The initial situation

For renowned credit institutions like DKB, reputation is an important economic factor. It also plays an important role in the process of granting a loan. DKB wants to be sure to grant loans to trustworthy companies that do business responsibly.

DKB has recognised that there is still technology potential here to improve the process of credit decisions. After all, a DKB bank advisor often has to decide in a short time whether a corporate customer will receive a loan. This is the only way to ensure a good customer experience.

But how can publicly accessible information about a company be included in the credit decision in the shortest possible time?

The solution approach

The bank advisor needs a software application to assist him. It works like this: The bank employee enters the company name into the software application. The application searches through millions of pieces of information about the company (e.g. comments, news articles or posts) from publicly accessible sources such as Twitter, Facebook or t-online. The application then performs a sentiment analysis using AI models (e.g. BERT, related to GPT) and gives the bank advisor an indication:

• “70% Negative”
• “20% Positive”
• “10% ambivalent”

At this point, it is important to mention that the results of the sentiment analysis are only used as guidelines. Nuances such as irony and exaggeration cannot be captured. Thus, this step is mainly about recognizing publicly known patterns of a company. Example: The sentiment analysis gives a predominantly negative indication. The bank advisor then uses the sources provided by the software for more detailed research and learns more about the background to the publicly shared news. In this way, DKB - in conjunction with all other available information - can minimize its lending risk.

This result is only part of the information that he can include in the credit decision. The software application accelerates this step in the risk assessment process. This is because the bank advisor does not have to search through all the sources independently. They can incorporate the generated results into their analysis and build on them. The better a bank advisor knows his customers, the better he can advise them.

Technological innovation

Doesn't the implementation sound simple? Yes, we think so too. But it wasn't always like that. To be more precise, the implementation effort has only been very low since December 2022. Why is that? ChatGPT was published in December 2022. And it, or BERT (the sister of GPT), is also used here. BERT stands for "Bidirectional Encoder Representations from Transformers" - a machine learning and pre-training technique for transformer-based natural language processing models (NLP models) developed by Google and introduced in 2018.

But let's take another step back: why was it so difficult to technically implement this use case pre-GPT in the first place?

What we need here: A model that indicates reliably enough whether a text is written positively, negatively or neutrally. "Conventionally", the case would have been processed using a natural language processing approach, for example, i.e. the "old school" approach with Python and the NLTK program library.

Alternatively, you would have had to laboriously train your own model. These are exactly the approaches that require a lot of data, time and AI experts. The results of our brief validation were also anything but promising. We would probably have quickly discarded the case or fallen into a costly trap with high (self-)training costs.

But luckily there is BERT. This gave us very promising results out of the box in just a few minutes. It was clear that the case would work and the effort would be low. What remains is: plugging together, implementing, prompt engineering and delivering - all things that SIDESTREAM is very good at.