ANTONIA - Automatic order processing

Quick Check

Initial situation

The Postbeamtenkrankenkasse (PBeaKK) has been providing benefits services for civil servants and pension recipients of the former Deutsche Bundespost for over 100 years and, since 2017, also for other contractual partners.

All core processes at PBeaKK are semi-automated and workflow-supported. The correct process type is selected on the basis of the classified incoming documents. If classification errors occur, the template must be changed. The orders for the change are made as a text note, which is interpreted and implemented at a central reworking point. Approximately 2,000 orders are processed by this unit every day.

To date, the cash register has used a prototype ("ANTON"). This has proven itself in the classification of simple cases, but cannot handle complex texts and is complex to maintain.

Solution idea

An intelligent system can interpret the text notes and automate the basic correction process. The cash register has stored many months of data on the start and end characteristics of each order including the text note, process type and associated document types. This can be used to train a machine learning model for flexibility in interpreting the text notes and automatically classifying the correct changes.

AI_Progress_Center_204_QC_ANTONIA_Abb1 — Figure 1: Approach to the automatic classification of orders and initial results based on representative data, own illustration

Benefit

The average workload of a full-time employee is a maximum of 200 orders per day. The personnel costs for reworking are disproportionate to the benefits, as very standardized orders have to be processed in some cases. With partial automation of the work, the employees can be deployed in other fields of activity.

In the future, existing patterns will also be recognized during reworking and taken into account directly during the initialization of the process flow.

Implementation of the AI application

The quality of the commissioned texts varies between the individual employees. The texts also contain specific terms that are used internally. This requires a system that takes into account the possible combinations of the text with other start parameters. There are over 60 possible target variables for the correct process type, many of which only occur sporadically. Therefore, the most frequently occurring types were analyzed in detail and the rest were bundled into a common category. As a precondition, the order text had to be cleaned up (typing errors, abbreviations). An analysis of data records over 4 weeks with spacy and scikit learn shows very promising results with a high hit accuracy.