Expand your moat with machine learning

Hey Xero

By Trevor Clarke, TRA Director

The idea of a ‘moat’ as a metaphor for competitive advantage in the business world isn’t new. But it is doing the rounds again.

Put simply, the wider your moat is the more advantage you have over existing and new competitors in your industry as they’ll find it harder to attack your position because they need to cross it.

Even though nobody has a literal moat anymore (aside from a few historical castles, of course), it’s an easy way to articulate competitive positioning.

There are, of course, many things that make up any given organisations’ competitive advantage. AI and machine learning are one area that is now starting to inform this as well.

The world of software providers is a good example. Every one of the largest software providers and Internet companies in the world are either using AI within their own products or bringing to market their own AI-based offerings. .

But it’s also smaller providers as well. Take Xero, the New Zealand SME SaaS accounting software provider who has grown in 10 years to having over 1 million subscribers worldwide.

In the past 12 months, the company has commenced seriously pursuing things like chatbots and machine learning.

The major thing Xero has done (according to interviews TRA conducted with Xero representatives), is it has added machine learning to its account coding for invoice line items. Effectively, Xero has taken out the need for people to manually do accounting when they're creating invoices. So they can just focus on creating the invoice and the machine learning takes care of the accounting in the background, once, of course, the system has learned the way that they operate and the way that the business operates. Yes, it’s a lot more involved than that, but for the sake of this article let’s just say they’ve used machine learning to automate some parts of invoicing in Xero.

The genesis of this development was a Xero-wide hack-a-thon called Explore, which it runs once or twice a year. One team in the hackathon, who had experience and interest in machine learning, had written a machine learning model that could do automated coding of bank reconciliations. When they gave the demo it showed how the technology could be applied. According to reports, it “brought the house down”.

Even internally the event was seen as a classic hack-a-thon project; it was largely smoke and mirrors. But it was enough to get machine learning really on the map internally at Xero. And now with a grassroots team of employees, many who had an existing interest in machine learning, the company is pursuing it for some of its customer challenges.

One thing to keep in mind, which Xero learnt, is that machine learning does a great job of learning how an individual business (or thing) works but when you try to relate that to what another business (or thing) does, there's often not really a lot of crossover. All the similarities are at a layer a bit deeper than that, and what Xero also found is that most of the commercially available models like AWS, Microsoft, IBM, and Google tend to deal with very large models. They might say, have data on a million different people and they can predict what one person will do based on what the other million people did.

But with small businesses – Xero’s core customer base – it's really not like that at all. They're all snowflakes.

The technique or the tools that Xero uses are therefore open source, but adjusted for its own unique needs. The way it implements the algorithm is unique in every case. It’s kind of like interior design. All the tools and objects are similar, but every space has its own design. So for Xero it has had to tailor the machine learning algorithms it uses for each and every customer. This means the developers had to not just have a deep understanding of the way businesses use its software and conduct their accounting, but an intimate picture of the next level down in terms of how they use specific functions.

In testing the machine learning model came out better after the very first invoice. After the first four invoices – where the algorithm was tested against what a user would do – it was getting up over about 80% accuracy, which is a good result. That was when a lot of the developers internally went “It's not just a buzz word; it's not just a thing, the latest fashionable thing; it really is actually probably the best way to solve this problem”.

Although customers have been advised the machine learning is now being used, Xero is keeping it low key. But it's like another moat for Xero, really. It is getting a head start on competitors by offering more advanced automation.

In the conversations TRA had with Xero, it was also clear that one of the other lessons is you can’t just hire a couple of data scientists and have this work. You have got to do the investment, and you need to be a mature product organization to be able to realize it. Machine learning solved an intractable problem – how to help users with their manual steps in invoicing. But it may not be appropriate for everything Xero needs to address. That’s a step by step process of evaluation.

The thing is, even if you don’t believe you are doing machine learning or using AI, today you likely are as a result of it being built into the software you already use (as Xero’s experience shows). Even if you don’t know it. Everyone who uses digital technology (or IT) is. And it’s only going to spread. The question is, what are the competitive implications of the right application of machine learning and AI to your organisation?

Yes, AI is going through a bout of hype. I’ve heard many talk about this being one of those 20-year cycles where technology like this comes back into fashion. I get it and I agree to some degree. But I also think there are many differences, especially in terms of the availability and accessibility of machine learning + AI tools, services and talent across the globe.

As Xero’s experience and that of many others we could introduce shows – it’s happening already. This doesn’t mean you need to go out and create the next “Jarvis”, “Hal 6000”, “Siri, “Cortana”, “Alexa”, or “Watson”.

You might find machine learning might be applied in a smaller, more modular fashion that offers up gains for you and your customer. But unless you go through the evaluation and ideation phases – and particularly if you don’t understand your customers’ journey - you won’t know. And you won’t know which AI and machine learning tools and services are right for your organisation. That’s never a good position for any IT or business leader to be in, especially if you already know that competitors (and their supply chains) are building their moats out of this kind of technology.