Xero's Global GM of Data & AI Products explains how the company uses millions of financial documents to build AI features small businesses can't create alone. Hubdoc runs OCR on bills and invoices, extracts key fields, then feeds the data into models. The hardest problem: spotting rare credit notes hidden among thousands of invoices, which if miscoded create large ledger errors. Xero built a detection model that flags them, paired with a human-in-the-loop prompt asking users to double-check before confirming.
Chapters
So you mentioned that at Xero you have millions of small businesses and you can see patterns across the data that an individual small business couldn't see. Can you give me an example of some of these patterns that appear?
Yeah, this is actually another really, really interesting one because it's actually quite tricky to build, but when we released it, folks were just in raptures. So we have a product in a number of the countries we work in called Hubdoc. And Hubdoc allows people to upload their financial documents and to have those financial documents automatically processed and added to their accounting software. And we have for a number of years had a machine learning driven pathway so that when you upload your financial document, be that a bill, an invoice, it goes through OCR, optical character recognition, it comes out as an electronic version of itself and we then go through and pick out the key fields so we can pick out what currency is that in, what date was it, who was the merchant and what is the invoice total, for instance, let's say. Now one of the most annoying things, and I'm not an accountant, so people will have to forgive me if I get this slightly wrong, but I'm fairly sure I am correct, is when what you think is an invoice comes through, but it is actually a credit note. So rather than somebody charging you money, they're actually giving you money back because you've overpaid a previous invoice or something like that. Now as you can imagine, if you have a document come through and you code it as an invoice, when it is actually a credit note in your back end, you actually end up with a double up problem because it was money that was supposed to go on one side of your ledger and it goes up on the other. So say it was 20%, 25 grand, you'll actually get a $50,000 offset of incorrectness. And where the money is supposed to be because you took negative 25 and put in a positive 25 effectively, so you've got a 50. So it is the bugbear of small business owners, accountants and bookkeepers everywhere when somebody miscodes a credit note because you have to go picking through the reconcile general ledger, understanding how on earth you have a $50,000 incorrect level in your bank account versus your reconciled ledger. Now what we are able to do by looking at millions and millions and millions of financial documents was build something called, okay, this is our great naming here, credit note detection. But it does exactly what it sounds like it does. It's not a simple problem. As your listeners will understand, because it's anomaly detection. And anomaly detection is hard because for every, for every credit note there are, I'm going to get this slightly wrong, but hundreds, if not thousands of invoices. So the documents look very, very similar. Some people will stamp credit note across them and some people won't. They look very, very similar. And 1% ish or a fraction of a percent ish are actually credit notes. So it's a hard thing to do. Well, it's also actually a hard thing for humans to pick up because we're really, really good at getting into the mode and just going, yep, click, click, click, click, click. But that is an example of where we can use the transactions of millions of people to pick up the subtle patterns that indicate this isn't an invoice, it's credit note and actually alert humans in the flow. So that's something I really like too, is actually the way we do it. It's very easy for humans to fall into the mode of I'm just clicking a button and we need to be reminded. You probably need to pay a little more attention before you click this button. And it's one of my sort of hobby horses is I get frustrated when I hear people think that an AI product is the algorithm. It's the sexy part, it's the part that gets talked about, it's the part that gets published, it's the part that gives people patents, but it's only part of a really functional AI product. And when I'm talking to folk about how to build really functional AI products, I'm always reminding them that it's a systems problem and it's the end to end flow, including the human. And in this case of credit note detection, the piece of the flow that actually we thought quite hard about, which is this is an impactful thing because if we get it wrong, it's really annoying and if the human doesn't notice, it's really annoying and ends up in the wrong place. So we actually subtly, but in a nice, in a way that who are human stands out because it is different, redesigned the interface at the point where we say, hey human. We're fairly sure that this is in fact not an invoice, a credit note, but we suggest that you now pause and take a little time to think and to genuinely verify our decision. And so that just that subtlety and redesign the interface and making that pop up to the person in a way that is noticeable was a really important part of building a really useful AI product. And we rolled credit note detection out. Gosh, time flies. So it's probably longer ago than I think it is now. Several months ago, and I kid you not, people love it to pieces. It was one of the most requested fake product features, and when we rolled it out, people were in raptures that we have now an anomaly detection system that allows them to avoid accidentally coding a credit note as an invoice.
