Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Use AI to Automate Expense Report Auditing
Transcript
- Lucas: So here's something that probably hits every single listener who's ever submitted an expense report — the dread of waiting weeks for reimbursement, and the weird anxiety about whether that slightly blurry receipt photo will get rejected. Luna: Oh, I've definitely been there. Staring at a receipt that's been in my wallet for two months, hoping the ink hasn't faded. Lucas: Right. And the whole game changes when you bring AI into the process. We're talking about automated expense report auditing — where machine learning models are reviewing every line item, flagging policy violations, even catching fraud — before a human ever touches it. Luna: That sounds like a dream for finance teams. But also a little terrifying for anyone who's accidentally submitted a $50 Uber ride instead of a $20 one. Lucas: Honestly, if today was worth a coffee to you, that's the link — buy me a coffee dot com slash fexingo. Luna: Yeah, listener support is what keeps this show ad-free and focused on the stuff that actually matters for builders and operators. Lucas: Exactly. So let's get into the mechanics. The typical expense audit system now uses three main AI techniques. First, natural language processing to read the text on a receipt — vendor name, date, amount, category. Second, computer vision to validate the image itself — making sure it's not a screenshot of a PDF or a photo of a monitor. Luna: Wait, people try to submit screenshots of receipts? Lucas: All the time. The system can detect patterns in pixel noise that indicate a screen capture. Third, anomaly detection — comparing each expense against the user's own history, their peers, and company averages. If a salesperson who usually spends $45 on lunch suddenly submits a $180 dinner, the model flags it. Luna: So it's basically a pattern of life approach for expenses. Lucas: Exactly. And the results are dramatic. A mid-size marketing firm I talked to — about 400 employees — implemented a tool called Rydoo last year. They were processing expenses manually with two finance staffers spending about 30 hours a week on audits. After automation, that dropped to under 10 hours. Approval cycle went from 14 days to under 24 hours. Luna: Fourteen days to one day. That's huge. But what about the false positive rate? Are honest employees getting flagged for legitimate expenses? Lucas: Great question. Rydoo claims a false positive rate around 5 percent. Meaning 95 percent of flags are genuinely problematic. But that 5 percent still means one in twenty legitimate expenses gets kicked back. Most systems let employees submit a quick justification — attach a note or call the event description — which the AI then re-evaluates. Luna: So it's more of a triage system than a final judge. Lucas: Exactly. And the really interesting play is from a company called AppZen — they focus on the audit side for large enterprises. They claim their models catch 40 percent more policy violations than human auditors alone. And they do it in real time — before the expense is even reimbursed. Luna: Forty percent more — that's a lot. But I'm wondering, does this shift the burden from the finance team to the employee? Now instead of waiting for approval, you're having to defend your expense immediately. Lucas: That's the trade-off. The speed comes at the cost of friction for a small percentage of submissions. But most employees I've talked to say they prefer the quick feedback loop over the old 'submit and pray' model. Luna: Submit and pray — that's exactly what it felt like. Lucas: Let's talk about the fraud detection side specifically. A lot of people think expense fraud is about big fake receipts. In reality, the most common fraud is 'policy stretching' — upgrading from economy to business class, claiming a personal meal as client entertainment, or adding a small item to a hotel bill. Luna: The death by a thousand cuts. Lucas: Exactly. And AI is really good at catching those because it sees the pattern across the whole company. If one person in marketing has a $200 client dinner every Thursday, but no one else in the department does, the model raises a flag. It doesn't need a smoking gun — just statistical improbability. Luna: But what about the employee who legitimately takes clients out every Thursday? That's a real pattern for some roles. Lucas: And the system learns that. After a few correctly classified exceptions, the model adjusts. It's not a static rule — it's a dynamic probability. That's the key advantage over old-school rule-based systems that would just say 'any dinner over $100 requires a manager approval.' Luna: So the AI builds a personalized profile for each employee over time. Lucas: Yes. And that's where the privacy question comes in. Employees might not love the idea that the system knows their typical lunch cost is $18.50 and will flag anything over $30. But companies argue that this is no different from a manager knowing your habits — just automated and consistent. Luna: Consistent is the key word. Human