Find out more about the latest developments in credit risk management for your industry.

Logo Created with Sketch. SCHUMANN - EN

How Do Chatbots Like Chat GPTChange Credit Risk Management?

ChatGPT, the AI language model developed by OpenAI, has been a topic of conversation for weeks. Based on patterns and information, ChatGPT can generate human-like answers and texts. Which changes will this bring to credit management in the future? Find out more in SCHUMANN Insights.
Schumann Insights, Blog Post
30.05.2023, Prof. Dr. Matthias Schumann

Background to ChatGPT

ChatGPT is the talk of the town, as the tool can be used to conduct human-like dialogues or generate texts depending on certain questions. It is actually a chatbot that is very powerful in communication and contains mechanisms to learn from conducted dialogues. A special language model is used for this purpose. But the question is whether such a tool will lead to changes in credit management.

Meanwhile, the AI language model has gone through various learning stages. In addition to supervised and unsupervised learning phases with very large data sets, reinforcement learning has finally been used, in which human testers provide feedback on the answers from ChatGPT.

Where can such approaches now be used in credit management? It should be noted at the outset that in order to make the AI language model powerful, it must be fed with sufficient training data, which has not yet been fed to the system in this form. Assuming this is the case, there is a broad field of application.

Possible Areas of Application for ChatGPT in Credit Risk Management

Such a system can be applied to answer corresponding queries, e.g., the payment behaviour of customers could be verbalised. Likewise, such a solution can be used to explain to customers whose orders are only processed against cash payment why this is the case. In complex application situations, where financing components, their term and prices depend on the prospective customer or customer risks, it would be possible for the system to explain the respective results of a corresponding simulation or to make alternative proposals. This would be a field of application for leasing or credit insurance companies, for example.

If there is sufficient training data, an annual financial statement analysis could also be explained verbally, or the examiner could ask the system for details and their justification in a dialogue. Taking these presentation possibilities, a step further, new forms of dialogue systems for decision support would be conceivable. The decision-maker asks about unusual situations that could represent a risk in the customer data. The system presents such cases and explains which risks are present based on the identified patterns. Going even one step further in the training, suggestions for solutions, e.g., for limiting or reducing the risks, could also be integrated. Likewise, it would be possible for the system to learn with the algorithms used with an assessment of the proposed solutions by the decision-makers. All in all, this could significantly advance decision support, especially for large customer portfolios. In the case of fully automated decisions, the affected parties could be informed directly if necessary. This would also be a gain in efficiency compared to the status quo.

Alongside direct risk assessments and decisions, the processing of complaints in relation to the handling of accounts receivable is also of considerable importance. In this area, chatbots could also control communication with customers or suppliers and thus contribute to streamlining and faster clarification of such processes.

Conclusion: Sufficient Data Stocks and Training are Prerequisites for Further Developments

Overall, however, it must be emphasised once again that sufficient data and cases must be available for training. This primarily determines the quality of such a system. However, the industry is still in the early stages of

collecting such specific data for training. In addition, despite all the speed in the development of ChatGPT, it can be assumed that not enough training data will be available in the near future. In this respect, it will still take time for such specific applications. In this sense, you as a credit decision maker are still dependent on your existing decision support systems for the time being.

About the Author
Prof. Dr. Matthias Schumann

Since 1991, Prof. Dr. Matthias Schumann has held a professorship in Business Administration and Information Systems (Chair of Application Systems and E-Business) at the University of Göttingen. He also heads the joint computing center of the Faculty of Economics and the Faculty of Social Sciences. He is a shareholder of Prof. Schumann GmbH.

Prof. Schumann's research interests include information systems at financial service providers and systems for credit management, as well as issues related to knowledge and education management. Prof. Schumann has a wide range of experience in consulting companies, extensive lecturing activities and more than 350 publications.

University of Göttingen

Matthias Schumann q