Tax codes: Reducing administrative costs with AI

Consequently, applying tax codes to your incoming invoices means selecting from countless options. Over 60 different codes can be in use at a medium-sized company with numerous relationships in Europe and the USA. Why so many?

Firstly, due to the different tax rates. Rarely there are 2 countries with an identical scheme (e.g., in Germany 0%, 7%, 19%, in France 0%, 2.1%, 5.5%, 10%, 20% ...). Fact: The 27 EU countries have 10 different standard VAT rates and 18 different reduced VAT rates (if you include 0). For number theorists: Of the natural numbers between 0 and 27% (Hungary being top in VAT) the 1, the 16, and the 26 are missing). Instead, there are some fractional numbers, e.g., France stands out with a sharply reduced tax rate of 2.1 percent - someone must have calculated for a long time.

However, it is not only the different tax rates that lead to an abundance of tax codes. Secondly, there are numerous legal requirements for reporting, allocation according to local and EU tax law, and other aspects such as refund requirements that must be considered when posting. Even a 0 percent tax rate can be ambiguous (small business, EU import of goods, reverse charge). In many countries, the type of spend must be considered when recording input taxes . In Switzerland, for example, a distinction is made between "cost of materials and services" and "capital expenditures and other operating expenses." Several different tax codes are then needed for the same tax rate.

Minor issues can multiply the effort

Tax codes are important. They are part of the central master data in the SAP system and enable input taxes to be posted directly to the appropriate tax accounts. Subsequent processes depend on the correct assignment, in particular the automatic processing of payments.

So where is the problem? The administrative jungle of tax codes makes automation difficult. It is true that the same tax code can be automatically assigned for all domestic invoices with the standard rate of 19% - easy. These tax codes can be easily set based on a rule, as the supplier country, invoice recipient country and tax rate can usually be captured automatically.

As soon as simple basic rules like this are no longer sufficient, traditionally programmed solutions reach their limit. In many countries it is almost impossible to avoid manually entering the tax code for each invoice. But once manual entries are necessary, all subsequent processes are now delayed. While accountants may be able to make the tax code decision quickly because of their knowledge and experience, every manual action can cause a significant delay:

• The invoice twiddles its thumbs for a while until an accountant the the open task in their backlog.
• Then: open invoice, understand, decide.
• Enter tax codes.
• Once the invoice needs a manual touch, typically, the next step will also need to be done manually (e.g., send approval workflow).

With automatic tax code assignment, the approval workflow would have started days before and the accountant would have been saved some work.

The AI that cuts through the administrative jungle

Naturally, it made sense to us to use AI to overcome the limitations with automatic assignment of tax codes . We have clearly succeeded with avvaneo Expedite. Our machine learning solution, which is constantly getting smarter, achieves measurable increases of over 90 percent when finding tax codes automatically from day one.

Our AI model was trained with the structured invoice data (of course) which is extracted automatically, as well as the whole invoice, with all the unstructured information that can be captured from it. Based on this historical data, the AI then detects the statistically highest correlations between all the data/information present as well as the tax codes assigned. Given enough invoices, this leads to reliable decisions.

To illustrate this with an example: Whenever ...

1. according to the structured data the country of the buyer was Germany,
2. the invoice showed a Belgian supplier,
3. a note like "Reversal of tax liability" was written on the invoice (in German, French or English)
4. and the tax rate read was 0 percent ...

... the tax code A4 was used. And that was correct.

An AI solution provides significant advantages over completely rule-based programming. Rule-based programming requires that all the information in an invoice must be available as structured data. This places very high demands on the capture software. Creating added sources of error as well as more need for manual corrections, simply because more fields must be captured.

Crossing tunneled, manual effort removed

As mentioned, the context of the complete processing of an invoice is important. With avvaneo Expedite, we provide this context for our customers to the extent that even more complicated invoices can now be processed completely automatically.

• avvaneo Expedite can set virtually all the more common tax codes, automating this step.
• At the same time, we also automate the preceding and subsequent steps.
• Result: fewer and fewer invoices require manual processing by an experienced accountant.

You can imagine the acceleration effect like a tunnel in the city. Where previously the flow of traffic was stopped at intersections, now everything flows through without congestion.

Even better, sloppy mistakes can also be highlighted.

Manual entries are never error-free. Our AI solution gives you the added advantage that manual entries are automatically validated by the AI, as if a second person were double checking the results, just to be on the safe side.

An example of where errors easily occur: In Germany, Romania and Cyprus, the standard VAT rate is 19%. If you mainly process German invoices, you could easily assign the VAT code for Germany to an invoice from Cyprus out of habit. Our AI solution will notice such discrepancies immediately. This happens during anomaly detection, which when the solution is implemented is always watching. Anomaly detection will compare the manually assigned tax code with the automatically generated proposal. If there is a discrepancy, an alert appears before the document is posted. This provides a significant advantage, as correcting errors after the fact is usually a particularly time-consuming task.