Skip to main content

Case studies

Credit assessment with ai support

An AI based virtual assistant can make the processing of loan applications faster, more efficient and reduce decision time.

With a 60-year history, Fókusz Takarékszövetkezet provides a full range of financial services to its customers through a network of nearly 20 branches.


In 2016, Fókusz Takarékszövetkezet implemented a loan approval support system that handled 800-1000 loan approvals per year. This included a forum where management members could ask questions about the loan and then vote on the loan application based on the answers and the information stored in the system. These human factors were the reason for the slow decision-making process, which reduces customer satisfaction.


An AI based virtual assistant can make the processing of loan applications faster, more efficient and shorten decision time by analysing the application data right after submission based on objective factors. The assistant is able to ask questions about the missing information or on a factor considered risky.


To speed up and support the approval process, the forum has been extended with the virtual assistant (Avatar), which is objective, fast and reduces the possibility of human error. The model behind the Avatar is not only based on structured data, but also on previous forum posts.
Based on this model, Avatar asks forum members about factors that are considered risky from a credit perspective. It then understands the response received in the forum and uses it into its decision. By asking these questions and making a decision, Avatar can support other decision-makers and draw attention to the risks of the credit.

1. Preparation

  • Understanding the credit assessment process and decision criteria
  • Building a domain-specific dictionary
  • Creating a text analytics corpus

2. Implementation

  • Integration of Avatar as a virtual forum member into the existing credit scoring system
  • Integrating Avatar with the risk analysis solution
  • Development of a "self-learning" algorithm