Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Use AI for Real-Time Payment Routing Optimization
Transcript
- Lucas: Every time you tap your card or click 'buy now', there's a silent war happening behind the scenes — a split-second battle between payment rails, each with different costs, different success rates, and different rules. Luna: And most of us never think about it. We just assume the payment goes through. Lucas: Right. But for the merchant, that choice of which rail to use — Visa, Mastercard, a local debit network, a real-time payment scheme like FedNow — can mean the difference between keeping two percent of the transaction or losing it to fees and failed attempts. Lucas: So a growing number of fintechs are using AI to make that routing decision in real time, optimizing for cost, speed, or reliability depending on what the merchant values most. Luna: And this isn't theoretical — Stripe, Adyen, Checkout.com, even Railsr are all building or already have these features. Lucas: Exactly. Let's focus on Stripe's 'Optimized Routing Suite' — they rolled it out in early 2025, and by mid-2026 it's processing billions of dollars in transactions daily. Luna: What's the core idea? Lucas: The core idea is a machine learning model that looks at every incoming transaction — card type, issuer country, time of day, historical success rate of that card at that merchant — and then picks the best path. Lucas: If a Visa card from a specific bank in Brazil has a 98% success rate on the Visa network but a 99.5% success rate on the local Elo network — and Elo's fees are lower — the model routes it there. Luna: And it makes that decision in milliseconds. Lucas: Yes. The latency requirement is brutal — the whole routing decision has to happen in under 50 milliseconds, because the customer is waiting for that spinner to stop. Lucas: Stripe says merchants using the suite see an average 1.2% increase in payment success rates and a 0.8% reduction in fees. For a merchant processing $10 million a year, that's about $200,000 in combined benefit. Luna: That's real money. And it's not just Stripe — Adyen has something similar they call 'RevenueAccelerate', and Checkout.com has 'Smart Routing'. Lucas: Right. But what I find interesting is that the smaller players, like Railsr, are actually going deeper into specific verticals. For example, they built a routing model optimized for travel bookings, where authorization rates are lower because of cross-border fraud flags. Luna: So it's not just about cost — it's about getting the payment through at all. Lucas: Exactly. And the model learns continuously. If a particular bank starts declining more transactions on a Sunday evening, the model adapts by Monday morning. Lucas: Now, there's a tension here. Some merchants want the cheapest route every time. Others want the most reliable. Stripe's model actually lets you set a preference slider — cost vs. success rate — and the AI optimizes within that constraint. Luna: Which is smart because it gives the merchant control without them having to understand the underlying complexity. Lucas: And that's the key — the merchant just sets their preference, and the AI does the rest. This is the kind of 'invisible fintech' that makes the whole system work better. Lucas: If these conversations are useful for what you're building or running — a couple of dollars a month is genuinely what keeps these going. Buy me a coffee dot com slash fexingo, if you've gotten something out of them. Luna: Yeah, it really does make a difference — keeps everything ad-free and focused on the actual tech. Lucas: Alright — back to routing. One of the more surprising applications is in B2B payments, where invoices are often paid via wire transfer or ACH, both of which have different cut-off times and settlement speeds. Luna: And B2B is a completely different beast from consumer payments — higher values, more fraud risk, slower rails. Lucas: Right. So Bill.com, for example, uses AI routing to decide whether to send a payment via ACH, virtual card, or wire based on the supplier's preferences and the payment amount. They claim a 3% lift in early payment discounts because the routing is faster. Luna: Let's talk about the elephant in the room — regulation. Does routing AI run into issues with, say, the European Interchange Fee Regulation or the Durbin Amendment in the U.S.? Lucas: It can. In Europe, for instance, the Interchange Fee Regulation caps interchange fees at 0.2% for debit and 0.3% for credit. So the routing AI might not save much on fee arbitrage there, but it can still optimize for success rates. Lucas: And in the U.S., the Durbin Amendment requires that debit cards have at least two unaffiliated networks available — so the AI has to actually choose between, say, Visa and Star or NYCE. That's a genuine routing decision. Luna: So the regulatory framework actually creates the opportunity for optimization. Lucas: Exactly. In some ways, regulation forces competition onto the merchant's side, and the AI is just picking the winner. Lucas: One more angle — there's a startup called Swapp that focuses purely on routing for subscription businesses. They noticed that recurring payments have a different pattern: the first payment might succeed, but the second and third fail because the card expires or the bank flags it as unusual. Luna: So they built a model that predicts which rail is most likely to succeed for a retry — maybe using PayPal as a fallback if the card fails. Lucas: Exactly. And they claim to recover 8-12% of failed recurring payments by smart routing retries. For a SaaS company with thousands of subscribers, that's huge. Luna: It's interesting how this technology, while invisible to the end user, is fundamentally changing the economics of online commerce. Lucas: It really is. And I think we're only at the beginning — as real-time payment rails like FedNow and the European Payments Initiative roll out, the number of options will grow, and the AI models will become even more valuable. Luna: So the question becomes: how do you build a system that can handle 20 different rails, each with different rules, different uptimes, different costs? Lucas: That's the engineering challenge. Some of the incumbents are actually partnering with cloud providers to run inference at the edge — closer to the point of sale — to cut latency. Lucas: Adyen, for example, uses AWS Lambda functions deployed in multiple regions so the routing decision happens within a few milliseconds of the data center closest to the customer. Luna: That's an infrastructure play as much as a data science play. Lucas: Exactly. And the winners will be the ones who combine the best ML models with the best distributed systems. It's like building a GPS for money — you need real-time traffic data, accurate maps, and a fast engine. Luna: I like that analogy. So where do you see this going in the next two years? Lucas: I think we'll see routing become a commodity feature — every payment processor will have it. The differentiation will be in the quality of the data: how many payment rails you connect to, how granular your success-rate data is, and how quickly you retrain your models. Lucas: Also, I expect to see routing AI start to incorporate fraud signals more directly — routing a transaction not just for cost and success, but for fraud probability. If the model thinks a transaction is risky, it might route it to a rail that gives you more time to verify, like a slower network. Luna: That would be a game-changer — merging fraud detection and routing into one decision. Lucas: Right. And some startups are already working on that. So the next few years should be really interesting. Luna: Well, as always, thanks for diving into the plumbing with us. Lucas: It's where the real value is.