Markus Mantere argues that for the professional accountant, self-learning technology means an opportunity to turn expertise into a scalable asset.
A technological shift is taking place that will have a positive impact on the daily lives of millions of finance professionals. The shift is about software that can learn without being explicitly programmed from its surroundings, and turning those lessons into value for people.
Let´s start by stating the obvious. Buzzwords are a thing in the world of accounting technology. Words like “robots” or “AI” risk causing confusion since they are used in different ways with different meanings. Therefore, let me define what I mean by self-learning technology.
Computer programming has for long been all about giving computers strict directives such as ‘if X occurs then perform action Y’. An example of this in an accounting software context is being able to connect a certain tax rate to a specific supplier. Getting the tax automatically calculated based on the value you key in the total sum field works well some of the time, but gives you the wrong result other times. It is predictable in its shortcomings – it will stay the same.
Software that learns, on the other hand, can change and get better with new information. Instead of having one strict rule to follow like a slave, learning software can handle a much more complex reality. In the example mentioned, it could learn to handle multiple tax rates on one bill, splitting an invoice on multiple accounts, just like a human would. I will explain how that works in practice later in this article.
World-leading companies leverage learning technology
Let’s start with an example from the modern world of business. Tesla aims to create autonomous, self-driving cars. The theory on how they use learning technology to reach that goal is rather interesting.
A car is packed with cameras and sensors, observing the environment surrounding it. As the car approaches a bridge, the frontal cameras pick up the bridge as an object ahead. It is difficult for a computer to understand if this object is above the road or on the road.
This is where it gets exciting. By simply observing the driver's choice of actions when approaching that bridge, Tesla learns that the object was perfectly safe to drive under. Hundreds of thousands of cars are on the roads all around the world, and this fleet of students are laying the fundament for self-driving vehicles by learning every day and uploading these lessons to Tesla´s central, collective intelligence.
Another interesting example, in a way closer to bookkeeping, is Netflix. Your account is personal and Netflix aims to give you good recommendations.
An old-school approach would be settings where “female, aged 40-50" would give you a recommendation for something they believe that group wants. If your taste in movies would differ from your group, you´d get lousy recommendations continuously.
Instead, Netflix learns. By simply observing what movies you watch, Netflix gets better at giving you highly individualised recommendations.
Learning technology can solve bigger problems in more complex realities. In the case of Netflix, much more granular categorisation is possible. Who would go into settings of their movie app account, scroll thousands of categories and manually choose “Deep sea horror movies” as a category they would want more recommendations from? And yes, that is an actual category; that is how granular it is.
Self-learning technology in bookkeeping
So, people choose movies in an app and that app learns from their choices. What about bookkeeping? Many have tried but the difficulties are plenty.
Developers with skills in AI are few and expensive. Lots of existing tools crave huge amounts of data but offer unpredictable outcome. The processing of big data sets takes time, costs money and is even harmful to the environment.
Perhaps the biggest problem is teaching a computer stuff that a person knows by simply looking at an invoice and knowing a client, like slight differences in the chart of account or granularity of bookkeeping.
Magical wands, Mechanical Turks and forced standards
A lot of solutions that claim to use magic wands (substitute this for the hype word of your choice) can be put in two broad categories. One talks more than it walks, and the other has a rather narrow capability.
The original Mechanical Turk was an 18th-century chess-playing machine. The con was simple: under the table with the chess board on it, a person was hiding. Perceived as an autonomous chess-playing machine, but with a human touch. Sounds familiar?
Another way of reaching high levels of automation is with forced standards. There have been countless bookkeeping startups claiming to automate everything with magic wands. But they often end up offering very niche solutions to a small customer group that adapts to the reality needed for the magic wand to work.
Basically, magic wands are often about moving a workload to another human or the strict application of specific – often simplified - standards.
Self-learning bookkeeping algorithm
For the past five years, a team of developers at Arkimera Robotics, a deep-tech company based in Sweden of which I am deputy managing director, have been working on a technology for adding self-learning to bookkeeping.
The description of self-learning technology in this article will focus on this solution, not as a way of self-promotion, but to demonstrate both the practical application and benefits of self-learning technology through a worked example.
We created an algorithm from scratch: not simply a model created with machine learning tools from one of the big tech companies. Hence it is perfected for one task: bookkeeping.
This technology can be described as an autonomous model builder capable of learning and automating data entry in a highly individualised way. While it gets better with more data, it is not dependent on vast sets of data to start performing well.
Learns by observing you – right where you are
The algorithm does not function by itself; it needs a home and it needs food.
A suitable home is any accounting software out there connected to the internet. There is no standalone app, no user interface. It is a piece of technology that needs to be integrated into some data entry software. Once integrated, it observes the human's data entry, learns and can then assist the human by doing it for her.
It also needs to be fed. Tesla can´t do much with information about a human's actions if they can’t see what the human is reacting to. The algorithm is fed with two pieces of information: what you bookkeep and how you bookkeep it.
What you bookkeep is, in this case, a file, such as a PDF invoice or a smartphone image of a receipt. How you bookkeep it is the different words and values you enter in fields in the data entry software. These are the lessons that the algorithm gains its knowledge from.
Learning from the first lesson and onward
The algorithm learns from very few interactions with data. With no prior training data, it learns from the very first categorisation of costs to the second. And onward. This rapid learning is necessary in order to pick up the slight differences between an accounting or bookkeeping firm's different customers when it comes to coding accounts and level of granularity.
It also makes it possible for smaller firms with less data to leverage learning technology, or for that matter for accounting software companies to do the same.
The importance of isolating knowledge
The rapid learning leads to a very important factor in bookkeeping: the isolation of knowledge on a company level.
A purchase of the same product from a certain supplier can be categorised differently depending on the purchasing companies’ use of the product. One customer´s desired granularity in the chart of accounts compared to another is another example.
The algorithm contains the knowledge on a company level, understanding the slight differences in the data entry between companies.
This collection of knowledge on how to bookkeep for a company is referred to as a robot.
Controlled knowledge cloning
Understanding the slight differences between companies is important. But sometimes a bookkeeper would also enjoy making use of the similarities.
Controlled knowledge cloning is a method to create progenitors. One progenitor gained knowledge on how to bookkeep for one company. This knowledge can be inherited in the blink of an eye by a new robot. This new robot could be the bookkeeper´s assistant, within the accounting software, in bookkeeping for a new customer in the same industry as the progenitor robot.
The similarities in the bookkeeping will be more than enough to clone the knowledge from the first robot, the progenitor. Since the learning is rapid, and the robots are isolated from each other, the bookkeeper can also make slight changes in the new robot – simply by bookkeeping differently.
No settings, just business as usual
How we self-learning technology working from an end-user perspective is very easy to explain. You keep working in the software you already are familiar with.
You enter words and values in the same fields as you did yesterday. The only difference is that the system of record has been turned into a system of results; the lessons you teach the software are returned in the form of automated data entry – in exact accordance with how you would have done it yourself.
If technology can learn, a good teacher makes a good product
Software companies have put a lot of effort into making bookkeeping apps so simple that anyone can bookkeep themselves. This comes at the cost of data.
With a self-learning algorithm doing the data entry in exact accordance to how its human bookkeeper teaches it, there is no longer a need to standardise or simplify. The algorithm takes the same number of milliseconds to do data entry on many accounts as on one.
With the mega-hype about advisory, or the scare tactics about robots stealing jobs, how about instead envisioning a future where accounting firms teach highly individualised industry robots and rent them out to companies on a subscription business model?