot of moving parts. Lucas: It is. And the winner in this space won't just be the one with the best AI model — it'll be the one that can integrate with the most payment systems while maintaining reliability. That's a hard engineering challenge, but it's also a huge opportunity. Luna: Any predictions on where we'll be in five years? Will SWIFT still be relevant? Lucas: SWIFT will probably still be the backbone for high-value and interbank messaging, but for retail and SME cross-border, I think we'll see most traffic routed through ai driven multi-rail engines. The costs should come down to near-zero for digital corridors. The real differentiator will be speed and transparency — knowing exactly when the money will arrive and how much it'll cost. Luna: Something to watch. Thanks, Lucas. Lucas: Thanks, Luna. And thanks for listening.