ClearFacts chose IxorDocs to improve the recognition of incoming documents. ClearFacts used a self-developed algorithm for document recognition in its pre-accounting application. This algorithm was replaced by Ixor's AI model in October 2021.
We interviewed Gert Vrebos, CTO ClearFacts about digital invoices, Machine Learning and the exciting challenges the company faces.
"IxorDocs has now become a key component of ClearFacts. It replaces a central element within our application."
Gert, how did you end up at ClearFacts? "Before I founded ClearFacts together with partners, 8 years ago, I developed all kinds of custom software. We went through a number of iterations, and that's how we grew into the current ClearFacts company."
Do you have an elevator pitch for us? "ClearFacts primarily simplifies the administration of the accountant and the SME. We make the flow of documents as simple as possible and support their processing in the accounting package. We ensure time savings and above all increased efficiency."
How do you do that in practice? "Our application digitises documents by placing an OCR text layer on top of them and immediately extracting the most important data from the document. Based on histories and forecasts, we make an accounting proposal for the accountant. At the time we founded ClearFacts, accountants were still manually processing all documents. The financial reports were often delivered to the SMEs with delay. We wanted to optimise this process. We find it important that our solution is and remains low-threshold. Our customers can upload directly or send invoices via e-mail, and we also offer a link with Peppol. the list of input channels is long. Our solution includes integrations with various invoicing packages, so that invoices can be delivered directly to the accounting system."
What role does Ixor play? "IxorDocs has now become a key component of ClearFacts. It replaces a central element within our application."
You mean the creation of the text layer? "Well, the creation of the text layer is a commodity. OCR has been around for years. But automatically recognising individual data on invoices such as invoice date, total amount, etc... is complex. Recognising this information automatically and extracting it correctly, that's real invoice recognition. We developed an algorithm for this in-house. That algorithm has functioned well for years, but it was time for an upgrade."
You sound almost nostalgic. Did you build that yourself? "That algorithm was one of the strategic components that I indeed developed myself. I maintained the algorithm until it was replaced by Ixor's invoice recognition algorithm. Our algorithm was rule-based. It was a tremendously complex system, very powerful, but fragile. It contained so many rules that adjustments or extensions became increasingly difficult to realise. Our algorithm could not evolve and be adapted automatically. That is why we looked for machine learning solutions. Such solutions can easily evolve and adapt to new invoice templates."
Did you search for a long time for a solution before you came to Ixor? "After an extensive preliminary process, we went ahead with Ixor. The recognition component is a crucial part ofClearFacts. Currently 50,000 SMEs use our application, some of which are large companies with sensitive files. All documents must be processed correctly. Our information to the accountants is built on recognition. That has to be absolutely right."
Why was the preliminary process so tough? "Because we developed the original recognition algorithm ourselves, we have a very valuable benchmark. The results of the potential solutions were compared to that benchmark. Initially, we looked at off-the-shelf products, the usual suspects, including international solutions. Several solutions achieved good results compared to our benchmark, but none matched our benchmark on Belgian invoices."
Then we are back to recognition quality? "Invoice recognition is language specific. The other solutions achieved good results on English-language invoices, but the results on French and Dutch invoices were disappointing."
A local specialist was needed. "Absolutely. The results of the IxorDocs recognition algorithm were the closest to our benchmark. Afterwards, we worked hard with Ixor to bring the quality of recognition up to the level of the original rule-based system."
When people at Ixor talk about you, they usually mention quality and feedback in the same breath. Can we already start calling it quality control as a service? "Benchmarking recognition quality is certainly not an easy task! For us, it is crucial that the new algorithm achieves the same quality as our initial algorithm. Before we took the plunge, we wanted to be sure!"
We had big shoes to fill. (Laughs) "You may rest assured! 50,000 active customers, millions of invoices... we wanted to ensure that the quality of recognition at least reached the original level. And we have succeeded. The recognition is already better than before the switch."
So fast? "So fast! That was one of the reasons to switch to a Machine Learning algorithm. The promise there is that the machine learns. Our customers expect our system to process corrections immediately and without error. With IxorDocs we can automate the quality control through the feedback loop. In addition, we have the option of adding custom requests ."
Even more scalability for ClearFacts. "Scalability and worry free! We can now focus on the further development of ClearFacts and the services we offer. Resources are also freed up in terms of infrastructure, since IxorDocs is a native Cloud solution. Recognising extra fields, breaking down VAT percentages, etc., are all now realised by IxorDocs. The additional possibilities offered by IxorDocs have convinced us to continue with the Ixor algorithm."
If I understand correctly, in a few iterations of our product we are at the same level as your old system, and each iteration that comes now brings improvement. How do your customers react? "Not! Amazingly enough! The little feedback we received was mostly positive. We have phased the changeover so that we could react quickly in case of negative feedback. We first switched to the Ixor algorithm for the purchase invoices. In a subsequent phase, we activated the Ixor algorithm for the sales invoices."
Is there such a big difference? "The difference between the two types of invoices is immense. All sales invoices for a certain customer are the same. But the purchase invoices exist in as many versions and layouts as there are SMEs. That is why we first switched to IxorDocs for the purchase invoices. When a purchase invoice is not recognised correctly, it hurts once. When sales invoices are incorrectly recognised, yes, the pain is much greater. In a second phase, we therefore switched to sales invoices. We have informed our customers of the switch to a Machine Learning system. They are also aware of the monthly updates released by Ixor."
You switched just before the VAT quarterly report for Q3 2021. We found that exciting. "We certainly found that exciting! But we were convinced of our decision. The quality of recognition was high and the result of the stress test was promising. The switch went smoothly! Afterwards, we hardly received any comments."
What are the next steps for ClearFacts? "We are planning another extension for custom keywords, for a limited number of our customers. The tests are going well, so I do not expect any issues with the roll-out of this extension. Afterwards, we want to tackle the recognition of new fields and the breakdown of VAT percentages as quickly as possible. The latter will be a crucial step.
Our final question: where do you think the market as a whole is headed? "I expect the move to a fully digital invoicing process. But we are not there yet. The legislation lags behind what is technologically possible. In Italy, they are already further along, but I suspect that the Belgian legislator will wait for the European legislation. Of course, the digital invoice puts pressure on us. But with our hyper-focus on the niche of pre-accounting and financial administration, ClearFacts will certainly be of value to accountants and SMEs in the years to come.
Would you like to test our models? Are you interested in a customised invoice recognition model?
