Machine learning is handy, not holy
eConnect is known as a provider of the Peppol infrastructure to service providers. Less known is our solution for hybrid invoice processing. A few years ago, eConnect used other existing providers in the market for converting PDF invoices and receipts, but the quality was poor. Therefore, we decided to develop this ourselves! We focused on the ideal approach based on the latest techniques.
Machine learning is handy, not holy
Machine learning (ML) doesn’t solve everything and doesn’t make everything better. It is particularly good at recognising patterns in chaos, but it is less suitable for making absolute decisions based on structured data. Machine learning is not suitable for generating a booking proposal, but it is very suitable for classifying attachments (images) in an email.
The approach of eConnect
Classification of email attachments
PDF invoices are often submitted via email, where not all attachments in the email need to be processed. Sometimes it’s a reminder, and sometimes it’s an invoice with an attachment. The first step in the process is to classify the attachments using machine learning. Logos are removed, and invoices are converted. Afterward, it’s convenient for PDF specifications to be automatically merged with the PDF invoice, so no separate handling is required.
Transforming xml invoices
Not all XML invoices received via email are based on UBL, so eConnect detects which XML standard is used. We often also see errors in the XML invoices and can automatically correct them to ensure as many XML invoices as possible can be processed. Machine learning is unreliable for these techniques. We develop rule-based software specifically for correcting errors in XML invoices.
PDF conversion
The PDF invoices are first provided with text using standard OCR technology. The interpretation of the PDF image and the recognised text is done based on machine learning to produce a proposal with recognised data. Then it gets interesting! At eConnect, we don’t ask the user to check the recognition; the system does that itself. Every value is compared with an external source or analysis, so we automatically know which recognition is correct. This way, we only focus on values that don’t pass this test.
When a found invoice number is assessed, the pattern and placement must ‘resemble’ previous invoice numbers from this supplier. This ‘resembling’ (pattern recognition) is based on machine learning.
Booking proposal**
Booking purchase invoices requires knowledge. We have combined the knowledge of humans and machines into a powerful booking robot that always makes a booking proposal for the invoice based on this knowledge.
The first version of this robot was good enough to win the ‘ICT Accountancy Award 2017’ and provides a booking proposal for every invoice. The basic booking proposal is made in RGS and can then be further refined, for example, with your own schema. Over the past few years, this booking robot has been further developed under the guidance of a group of accountant users.
Stop accounting
The mission of eConnect is to achieve automated accounting for everyone. The technology has already compared the invoices with other sources or analyses, so it’s no longer necessary to review them again. All invoices are directly posted to the accounts, so outstanding items are updated—even if you’re on vacation. The administrator (or accounting firm) periodically reviews the figures and provides feedback to the rule-based software.
Independent
eConnect’s solutions operate independently of accounting software. We now have connections and integrations with more than 40 financial software packages. Thanks to the large number of connections and integrations, switching to another financial application is, in most cases, simply a matter of ‘flipping a switch’ in the eConnect platform. This prevents a lock-in with a software company, as switching between different software packages is very easy. It’s also possible to combine industry packages with generic financial software, for example, sending invoices from your own industry package and posting these invoices to the accounting system at the accountant. The solution’s functionality remains the same, and the robot’s accumulated knowledge is not lost. In this way, more and more organisations across various industries can benefit from an automatically posted administration.