In addition to conventional decision logics for determining the next step, there are many other exciting applications in receivables management, including the automation of complex written communication with end customers and third parties. This requires the accurate recognition of incoming queries and the correct deduction of next measures. Performing this assessment using artificial intelligence enables the direct initiation of steps, respectively context-dependent forwarding to the right administrators. By combining machine learning with human experience, this allows queries to be answered effectively and as quickly as possible.
Decision-making scope of artificial intelligence
After deciding to entrust manual processes in debt collection to artificial intelligence, two important topics need to be considered: firstly, the quantity, nature and quality of the database that can be used for modelling. Secondly, there is the question of the target process: is the automated decision a single point in the process that should recognize and assess every special case to initiate the next action? Or will the decision-making authority of the machine be restricted to the most frequently occurring possibilities?
The answer to this question regarding the optimum degree of automation is largely determined by the extent and reliability of the data available for training the model. If there are only limited changes to the processes in the course of digitization and the historical data sufficiently covers future decisions, directly modelling all potential cases is quite possible. This hugely simplifies technical implementation with fewer modules and interfaces.
However, if future decisions are defined differently or the data quality is poor or inconsistent, the AI should only initially take basic decisions. Complicated matters should still be handled manually.
Striking a balance with successive model use
In the example of mail classification using machine learning with hundreds of queries to be identified regarding receivables management, at Arvato Financial Solutions we opted for a balance based on our many years of experience: to keep the degree of automation as high as possible despite a number of altered processes and many rare cases.
The implementation of the recognition logic after scanning encompasses three consecutive steps:
1st model – readability:
If a written document produces a poor digital scan due to weak toner or handwritten notes, the automation process first allows for specific points to transfer the query to manual processing. An application-based readability index was developed using statistical methods for this purpose. Only a sufficiently readable document will move onto the next step or model.
2nd model – recognition of frequent queries (e.g. claim questions, instalment agreements, complaints):
The following basic model is mainly based on a three-level neural network. It first only predicts the incidence of frequently occurring debtor queries. The several hundred rare queries are aggregated into pools. That’s because they pose a statistical challenge particularly in a text mining context: many of these letters contain lengthy texts with few keywords, and the training quantity from the past is often extremely limited.
3rd Model – recognition of rare queries (e.g. reduction arrangements):
If a rare query is predicted in step 2, corresponding special models will then be activated to precisely identify the contents.
Even though this cascading implementation with branches and the integration of several models appears complicated at first glance, there are clear benefits: separate management allows for the discrete measurement of the recognition quality of the individual models. For the team comprising the process manager, technical expert and data scientist, this simplifies the readjustment of threshold values and the independent improvement of individual areas.