How can the accuracy of invoice automation artificial intelligence be tested before the purchase decision?

Utilizing artificial intelligence in automating the purchase invoice process is a relatively new thing. The use of artificial intelligence is becoming more common all the time and its benefits are clear, but the truth is that the performance of artificial intelligence varies from company to company.The performance of artificial intelligence is affected by many things. For example, the quality and consistency of a company’s purchase invoice data, the complexity of the dimension structure, the number of individual dimension values ​​(e.g., the size of the chart of accounts), the consistency of the accounting history, the number of invoice auditors, and so on.

At best, almost 90% of our customer’s invoices are correctly posted for using artificial intelligence, and at worst, the level can be 50% – 60%. However, prediction accuracy is constantly improving as purchase invoices flow through artificial intelligence. Although our pricing is based on successful predictions, our customers are very demanding in terms of accuracy.

Testing artificial intelligence as a new customer

We constantly meet with companies that are interested in using artificial intelligence to automate the manual posting and routing of purchase invoices. We often hear the headline question: “Can we test the artificial intelligence with our company’s data before making a purchase decision?” Our answer is, of course, you can.

If a company could benefit from artificial intelligence, we will do the trial before they need to make a purchase decision. This means that the customer provides us with a data set that includes copies of the history of purchase invoices (XML) and the corresponding postings for a period of approximately 3 months to 12 months. Invoice data can be obtained either directly from the purchase invoice processing system or by requesting it from the system supplier.

We use about 70% of the data to build an artificial intelligence model for that company. The result of this is an artificial intelligence model that is specialized in automating the purchase invoice process for that company.

After this, we give the remaining 30% of the purchase invoices to the AI ​​model as “new purchase invoices”. The task of the artificial intelligence is then to predict the accounts for these invoices as well as the purchase invoice auditor. In the end, we compare the predictions made by the artificial intelligence with the original postings as well as the auditor data. This allows us to verify what the level of Snowfox.AI prediction accuracy is for the customer.

Trial is used to determine prediction accuracy and costs

Trial results are always reported very comprehensively with a visual report. We show e.g. analysis of purchase invoice mass, dimensional prediction accuracy, invoice prediction accuracy, and dimensional confidence value distribution by invoices.

Based on the results, we are able to provide an accurate price estimate for the use of the service. Our goal is to show the customer how much Snowfox.AI is able to increase the level of automation in the purchase invoice handling process, ie reduce manual work and how much it costs to use the service. This way, there is no need to buy artificial intelligence without validation that it actually works.