Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Use AI for Real-Time Treasury Management
Transcript
- Lucas: Luna, I was looking at a report from a treasury technology firm called Trovata — they analyzed data from over a thousand corporate finance departments. And one number jumped out: the average company still spends 40 percent of its treasury team’s time just pulling data together. Not analyzing it. Just gathering it from bank portals, spreadsheets, ERP systems. That’s the part AI is quietly killing. Luna: 40 percent — that’s basically two days out of a five-day work week. So the value add isn’t just faster forecasts, it’s freeing up people to actually think about strategy. Lucas: Exactly. And that’s the shift happening right now in corporate treasury. For decades, the treasury function was essentially backward-looking reconciliation. You’d close the books, see where cash was yesterday, maybe project out a few days. But ai driven treasury platforms are turning that into a real-time, forward-looking operation. I’m talking about cash flow forecasting at 95 percent accuracy for the next week, automated hedging recommendations, even intercompany loan netting. Luna: Intercompany netting — that’s one of those things that sounds boring but saves millions if you have subsidiaries in 15 countries. Manually tracking who owes whom, in different currencies, with different settlement dates... it’s a nightmare. Lucas: It’s exactly the kind of messy, data-heavy problem where machine learning shines. And the fintechs attacking this space aren’t the big core banking vendors — it’s startups like Trovata, Kyriba, and even newer players like Treasury Prime that are building on top of open banking APIs. They connect directly to your bank accounts, pull transaction data in real time, run it through a forecasting model, and spit out a recommendation. Luna: And they’re not just for Fortune 500 treasuries anymore. I spoke with the CFO of a mid-market manufacturer — about 300 million in revenue — who implemented one of these platforms six months ago. He told me they reduced their FX hedging costs by 30 percent because the AI was able to predict their euro and yen exposures more accurately than their previous monthly manual exercise. Lucas: That’s a huge concrete outcome. And the interesting thing is, the AI models are getting better because they have more data to train on. Every time a company connects a new bank account or a new ERP module, the model learns the payment patterns. It learns that payroll hits on the 15th and last day, that supplier X always invoices on Tuesday, that tax payments spike in April. The forecast becomes more granular over time. Luna: It’s almost like the model builds a digital twin of the company’s cash flows. But I’m curious — how do these models handle the unexpected? A big customer pays early, a supplier delays shipment, a currency moves 2 percent overnight. Does the AI adapt in real time? Lucas: The best ones do. They’re constantly updating their predictions as new data comes in. If a wire hits the account, the model recalculates the day’s ending balance and the next week’s forecast. Some platforms even send alerts to the treasurer’s phone: ‘Heads up, your euro balance is projected to dip below your target buffer in three days. Consider a spot trade now.’ That kind of real-time decision support is new for most companies. Luna: It’s a big leap from the old model where the treasurer would get a PDF report every morning at 9 AM that was already 12 hours stale. Now the question is — how do you get the data clean enough? Bank portals are not known for their API quality. Lucas: That’s the hard part. And this is where the fintechs have had to get creative. Some use screen scraping as a fallback. Others, like Trovata, have built direct API integrations with the major banks — JPMorgan, Bank of America, Citibank — and also with ERP systems like SAP and Oracle. But it’s still fragmented. There’s no universal standard for corporate banking APIs. So part of the AI’s job is actually data normalization — figuring out that ‘Payment ID 12345’ in one bank’s feed is the same transaction as ‘REF:INV-9876’ in the ERP. Luna: It’s a data engineering problem as much as a machine learning problem. And that’s probably why adoption has been slower than you’d expect. But the ROI is so clear — 30 percent savings on FX, 40 percent less time on data gathering. What’s holding back more companies? Lucas: I think two things. First, trust. Treasurers are naturally conservative — their job is to make sure the company doesn’t run out of cash. Handing over forecasting to an algorithm takes a leap of faith, even if the algorithm is demonstrably better. Second, integration complexity. A lot of mid-market companies don’t have the IT bandwidth to connect 12 bank accounts and an ERP system. That’s why we’re seeing fintechs offer white-glove onboarding as a service. Luna: It reminds me of the early days of cloud accounting software. Companies were hesitant to put their financial data in the cloud, but once they did, they never looked back. And you know, on that note — we do something similar here at Fexingo. We keep this show ad-free because we believe the content should speak for itself. If you find these conversations useful for what you’re building or running, the way to support that choice is buy me a coffee dot com slash fexingo. Simple, no strings attached. Lucas: Yeah, it’s a small way to keep the signal clear. Anyway, back to treasury — I wanted to mention a specific case study that I think illustrates the power of real-time forecasting. A mid-sized consumer goods company with operations in the US, UK, and Japan. They used to run a weekly cash flow forecast that took three people two days to produce. After implementing an AI treasury platform, they now get a daily forecast in 15 minutes, and the accuracy improved from 80 percent to 94 percent over six months. Luna: And I imagine that