We use cookies to personalise the website and offer you the greatest added value. They are, among other purposes, used to analyse visitor usage in order to improve the website for you. By using this website, you agree to their use. Further information can be found in our data privacy statement.

Automation in BPO: Process optimisation in Accounting – Part 3


published on 20 January 2022 | reading time approx. 7 minutes


Imagine you are in the process of renting a movie. Back in the days you had a DVD or Blu-Ray player and then you would go to DVD shop to rent a movie. The process would require your physical journey from your place to the renting company and back. You could optimize the process by going by car instead of walking, checking what options they have via phone in advance, is your movie available, what are other movie options. The DVD shop can offer externally placed return boxes that you can use outside of their business hours. This is process fine tuning – giving you the 10 to 30 per cent im­provement. But it wasn’t until the online streamlining service was introduced that the rules of the game changed, that you could suddenly dramatically improve the process by over 95 per cent. In every business you have incremental changes and these big, radical changes that are process disruptions changing the rules of the game.



With the introduction of AI in accounting we are more and more interested in the concept of the blind booking. What is the blind booking?
Blind booking is a situation, where an economic transaction is being recognized by the accounting proof that is automatically reflected in the chart of accounts. In our example the accountant will not touch or even see the invoice at all unless he or she explicitly wants to. The invoice will enter the system, have all of the required information extracted, all validation compliance checks performed and entered straight into the accounting software where it will be saved in the form an accounting record. This could be the next big game changer that we all look for. But imagine a world where invoices post themselves. It is so difficult to trust this system. There are all kind of different emotions in out head, especially if we are connected with accounting. Let us look at the new process step by step and try to understand it a little bit more.


1. Invoice enters the workflow system

That is the traditional step – documents are put into the workflow software. There are several ways this can be achieved, but let’s assume it was sent by email to a special customized email inbox to then be automatically transported to the workflow system.

2. Main classification decision

This step is already connected with an OCR software in place.

  • The OCR reads all of the information from the document
  • The Computer Vision algorithm recognizes all of the fields on the document and captures their position coordinates
  • Word Embedding makes a logical understanding of the document enhanced with data enrichment, related items position are being recognized in the form of a flexible graph
  • A Machine Learning model runs probability matches with an invoice meta characteristic model and specific domain sub-models. A recommendation is made what type of document it is – in our case it is a purchase invoice – and whether it is a duplicated document that is already in the system.
  • After clearing this the workflow route for purchase invoices is classified. 

3. Compliance validation checks

As explained in the beginning, there are some elements that have to be there for an invoice to be formally valid. Employees usually check for these manually and if a missing piece is to be found, the process of solving this problem starts. Usually, you would need to correct the invoice which takes more time to manage, write to the contractor and coordinate explanations. Instead, formal rules and Machine Learning can conduct this compliance entry check automatically and provide a direct feedback e.g. via email to the email address that the invoice was sent from explaining what the problem is and what information needs to be found. One example would be – “Tax identification number not matching with supplier information” with detailed explanation what is the Tax ID and what was found in the Public Tax ID database. There are usually several compliance checks, among which are:

  • Tax identification number
  • Supplier information
  • Supplier bank number validity
  • Recipient information
  • Invoice number
  • Invoice date
  • Invoice item
  • Sales date
  • Net amount
  • Gross amount
  • Tax rates
  • Gross amount calculation
  • Currency exchange

For each of these points a Yes/No check is performed. With some fields such as supplier information you could have different formatting in a database and in the invoice so you would use Machine Learning to give you the probability match then you could say if it is > 99 per cent that means compliance check passed. Our data shown 2.46 cent of invoices are corrections.

4. Accounting classification

When we have confirmed that an invoice is valid, we have to assign a booking entry for this invoice. We would therefore feed our Machine Learning model with a vector of numerical historical information to assign a booking account for this incoming type of economic transaction. But there is one more hurdle yet to overcome. And that is how many accounting entries should we create for this invoice? Sometimes, when the invoice is encompassing different topics, we would like to enter each of the line items as a separate booking. Here one good example would be an invoice from an office supplier company where we have plenty of different office supplies, cookies and cleaning materials. This invoice would usually have 3 booking entries.
Another example would be a fuel summary invoice where you would have many positions of fuel costs but also shield spray and spare parts. Normally accountants group these individual line items to put them all together for one account. Our Machine Learning is doing exactly the same. 

