Extending Contract Accounting with Artificial Intelligence

The following table lists out the central concepts of Machine Learning training.

Adding to what we learned previously about the SAP Cash Application and the Machine Learning concepts, this figure shows the process steps in more detail.

1. Payment lots are created from bank account statements which we receive typically daily. These lots contain the detailed payment information and will be processed in Contract Accounting based on a background job. Remittance information, as provided with the payment, will be submitted from the bank statement and will be captured in the payment lot. This might be for example, a reference number or invoice number. Based on this information, payments will automatically be posted in FI-CA and assigned to a suitable open item based on the defined logic in clearing control. These rules already cover a complex logic and consist of various steps.
2. Items which cannot be cleared based on this approach, end up as clarification cases. In the past, these cases have been clarified based on manual analysis. Depending on the information available – or not available – this has often been a complex process with time consuming process steps and individual decisions.
3. Now, with SAP Cash Application these items can be sent to the Cash Application tool for further processing. The Machine Learning logic will be applied and tries to find a suitable matching proposal.
4. Together with this proposal, a confidence level will be provided. Based on this level, the clearing document can either automatically be posted or it can be posted and displayed for further checking by the AR accountant. If the confidence level is even lower, the clearing can just be proposed and the accountant would initiate the clearing, or if the confidence level is quite low the proposal will be skipped and not shown at all.

All these confidence levels are configurable and can be defined according to the individual requirements. Only items without any matching proposal remain as a clarification case for manual analysis.

Now let’s have a closer look at the training process of the SAP Cash Application.

• In the SAP Cash Application training process, two models of the matching criteria based on historical data and on automatically cleared items is built. The historical data is used during the training process to create two models that are later used by the SAP Cash Application when in a productive environment.
• In the training process, Machine Learning can identify patterns in knowledge-intensive processes, without explicitly defining the patterns by rules. The SAP Machine Learning engine learns from historical transactions during an initial training period. With further training periods the model gets adapted in case of changed system behavior.

Inputs to the training process are the following:

• Historical information from electronic bank statements, note to payee information or payer bank account information. In addition details of the payment lot items, including selection categories and selection values are used. Further manually and automatically cleared selection categories and selection values are included.
• Also, master data information, such as business partner details, contract account details or bank master data will be used.
• Customizing data such as selection category and selection field names are inputs to the training process as well.

During the training process, two models are built; one model called GLIM model and a second model called BER model. Both models are then used in the inference process in parallel when the SAP Cash Application process is live in production. If each model generates a proposal, a final proposal is assembled in FI-CA post processing. To be able to start the training process sufficient volume of historical information is needed as input. At the moment the SAP Cash Application requires around ten thousand historical payment items at least in one company code per country in order to be successfully trained.

During the training process, the Machine Learning platform then automatically selects and arranges decision criteria and values. Based on these criteria, the Machine Learning platform develops a model for the payment assignment.

To achieve a good model, which is one that predicts well, it must learn implicitly through the provided data. The provided data must not only contain perfect match cases, but it should contain imperfect matching cases as well. This is necessary for the Machine Learning to learn from these imperfect cases to make a correct prediction when it encounters similar errors during the production inference executions.

The SAP Cash Application needs to learn the correct behavior from both the perfect matches as well as those with imperfectly provided data.

After the Machine Learning training process built up the model for matching the payments, it is necessary that the model gets evaluated as well. Therefore, the historical data with the different bank statement items, master data and selection categories can be divided into subsets of data. The bigger subset of data is used for the training process itself – here called training set and the smaller subset of data – Testing Set - is used to test the trained model. With this approach it is possible to evaluate the correctness of the model as you can directly compare the proposed matching with the real, historic matching that took place in the system already.

This evaluation process consists therefore of the following four steps:

1. Split the historic data into a training and a testing subset
2. Train the Machine Learning model with the larger training subset
3. Execute the inference process for the testing subset and receive the results
4. Compare the matching proposals to the true clearing that happened already

We have been using computers for decades now. We are used to seeing the computer take over certain tasks which were done by humans.

There is one major difference between a human and a computer in that humans can adapt to changes, and we constantly improve ourselves because we can learn. We learn from experience. Computers on the other hand have to be explicitly programmed. The software consists out of a chain of commands, which only work for the predefined situations. Therefore, humans can handle things that were not pre-thought of by the developer who created the coding.

This is now changing with Machine Learning. Just imagine a computer that could learn; not learn from experience but learn from data. That is the idea of Machine Learning that opens totally new doors for us.