Why we are doing AI based document extraction

When we started building our technology for the autonomous processing of accounting tasks 2 years ago, we thought that the extraction of accounting documents was something that the industry had already solved. As we realized, this is not the case.

(Reading time 3 minutes)

When we looked at the available solutions in detail, we quickly found out that a really good extraction quality can only be achieved in a very cumbersome and laborious way. And that many software solutions are expensive to procure and maintain.

One template per document type

Excellent extraction quality can only be achieved with conventional OCR technology if a template is created manually for each document. Unfortunately, this circumstance prevents truly autonomous processing of documents.

For this reason, we started very early with the development of a Machine Learning Cluster which can process documents without any manual intervention. At the beginning of our development, we realized that it was significantly easier than we initially imagined. But we quickly learned that it is a complex, marathon-like task. And, once into the process we learnt a lot about accounting documents.

Looking at documents like humans do

Our approach is fundamentally different from that of the well-known OCR technologies. Basically, our technology looks at documents like humans.

If, for example, we humans receive an invoice that was issued in a foreign language, we are able to recognize that it is an invoice at all. We are also able to sort of intuitively find relevant data points such as date, payment date, address etc.

This is possible because we usually have a lot of experience in dealing with these documents (if you haven't worked as, say, a gardener all your life). At first, we visually capture the document and can classify it according to the rough scheme. We don't have to read the whole document in detail to know where the data sits.

In a next step we are looking for values on the document that belong together and only then do we read the individual values in context to each other. It’s a very efficient process and methodology to get the job done.

Our technology basically does exactly the same.

Learning from people

The machine learns particularly well when we humans help it to learn. Show it which datapoints are important, what is right, what’s wrong. This is no different with accounting documents.

So we create secure, highly accurate, “resilient” data images from the documents which we validate and correct manually. We use this set of perfect data to train the machine. It is important that the literally every comma and every dot is correct. We quickly realized that it makes a lot of sense to capture documents with an enormous amount of detail. On the cost side, that comes with huge headaches.

Many people in the machine learning sector make the mistake of including a lot of data material, but paying too little attention to ensuring that this data material is of high quality. Only to be left surprised that the output is not of superior quality.

The solution to the extraction of accounting documents is the first important step towards autonomous accounting systems.

There is a simple reason why we deal with the extraction of accounting documents: We cannot skip doing it if we want to implement the autonomous accounting engine. The automatic posting of cash account transactions and the reconciliation of this stream with the receipt data is a comparatively simple task, but simply impossible without perfect receipt data.

Unfortunately, it is also not foreseeable that the electronic invoice formats would relieve us from this step within a reasonable period of time. In no future format will line items, i.e. invoice items, be mandatory. This means that, although we receive the standard information in a structured way, the details relevant for accounting are not provided in a structured form. Ironically, the standard information like vendor, dates, payment information etc are the easiest to extract.

It is time to radically reduce manual work in the accounting department. This has been clear for a long time and it seems that the industry has slowly but surely woken up and is increasingly working in this direction.

But cost and efficiency gains are only the first step, much more important is the creation of the basis for a future accounting that is casual and with a then radically improved data basis enables completely new decision data to entrepreneurs and managers.

We see Robo-Accounting in this context. To break new ground in accounting in the long term. In order for this to be possible, the supposedly banal, such as the autonomous extraction of data from paper documents, must first be solved. That is why we are into this.