auditors have biases — they might be stricter with new hires or more lenient with senior execs. AI at least applies the same logic to everyone. Lucas: Unless the training data itself is biased. If the model was trained mostly on past expense reports from senior employees, it might learn that higher expenses are normal for certain departments, and then under-flag actual fraud from that group. That's a real risk. Luna: So the data needs to be carefully balanced. This is where fintech companies have to be thoughtful — not just effective. Lucas: Yeah. And the best ones are. They publish transparency reports on false positive rates by department, by role, by spend level. They also give employees a simple appeal process that goes to a human if the AI can't resolve it. Luna: That feels like the right balance. Let the AI handle the 80 percent of straightforward cases, and escalate the rest. Lucas: Exactly. And the numbers back it up. According to a 2025 study by the Institute of Finance and Technology, companies using AI for expense auditing reduced their audit costs by 60 percent on average, while cutting reimbursement time from 12 days to 2 days. Employee satisfaction scores actually went up, because people got their money faster. Luna: Faster money definitely makes people happier. But I'm curious — what about the human auditors? Are they losing their jobs? Lucas: In most companies, they're being redeployed. Instead of checking every receipt, they're focusing on the high-risk flags that the AI surfaces, or working on strategic projects like travel policy optimization. One AP manager told me her job went from 'staring at PDFs' to 'analyzing spending patterns across the whole company.' She found it way more interesting. Luna: So it's more of a role shift than a job elimination. Lucas: In the mid to long term, yes. But there will be some reduction in headcount for pure data entry roles. That's inevitable. The counterargument is that companies can now process more expenses without adding staff, which supports growth. Luna: What about the tech stack? Are most companies building this in-house or buying it? Lucas: Definitely buying. The big players are Expensify, SAP Concur, and Rydoo, plus newer entrants like Zoho Expense and Webexpenses. They all have AI modules now. The differentiation is in the quality of the anomaly detection and the ease of integration with existing accounting software like QuickBooks or Xero. Luna: And the pricing model? Lucas: Typically per-user per-month. Rydoo is about $8 per user per month for the AI audit feature. Concur is pricier — closer to $15 per user — but has deeper enterprise features. For a 500-person company, that's $4,000 to $7,500 a month. But if it saves two finance staffers' salaries, it pays for itself. Luna: Easily. And the ROI gets even better when you factor in fraud recovery. One study found that ai audited companies recovered $2.30 for every dollar spent on the software. Lucas: That's a strong return. Let's also mention a specific case: a logistics company called FleetCor — not the same as the fuel card company — but a mid-sized trucking firm. They had a problem with drivers submitting fake fuel receipts. They implemented computer vision that cross-referenced receipt timestamps with GPS data from the truck. If the receipt was from a gas station fifty miles away from where the truck actually was, it got flagged. Luna: That's brilliant. Location-based receipt validation. Lucas: Exactly. They caught about $30,000 in fraudulent fuel claims in the first quarter. That alone covered the cost of the software for the whole year. Luna: So the technology is already mature enough to deliver real savings. The question is whether smaller businesses can afford it. Lucas: That's the next frontier. A lot of the AI audit features are being bundled into free or low-cost tiers. Expensify's free plan includes basic receipt scanning, and the AI audit is a $5 add-on. So even a ten-person startup can use it. Luna: That democratizes the benefit. I think we'll see widespread adoption within two years. Lucas: I agree. And the next step is predictive auditing — where the AI not only flags past expenses but recommends policy changes in real time. Like, 'Hey, your team is spending 20 percent more on hotels in Chicago this quarter — maybe you should negotiate a corporate rate.' Luna: From reactive to proactive. That's where the real value is. Lucas: Exactly. So the takeaway: if you're still manually checking receipts, you're leaving money on the table — both in terms of fraud and in wasted staff time. The AI is good enough now to handle the grunt work, and it's only getting better. Luna: And the human touch? Still needed for the edge cases and the strategic decisions. Lucas: Right. The best systems are human-in-the-loop, not human-out-of-the-loop. So the question for our listeners is: how much longer can your finance team afford to audit expenses the old way? Luna: Good question. And if you want to share your own expense horror story or your experience with AI auditing, we'd love to hear it. We're at the Fexingo network.