Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Use AI for Real-Time Expense Fraud Detection
Transcript
- Lucas: So we talk a lot on this show about AI in lending, in payments, in credit scoring. But there's one area that's quietly saving companies hundreds of millions a year, and it's deeply unsexy: expense report fraud. Luna: Expense reports. The bane of every finance team's existence. But how big is the problem, really? Lucas: According to the Association of Certified Fraud Examiners, organizations lose about 5 percent of annual revenue to fraud, and expense reimbursement fraud is one of the most common schemes. That's trillions globally. Luna: And fintechs are going after that with AI? How does that work in practice? Lucas: Take AppZen, which is probably the poster child here. They use natural language processing and computer vision to analyze every line item on an expense report in real time. They check receipts against policy, flag duplicates, and even spot fabricated receipts. Luna: Fabricated receipts? Like someone photoshopping a receipt for a fancy dinner? Lucas: Exactly. Or using a receipt from a different date, or altering the amount. The AI can detect pixel-level inconsistencies that a human auditor would never catch. And it does it in milliseconds. Luna: So it's not just about catching fraud after the fact—it's preventing it from being approved in the first place. Lucas: Right. Real-time detection changes the game. Instead of a monthly audit where you find a few bad apples, you're flagging suspicious submissions before they're paid. That changes the incentive for employees. Luna: But doesn't that create friction? If every legitimate business lunch gets flagged, employees get frustrated. Lucas: That's the balance. The best systems learn from context. For example, if an employee always submits the same amount for a client lunch with a specific client, the AI learns that pattern. It only flags anomalies. Luna: So it's adaptive. What kind of anomalies are we talking about? Give me a concrete example. Lucas: Sure. One common one is duplicate submissions—someone submits the same taxi receipt for two different trips. The AI checks the receipt ID, the date, the amount, even the GPS coordinates if they're available, and flags it. Luna: GPS on receipts? That's a thing now? Lucas: More and more corporate cards provide transaction data with location. So the AI can cross-reference the restaurant location with the claimed purpose of the expense. Luna: Okay, that's pretty powerful. And this is all happening in real time, as the employee submits the report? Lucas: Yes. Platforms like AppZen integrate directly with expense management systems like Concur or Expensify. As soon as the report is submitted, the AI scores each line item for risk. High-risk items go to a human auditor, low-risk ones are auto-approved. Luna: So the human auditors focus on the edge cases. That makes sense. But I wonder about false positives. If an employee has an unusual but legitimate expense, do they get stuck in audit hell? Lucas: It can happen. But the systems are getting better at explaining why something was flagged. And they allow employees to attach additional context. The key is that the AI reduces the noise, so humans can focus on the signal. Luna: And I imagine for companies that spend millions on T&E, even a small reduction in fraud pays for itself. Lucas: Absolutely. AppZen claims their customers see a return on investment of up to 10 times within the first year. And that's before you factor in the soft savings from policy compliance. Luna: What about the cat and mouse game? Fraudsters will adapt. If they know the AI is looking for pixel inconsistencies, they'll get better at faking receipts. Lucas: That's an arms race, for sure. But the AI also evolves. It learns from new fraud patterns across the entire customer base. So if someone at one company tries a new trick, the model gets updated for everyone. Luna: That's a network effect. And it's not just receipts, right? There's also the policy side. Like, 'I bought a $500 bottle of wine for a client'—is that allowed? Lucas: Exactly. The AI can read the company's expense policy and compare it to each line item. So if the policy says no alcohol, or a maximum of $50 per person for meals, the AI flags it. That's where natural language processing comes in—it understands the policy text. Luna: So it's not just about fraud, it's about compliance too. That's a huge timesaver for finance teams. Lucas: And for auditors. Instead of manually checking random samples, they get a curated list of high-risk items. The efficiency gain is enormous. Luna: You know, it's interesting to think about how this ties into the broader fintech trend of embedding intelligence into everyday business processes. It's not a separate fraud detection tool—it's built into the expense workflow. Lucas: Speaking of which, this is exactly the kind of conversation that makes running a business smarter. And if these conversations are useful for what you're building or running, we hope you'll consider supporting the show. Luna: Yeah, we deliberately don't run ads on these episodes. We want to keep the focus on the ideas, not on selling you something. Lucas: Exactly. So if you want to support that choice, the link is buy me a coffee dot com slash fexingo. It's a small gesture that helps us keep doing this. Luna: Alright, back to the arms race. Lucas, you mentioned the AI learns from fraud patterns across customers. How does that work with privacy? Companies are sharing their expense data? Lucas: That's a great question. Typically, the data is anonymized and aggregated. The AI learns patterns like 'receipts from this particular taxi company are frequently forged' without seeing who submitted them. It's a balance between collective intelligence and data privacy. Luna: And what about companies that are concerned about sharing any data at all? Do they get a weaker model? Lucas: Some fintechs offer on-premise or private cloud deployments where the model is trained only on that company's data. The trade-off is that you lose the network effect, but for highly sensitive industries like defense or pharma, that's the right call. Luna: So there's no one-size-fits-all. But the trend is clear: real-time AI expense fraud detection is becoming table stakes for any enterprise expense management platform. Lucas: Absolutely. And I think we'll see it expand beyond T&E into procurement and accounts payable. Anywhere money is being requested for reimbursement, AI will be watching. Luna: That's both exciting and a little creepy. But for finance teams, it's probably a relief. Lucas: Yeah, I think the best way to look at it is: the AI handles the boring, repetitive stuff, and humans handle the judgment calls. That's a good division of labor. Luna: Alright, so to wrap up: if you're a CFO or a finance manager, what's the one thing you should take away from this? Lucas: That expense fraud is likely costing you more than you think, and the tools to catch it in real time are already here. It's not about punishing employees—it's about making the system more efficient and fair for everyone. Luna: And the AI is getting smarter every day. Thanks, Lucas. Lucas: Thanks, Luna. Talk to you next time.