Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Are Using AI for Real-Time Merchant Cash Advance Underwriting
Transcript
- Lucas: If these conversations are useful for what you're building or running, today we're looking at something that sits right at the intersection of payments data and lending: how fintechs are using AI to underwrite merchant cash advances in real time. Luna: Merchant cash advances — so, not quite a loan, but an advance on future card sales. That space has always felt a bit high-touch and slow. What's changed? Lucas: What's changed is that the data pipes are now fast enough — and models are good enough — that a fintech can pull a merchant's daily card-processing volume, look at bank account transactions, and make a funding decision in under an hour. No human underwriter, no stack of tax returns. Luna: And whose model are we talking about? Who's actually doing this at scale? Lucas: The poster child I want to focus on is Forward Financing. They're a Boston-based fintech that's been originating MCAs since 2012, but in the last three years they've rebuilt their underwriting engine around machine learning. They now process thousands of applications a month with an automated approval rate north of 60 percent. Luna: Sixty percent is high for MCA. Historically, a lot of those applications would have gone to a manual desk because the risk is so opaque. Lucas: Exactly. And the key insight is that the traditional FICO score is almost useless for a restaurant or a retail shop that runs most of its revenue through card terminals. Their credit score might be mediocre because of personal debt, but their daily card swipes show consistent cash flow. So Forward Financing built a model that ingests up to 18 months of daily processing data from the merchant's payment processor. Luna: They're looking at transaction counts, average ticket size, any seasonality — things like that. Luna: And the merchant doesn't have to upload a single PDF. That's a huge friction reduction. Lucas: It's the difference between a week-long process and a lunch-break decision. And for the fintech, it means they can price risk more granularly. Instead of one flat factor rate for everyone, they can offer better rates to merchants with strong, predictable volume — and higher rates to the volatile ones. Luna: But I've seen critiques that this speed could lead to adverse selection. If you approve too fast, you might attract merchants who need cash urgently because they're already in trouble — and the model might not catch that if it's only looking at backward-looking data. Lucas: That's a real concern, and it's why Forward Financing doesn't stop at the score. They've built a second-layer model that runs in parallel — a 'stress detector' that looks for signals like a sudden drop in average ticket size, or an increase in weekend transactions that might indicate the merchant is running personal expenses through the business card. Luna: So the AI is effectively looking for signs of desperation before the advance is funded. Lucas: Yes. And if the stress detector flags a merchant, the application gets routed to a human underwriter for a manual review. It's not fully automated for everyone — it's automated for the safe zone, and augmented for the edge cases. Luna: What about the cost side? Building and running these models isn't cheap. Do you know what their tech stack looks like? Lucas: From what they've published, they use a combination of gradient-boosted trees for the core scoring model — XGBoost, specifically — and a neural network for the stress detector. They run on AWS, and they've said their inference cost per application is under three cents. So the unit economics work. Luna: Three cents per application is almost nothing when your average advance is maybe 30 to 50 thousand dollars. Lucas: Exactly. And the broader market context matters here too. After the regional banking turmoil in 2023, a lot of small businesses found their credit lines frozen or their relationship bankers gone. Fintechs like Forward Financing stepped in with a faster, data-driven alternative. That accelerated adoption of AI underwriting across the entire MCA industry. Luna: So we've seen a shift from relationship-based lending to data-based lending, at least in this corner of small business finance. Lucas: And it's not just Forward Financing. Companies like Libertas Funding and Rapid Finance have rolled out similar models. Even some traditional lenders are starting to experiment with card-data underwriting. Luna: But there's a regulatory cloud — the CFPB has been circling MCA for years, and in 2026 they're pushing for more transparency around factor rates and payment structures. How does AI underwriting fit into that? Lucas: That's a great question. If the CFPB requires lenders to explain adverse actions, the 'black box' problem becomes a liability. But Forward Financing's model is interpretable — they use SHAP values to show which features drove the decision. So when a merchant is declined, the system can say 'your daily processing volume declined 20 percent in the last quarter' instead of just 'no.' Luna: That's actually a better customer experience too. A merchant who hears that has a clear action to improve. Lucas: Precisely. And I think that's the direction the whole industry is heading — not just faster approvals, but more transparent, more educational lending. The AI isn't just a gatekeeper; it's a coach. Luna: Speaking of useful — if today's episode gave you a new angle on how fintechs are opening up credit for small businesses, that's exactly the kind of conversation we try to have here. And the reason we can do it without ads is that some listeners choose to support the show directly. Lucas: Yeah, if you found this useful and want to keep it going, you can do that at buy me a coffee dot com slash fexingo. It's a simple way to say this is worth your time. Luna: And we really appreciate that. Now, back to the data — one thing I've been wondering: how often do these models get retrained? Because merchant behavior changes, especially after a shock like the pandemic. Lucas: Forward Financing retrains their core model quarterly, but the stress detector gets updated monthly. They also run a constant monitoring pipeline that tracks model drift — if the distribution of scores starts shifting unexpectedly, the data science team gets an alert. Luna: So they're not just setting and forgetting. That's important because the alternative data signals can be noisy. Lucas: Right. A merchant might have a great June because of a local festival, but that doesn't mean July will be the same. The model needs to be seasonality-aware. They explicitly engineer features like 'same-store sales versus last year same month' to smooth out those spikes. Luna: That level of granularity is really only possible with AI. A human underwriter couldn't track 18 months of daily data for every applicant. Lucas: No, and that's the whole thesis — AI scales the underwriting capacity. A small team can handle thousands of applications a month, and the models get better as they see more data. Luna: What about the repayment collection side? MCA is typically repaid via a fixed percentage of daily card sales, so the fintech gets paid when the merchant gets paid. Does AI play a role there too? Lucas: It does. Some fintechs use predictive models to forecast the merchant's future daily sales and adjust the withholding rate dynamically. If the model predicts a slow week, it might lower the percentage to avoid squeezing the merchant's cash flow. That reduces defaults and improves retention. Luna: So the same data that's used for underwriting is also used for servicing. That's a virtuous loop. Lucas: It is. And it's why the incumbents are paying attention. Banks like JPMorgan and Wells Fargo have started piloting similar cash-flow-based lending products, though they're moving slower because of legacy systems. Luna: What's your prediction for the next two years? Will AI underwriting become table stakes for any small business lender? Lucas: I think within two years, a small business owner will expect a funding decision in minutes, not days. The lenders that can't deliver that will lose share. But the winners won't just be the fastest — they'll be the ones who combine speed with fair pricing and clear explanations. That's where trust is built. Luna: And trust is the real moat, especially in a space that's historically had a reputation for predatory practices. Lucas: Exactly. The fintechs that use AI to be both faster and fairer — that's the combination that wins.