Technology at its best is designed to augment the tasks you do everyday. Quietly making your life easier, until eventually you wonder how you ever did without it. That is exactly what Xero has achieved with personalised machine learning applied to the sales invoice process.
The machine learning system is designed to address a common problem - small businesses struggle with getting account codes right. Which means that accountants and bookkeepers spend time recoding transactions. And time is money.
While Find & Recode helps fix this by providing a way to recode consistent mistakes in bulk, there have been over 3 million recoded transactions since its launch. This still takes too much time, and doesn't eliminate the mistakes the small business owners make. The question is - how do you address the behaviour that allows the mistakes to be made to begin with?
Solving that problem is no mean feat
There are more than 10.1 million unique account codes in Xero alone. Each business is different, using their account codes in different ways. Weird, and wonderful ways. This individuality guided Xero down the path of developing their own solution as "off-the-shelf products aren't built to deal with the complexities found in accounting". This first step is the result of that work. A specialist, personalised machine learning system that helps business owners correctly code their sales invoices.
"The machine learning technology developed by Xero takes an incredibly complex problem and delivers a solution which will directly help our customers and end-users from day one."
How it works is simple
Small business users continue to code their accounts as normal, making the same or similar mistakes to those they've made in the past. Accountants continue to recode the accounts, as they have always done. The machine learning system learns what the accountants change in the system and what it relates to. An invoice for time spent on site recorded against Sales - Labour and not Sales - Materials for instance. When the small business comes to create their next invoice, all they need to do is enter a few key details - the customer, quantity and a description. Xero automatically suggests the account code so that they don't inadvertently make the same mistake.
This is where the real magic happens. The suggestions start to shape the behaviour of the user. Technology becoming the teacher, the frequency of mistakes decrease as new habits are formed, reducing the need to find and recode. Allowing advisors to do what they do best.
The best part is, you aren't even aware it's happening
When we first tested it on our own Xero organisation, we clicked on new sales invoice', our breaths catching with wide eyed anticipation. This quickly deflated when it looked as if nothing was different. I'm not sure what we expected. An icon of a robot perhaps? Something to hint that there was some cleverness at play? It wasn't until we typed bookkeeping' in the description field of an invoice line, not having chosen an inventory item, and Xero happily blinked back Fixed Packages' that we knew we were in love.
"We're not asking small businesses to learn something new - we're saving them time and money by ensuring the system learns from and for them."
For the technically minded, the predictive model is trained on the last 200 invoices and kept for 3 days - long enough to not melt the servers, short enough to be responsive to new information. If an advisor has recoded any invoices, these are taken into account the next time the model is built. What this means for advisors is that it works best if you are active every few days initially - to ensure that the model is accurate early and guides business owners early as to best practices.
In cases where default accounts have been set up - Xero defers to them still. In time these defaults will give way to the machine learning models, simplifying accounting for businesses and their advisors. While the true impact of the technology will only be learnt when lots of small businesses start using it, testing of Xero's early machine learning implementations were accurate over 80% of the time by the fourth invoice, and consistently over 90% of the time by the 50th.
If Xero's proclamation of artificial intelligence changing the face of accounting was thunderous in resolve, the first look at machine learning in practice is a study of understated elegance. If you didn't know what you were looking at, you wouldn't even know it was there. And that's the beauty of it.