Analysis of past data shown that more than 80 per cent of purchase invoices is being assigned just with one accounting entry. That is telling us that there is a stable stream of purchase invoice documents that is simple in its nature of classification.
When we forecast how many booking entries should be created for an invoice, then we need to forecast a list of accounting entries from the most probable to the least probable for each of the line items and cut off the list by the number of booking accounts forecast in the previous step. Then we should group the booking entries based on the forecast accounts and voilà, we have a booking entry ready to be sent to our accounting system.

5. Transporting information to the accounting system

The last step would be to send the information regarding the booking entry to the accounting system and save it there. This can be done using 4 main methods:
a. Human interface (known also as the “protein interface”) which means that a person manually opens the workflow system, opens the accounting system and copies information from one system to the other.  This may be funny to do in this setup but if you look at Robotic Process Automation (RPA) solutions, this is exactly what’s being done with the difference that the copy and paste is not done by a human but by preconfigured software emulating mouse movements and controls.
b. Flat file upload – you can use e.g. excel files or csv/text files to download the information from the workflow system and then upload it to the accounting system. This is being done in the form of a file that users are familiar with and know how to modify. Therefore, when the excel file is being created and downloaded from the workflow system it usually has all of the booking entries created and the accountant can just upload it to the accounting system – provided the accounting system has this option of uploading and reading the interface file.
c. Direct database connection – you can connect directly to the database of the accounting software. That way, when the processing of your workflow is done, you can post this information regarding the new booking entries directly into the database. This approach is however found not secure. You should never use it in practice and always limit access to databases to a bare minimum.
d. Application Programing Interface – you can have access points on behalf of the accounting system that you use, to automatically inject booking entries in a safe way. The API gives you the controlled and usually well protected access to sending and receiving information which automates integration tasks 100 per cent.
If an API is used, the booking is injected directly to the accounting system where it is saved. So for how many documents would this new flow of “blind posting” be recommended and feasible?

When we looked at historical data, purchase invoices are being accounted to these main account categories:
0 Group includes fixed Assets accounts:

  • tangible fixed assets
  • intangible assets
  • financial fixed assets
  • depreciation of property
  • investments


4 Group includes synthetic cost accounts:

  • depreciation
  • material and energy consumption
  • outside services
  • taxes and charges
  • remuneration
  • social security and other benefits
  • other costs by type
  • settlement of costs

7 Group includes revenue, income and expenses accounts:

  • revenues and costs of sales of products, goods and costs of financial operations and other operating revenues and costs
  • taxes not included in account 403
  • grants and incentives received

Then we would need to focus on recommending:

  1. For the selected purchase invoice, how many booking entries should be made in the accounting system?
  2. What are the booking accounts that should be chosen for these entries?

If we look at historical data, we can see most of the documents result in one accounting entry. That is encouraging because we can see, there is a stream of documents that flows through the process with minimal complication. It is way easier to recommend a booking entry for an invoice that should have just one accounting entry. This is the “blind booking” opportunity for us.
Once we have this we can finalize the booking by recommending the booking account. If this is done with a >99 per cent confidence level, the booking entry is directly injected to the accounting system. If the confidence level is lower, the booking is directed for review. Then an employee looks at all of the assumptions and directs them. So far our Artificial Intelligence algorithms used in the production testing phase reached 78% accuracy for recommendations among 465 booking detailed accounts. 

6. Return of investment

Remember our video rental business idea? You can now imagine what we mean by applying process optimization vs. process disruption. 

The average booking of an invoice takes 4 min 36 sec and you are processing 20 000 invoices per month. Introducing a blind booking system with 60 per cent streamlining would let you save 920 hours per month, that is around 7 FTE’s which is an equivalent of 350 000 PLN saving annually.
But that’s not the beauty of this disruption. The beauty of it is that you can actually redirect these employees to be process experts and AI experts instead of dealing with repetitive tasks. It is the responsibility of every employer to support growth all of the engaged employees. The introduction of blind booking is a way to do so. After all, the accounting business is not an easy thing to do.
Read Part 1 and 2



NB: Graphics are all sourced from internal tools in use at Rödl & Partner.


Deutschland Weltweit Search Menu