Intelligent Invoice Scanning enhancements 2105
In this blog post I will have a closer look at the new ‘Intelligent Invoice Scanning’ enhancements, introduced with release 2105. For an overview of the ‘Intelligent Invoice Scanning’ please also have a look at the following blog posts:
Already with release 2102 we have started to incorporate new features and round-offs for the intelligent invoice scanning solution, that was introduced with release 2008. You can find the details here:
Now, with release 2105 we continue this by adding a set of new capabilities:
- Automatic synchronization of supplier and company master data. Before release 2105 it was necessary to manually synchronize supplier master data in case of scanning-relevant changes. We now run a daily job that uploads the supplier and company master data to the Document Information Extraction Service. Once the ‘PDF-based Invoice Scanning’ feature is scoped, a background job will automatically be scheduled at 2.30 am. The customers do not need to take care of the scheduling of the job. Nevertheless, the synchronization of supplier and company master data can still be triggered manually in addition, if needed. It might take up to 4 hours until updated master data are reflected in the scanning result.
- Collection of corrected scanned values for retraining. The PDF information extraction is based on a pre-trained model. Pre-trained means that SAP has used a set of invoices to train the model upfront. The key benefit of this approach is that there is no need for customers to first train the model upfront before reasonable scanning results are achieved. To further improve the quality of this global pre-trained model we now allow users to send corrected values back to SAP. To activate this the question ‘Do you want to train the ML model with the user values via feedback integration’ in Business Configuration needs to be answered:
Once this is set in scope, a new flag is available in the review UI to let the user confirm, that the review document does not contain any personal data.
This step is needed in order to stay compliant with the EU GDPR regulation, because we can only consider those invoices for retraining of the ML algorithm that don’t contain personal data.
Now, any corrected ‘Reviewed Value’ is sent to the Document Information Extraction Services.
Here we apply another check for personal data using an algorithm. In case any personal data are detected the document is immediately removed. Another post processing step is the UI annotation before the corrected document can be used for retraining. With the post-processing work by SAP, currently retraining of the model takes three or more months. In addition, if only a small set of invoices get corrected, the re-training effect might be small, because the model got trained with a huge number of invoices. We will investigate further in the upcoming releases, if this post processing duration can be reduced, e.g. by providing an annotation UI. Since SAP has already used a high number of invoices for retraining, an individual invoice correction might only have a very small impact on the model. Nevertheless, over time we expect an overall model improvement for all customers based on the re-training.
You can monitor the re-training progress with the newly introduced ‘Retraining Status’ value in the ‘Invoice Scanning’ work list.
- Creation of Credit Memos from scanned documents. Many customers requested to also allow creation of a credit memo from a scanned document. With 2105 we now introduced a radio button on the Review UI.
A user can now select to create a credit memo using the ‘Create Invoice’ action. A pop-up will allow users to optionally enter a supplier invoice ID, in case the credit memo references to a supplier invoice. Also, the purchase order ID will be used to identify if credit memo items can be mapped to a purchase order item. This will also be used to enhance the credit memo with additional information like the account assignment information.
Please also have a look at the video.