The processing of purchase invoices has traditionally been a very manual and labor-intensive process. The most time-consuming tasks in the processing of purchase invoices are manual coding and routing the invoices for review and approval. Often, the actual physical clicking work is not the most labor-intensive part; rather, it's the thinking work behind it. When a new invoice arrives, it need to be determined who should review and approve the invoice and how the invoice should be coded according to the company policy. Naturally, the more purchase invoices received, the greater and more complex the work becomes in their processing.
Accurate account coding is crucial in order to properly manage taxes and accurately report the company's expenses. Any errors in coding invoices can have detrimental effects on VAT payments and cost reporting, making it imperative to ensure precision in this process.
For these reasons and more, an increasing number of organizations have decided to incorporate artificial intelligence into their purchase invoice processing. As an example, in the spring of 2023, the City of Helsinki and the Evangelical Lutheran Church of Finland made the decision to incorporate purchase invoice AI into their processes, ultimately selecting Snowfox.AI as the provider of their innovative solution.
Over the years, organizations have attempted to achieve automation solely through manually constructed rules. Contracts have been created in the system for recurring invoices. Templates for coding have been developed for specific suppliers, order matching has been implemented, and so on.
As a result of diligent automation work, functional solutions have been built for some purchase invoices. However, automating a large volume has almost always remained within the realm of manual handling, despite diligent efforts. I have written a fairly comprehensive blog about this previously, which you can access here.
The construction of automation rules usually starts with the easy part, often with invoices that have the same repeating patterns. Rules are fairly easily established based on these, enabling, for example, the automatic coding of regularly recurring rental invoices in a predefined manner.
In practice, every purchase invoice system offers the possibility to build rules, and this is well-known to everyone. The challenge with rules is that there are surprisingly few truly identical recurring invoices. Most invoices do not repeat in the same way over and over again; a significant portion consists of various miscellaneous invoices. Automating these using traditional methods has proven to be very challenging.
The question then is how many identical invoices would need to be received on an annual basis to make it worthwhile to build automation rules in the purchase invoice system. Automation rules also need to be maintained; otherwise, your system will be filled with outdated rules that no longer apply in the present day.
Artificial intelligence as an automation tool significantly differs from traditional rule-based automation. While traditional rules need to be individually constructed and maintained, AI builds its automation rules entirely independently by utilizing historical invoice and coding data. AI is usually trained using about a year's worth of processed purchase invoices, including their coding and reviewer information.
The AI goes through each of these invoices one by one, after which it retains the content of each invoice and the information about how the invoices were previously coded. During the training phase, the AI utilizes every field on the invoices to gain a thorough understanding of their content and coding. Snowfox.AI's AI also reads the images on the invoices and uses the image data for learning as well. This is why our AI service works with PDF invoices as well, not just e-invoices.
After the training phase, the AI has comprehensive knowledge of invoice coding and routing, allowing it to automatically handle newly arriving purchase invoices.
The trained AI service is integrated into the invoice processing system and functions as follows. When you receive a purchase invoice for which there's no existing automation rule in the system, the AI takes care of coding it and routing it for review. The AI first examines the entire content of the invoice and then makes decisions based on that content about how each coding field (such as account, cost center, VAT) should be selected, and to whom the invoice should be sent for review and approval.
Once the AI has provided a coding suggestion for the invoice, it goes through the normal process of review and approval. If the AI has suggested an incorrect cost center, for example, the invoice reviewer corrects it in your purchase invoice system.
Once the invoice has gone through your standard approval workflow and is ready for accounting and payment, the AI receives information about the final coding of the invoice. If any dimension value (e.g., cost center) was changed during the review process, the AI receives feedback that the value was incorrect, and it learns from this.
The AI uses the coding information from each completed invoice as further training data, constantly learning and improving itself automatically. As a result, you don't need to manually update the AI like you would with rule-based automation; the AI learns automatically.
AI performance is monitored and evaluated using advanced analytics. The analytics show for each predicted dimension value whether it was correct or if a human needed to make adjustments. You can easily track how many invoices the AI has accurately processed without human intervention. The AI's progress over time can also be easily tracked through analytics.
Our goal from the beginning has been to make AI's operation as transparent as possible, so you don't need to guess its capabilities. When you can monitor performance through numbers, you can feel confident, and both capabilities and errors become visible.
The purpose of the AI is to streamline, simplify, and expedite the processing of purchase invoices, while reducing the amount of manual work. Each purchase invoice automated by the AI translates to less time spent on manual processing.
Our pricing model is built on this principle. Our clients pay only for the predictions made correctly by the AI (e.g., account, cost center, VAT, project, invoice reviewer, etc.). If the AI makes an error, for example, regarding the accounting account, our clients won't be charged for it. Our promise is also that each purchase invoice automated by the AI is significantly more cost-effective than manually coded invoices.