accuracy improvement directly impacts borrowing costs. If you know exactly how much cash you’ll have next week, you can minimize idle cash and avoid expensive overdraft facilities. Lucas: Exactly. The platform’s model actually recommends optimal cash positions across accounts — move excess from the operating account to an interest-bearing sweep account, or pay down a credit line early. Some of these platforms even execute the transactions automatically, with the treasurer approving a daily ‘suggested actions’ list. It’s a bit like a robo-advisor for corporate cash. Luna: One thing I’m wondering about — what about fraud? If you have AI automatically initiating payments or moving cash, that’s a huge security risk. How do these systems handle that? Lucas: They layer in anomaly detection. The same AI that forecasts cash flows also learns normal payment patterns. If a request to wire 500,000 to a new vendor in a high-risk country pops up, the model flags it and requires manual approval. Some platforms even integrate with external fraud databases. So the AI is doing double duty — optimizing liquidity and watching for suspicious activity. Luna: That’s smart. And it’s a good example of how AI in fintech is becoming more multifunctional. It’s not just one model doing one thing — it’s a suite of models working together. Lucas: Right. And I think we’re still in the early innings. The next frontier is probably real-time multi-currency netting for supply chains. Imagine a global manufacturer that can automatically net intercompany payables and receivables across 50 entities every day, using AI to optimize the settlement currency and timing. That could save billions in transaction costs and FX spreads across the industry. Luna: And it would make working capital management truly dynamic. You could adjust payment terms dynamically based on real-time cash positions — offer a discount for early payment if you’re flush with cash, or extend payables if you’re tight. That’s a level of sophistication that only the largest companies have today. Lucas: But with fintech, that capability is trickling down. I saw a demo from a startup called Cashforce that’s targeting companies with as little as 50 million in revenue. Their pitch is essentially ‘treasury intelligence as a service.’ No need to hire a PhD in quantitative finance. The AI does the heavy lifting. Luna: That’s the real story here — the democratization of treasury management. Five years ago, only a company with a dedicated in-house treasury team could do daily cash forecasting and automated hedging. Now, a CFO at a 100-million-dollar firm can get similar capabilities from a SaaS subscription. Lucas: And that changes the competitive landscape. If a smaller company can manage its cash more efficiently, it can offer better terms to customers, invest more in growth, and weather downturns more easily. It’s one of those behind the scenes fintech innovations that doesn’t make headlines but has real economic impact. Luna: I want to come back to the forecasting accuracy number — 95 percent. How do they validate that? Because it’s easy to claim high accuracy if you’re only forecasting one day out. But a week out, with all the variables, that’s harder. Lucas: Good question. The platforms I’ve looked at measure forecast accuracy using a metric called ‘mean absolute percentage error’ — MAPE — and they report it separately for different horizons. One day out, some are hitting 98 percent. One week out, 95 percent. One month out, it drops to maybe 85 percent, which is still better than most manual forecasts. And they’re transparent about it because that’s how they build trust with treasurers. Luna: It’s also worth noting that the models improve over time. The more data you feed them, the better they get at understanding seasonality, payment cycles, even macroeconomic correlations. Lucas: Absolutely. And that’s the flywheel effect. As adoption grows, the models get more data, the forecasts get better, and more companies adopt. It’s a classic data network effect. I think in five years, using AI for treasury management will be as standard as using cloud accounting software. Luna: And the losers will be the companies that stuck with the Excel spreadsheet model. Not because Excel is bad, but because it can’t scale to real-time, multi-source data integration. Lucas: Right. Spreadsheets are great for analysis, but they’re terrible for real-time data ingestion and automated decision-making. The fintechs are building on a new stack — APIs, cloud data warehouses, machine learning models — and that stack is becoming a competitive necessity. Luna: One last thing — what about regulatory considerations? In some jurisdictions, moving cash across borders or using AI for hedging decisions might trigger reporting requirements or require board approval. Lucas: It’s a real consideration. The platforms I’ve seen build in compliance modules that flag transactions requiring regulatory reporting — for example, large cross-border wires that need to be reported to the local central bank. And some treasurers choose to keep the AI in ‘advisory mode’ rather than ‘execution mode’ to maintain human oversight. But the trend is definitely toward more automation as trust builds. Luna: It feels like we’re at an inflection point. The technology is proven, the ROI is clear, and the barriers are mostly about change management and integration. I suspect the next few years will see a wave of adoption. Lucas: I agree. And it’s one of those quiet revolutions that makes the financial system more efficient, one company at a time. I’ll be watching to see which industries adopt fastest — I’d bet on manufacturing and retail, where margins are thin and cash cycles are complex. Luna: Good call. And on that note, I think we’ve covered treasury management pretty thoroughly. Thanks for the deep dive. Lucas: Always a pleasure. That’s it for this episode of Fintech Conversations. Until next time.