Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / How Fintechs Are Using AI for Real-Time Accounting Reconciliation
Transcript
- Lucas: If these conversations are useful for what you're building or running, you're in the right place. I'm Lucas. Luna: And I'm Luna. Today we're talking about one of the most boring but painful jobs in finance: accounting reconciliation. Lucas: Right. Anyone who's worked in a finance department knows the drill — you've got bank statements, invoices, payment receipts, and you need to match them all up. It's months-end, everyone's stressed, and it usually takes days or even weeks. Luna: But some fintechs are now using AI to do it in real time, or at least in hours. Let's talk about a specific case: a startup called AutoRecon. Lucas: AutoRecon is a small company — about 40 employees — but they've built a platform that uses a combination of natural language processing and neural networks to match payments to invoices automatically. They recently deployed at a mid-market retailer that processes about 40,000 transactions a day. Luna: Forty thousand — that's a lot of paper cuts. What was the before and after? Lucas: Before, their accounting team had four people spending roughly two solid weeks each month just on reconciliation. After AutoRecon, that dropped to under an hour of review time. The AI flags only the exceptions — maybe three to five percent of transactions — and a human looks at those. Luna: So the machine handles the straightforward matches, and humans focus on the edge cases. That's the classic human-in-the-loop model. Lucas: Exactly. And the tech stack is interesting. For parsing invoice text — like line items, totals, tax amounts — they use a fine-tuned version of a transformer-based language model, similar to what you'd use for document understanding. On the payment side, they use a graph neural network that learns the relationship between payments and invoices over time. Luna: Why a graph network? I would have thought a simple matching algorithm would suffice. Lucas: Good question. The thing is, payments don't always match one-to-one. Sometimes a single payment covers multiple invoices, or a partial payment plus a credit note. The graph network can model those complex relationships — like who paid what, when, with which reference — and find patterns that a rule-based system would miss. Luna: So it's not just about matching amounts; it's about understanding the entire payment graph. Lucas: Right. And the result is that the retailer's CFO told me their close process used to be a bottleneck for month-end reporting. Now they have a real-time view of their cash position every single day. That's a huge change for a company that was previously operating with a two-week lag. Luna: Now, is this only for big companies? What about a small business with, say, a hundred transactions a month? Lucas: AutoRecon's pricing is per transaction, so it scales down. But I think the bigger point is that the technology is becoming accessible. Five years ago, you needed a team of data scientists to build something like this. Now, there are off-the-shelf APIs for invoice parsing and matching models. The barrier to entry has dropped dramatically. Luna: So we're likely to see a wave of startups offering reconciliation as a service, similar to what happened with payroll and expense management. Lucas: Exactly. And the incumbents — like the big ERP vendors — are under pressure. Their systems were built for batch processing, not real-time. They have decades of cruft. A nimble fintech can eat their lunch on specific use cases. Luna: But what about the auditors? If the AI is doing the matching, how do they verify the process? Lucas: That's a great point. AutoRecon provides an audit trail — every match has a confidence score, and the system logs the features it used to make the decision. An auditor can drill into any exception and see why the AI flagged it. In fact, some auditors I've spoken to say this actually improves audit quality because you're reviewing all exceptions, not just a sample. Luna: So it's not just faster; it's potentially more thorough. Lucas: Right. And this ties into a broader trend: continuous accounting. Instead of closing the books once a month, firms are moving to a model where the books are always up to date. AI reconciliation is a key enabler. Luna: Before we go further, I want to acknowledge something. If this episode gave you a concrete idea — something you could apply or share — that's exactly what we're here for. And the way these episodes stay ad-free, with no sponsors interrupting, is through listener support. If you find value in shows like this, you can go to buy me a coffee dot com slash fexingo. Lucas: Yeah, that's how we keep doing what we do, without selling your attention. So, back to reconciliation — one thing I find fascinating is the data quality challenge. The AI is only as good as the data it's trained on, and many companies have messy data. Luna: What do you mean by messy? Duplicate invoices? Missing references? Lucas: All of that. For instance, the retailer had about eight percent of their invoices with no purchase order number, and payment references sometimes had typos. The AI has to be robust to that. AutoRecon's model was trained on synthetic data that included noise — misspellings, missing fields, partial amounts — so it learned to handle the real-world mess. Luna: So it's not just about having an AI; it's about training it on the right kind of messy data. Lucas: Exactly. And that's a competitive moat. A startup that's seen a thousand different payment patterns can generalize better than one that's only seen clean data. Over time, the data network effect becomes powerful. Luna: So where do you see this going in the next two years? Will reconciliation become fully automated? Lucas: I think for routine matching — say, ninety percent of transactions — yes. But the tail of complex cases will still need human judgment, especially when there are disputes or fraud. However, the role of the accountant will shift from doing the matching to managing the exceptions and improving the AI. Luna: That sounds like a much more interesting job. Lucas: Exactly. Less drudgery, more analysis. And that's ultimately what fintech should do — free up humans for higher-value work. Luna: Alright, that's a good note to end on. Thanks for listening, and we'll see you next time on Fintech Conversations.