Different AI models and their areas of application
When AI is discussed today, everyone immediately thinks of large language models such as Chat GPT, partly due to their personal experiences. However, when it comes to AI models, it is important to differentiate between which method is suitable for which task.
AI for scoring and rating
There is often discussion about the fact that AI solutions can be used for scoring or rating. This involves making predictions about the future economic situation. For example, a rating for a company or a person is to be predicted. To do this, forecasts must be derived from time series of data. Recurrent neural networks are used for this purpose. Complex combined decision tree methods such as XGBoost are also an alternative. These methods achieve good results for such tasks and allow the relevance of the influencing factors to be easily interpreted. However, research results also show that training on the data sets used to date yields hardly any improvements in discrimination and thus performance compared to established statistical methods. Advantages only arise when additional data, such as competitive intensity, employee satisfaction, etc., are included in the analysis.
Regulatory requirements under the AI Act
Due to the AI Act applicable to these tasks, a distinction must also be made between the rating of legal entities and natural persons. Without the use of personal data, the application of these methods is unproblematic. However, if personal data is integrated and rated, the application is quickly classified as high risk and extended documentation, reporting and auditing requirements must be taken into account.
Requirements for other AI forecasting systems
Of course, if time series or a large amount of experience are available, forecasting systems can also be developed for other tasks. The prerequisite is always the availability of a sufficiently large training data set.
Generative AI and LLMs in credit management
Generative AI in the form of large language models (LLMs) has different characteristics and, due to its ‘language capabilities’, is suitable for other tasks. Today, credit officers often compile so-called credit templates for major credit decisions from a variety of individual pieces of information and assessments, and also describe them verbally. This is an example of how LLMs can be used to create an initial version of such a template. The credit officer then has to review, revise and supplement it. With the appropriate performance of such a system, which requires only minor changes, efficiency gains can be achieved in the creation of credit templates.
LLMs for document analysis, extraction and summarisation
LLMs can also be used to instruct them with suitable inputs, known as prompts, e.g. to extract information from PDFs and then process it automatically in a structured manner, e.g. for an annual financial statement analysis. LLMs can also create summaries to summarise the essential aspects of a text template. This could be helpful when interpreting appendices or analyst reports.
Natural language information systems with RAG
If you want to create information systems in credit management that can communicate in natural language, you need LLMs that form the interface and communicate with a knowledge base in which the answers to the questions asked are to be found. To filter the right answers here, additional pre-processing steps are necessary. This is done using what is known as ‘retrieval-augmented generation’, which supplements the LLM. In addition, further ‘tuning measures’ could be used to improve the quality of the solutions.
Conclusion: AI is not just AI
These selected examples show that AI is not just AI, but that it is necessary to differentiate very precisely which tasks are to be supported. The selected models must then be trained for this task and their parameters adjusted in order to achieve the best possible results. Training and validation data sets must be used for this purpose. This is a process whose complexity should not be underestimated and must be weighed against the achievable increases in efficiency or improvements in quality. If this and the data set required for setting up the AI are taken into account, AI can make a valuable contribution to credit management.