Clerks spend a lot of time manually categorizing keywords and formulating responses to citizens' concerns. The aim was to enable these employees to produce high-quality responses more quickly in order to reduce the time-consuming manual effort.

Key questions were:

• Can keyword categorization be automated and integrated more seamlessly into the work process?
  
  
• Can the selection of text modules and their content be partially automated?
  
  
• Can the AI suggest suitable answers to requests to avoid the "blank page" problem?
  ‍

Background: The BKAmt also wanted a quickly realizable, presentable project that would motivate other units and make the added value of the unit visible.

The proof of concept (PoC) followed a multi-stage approach based on existing and new data, user feedback, design innovation and extensive testing.

core functions of the solution:

• Immediate threat assessment: Incoming texts are automatically checked for threats or dangerous content.
  
  
• Decomposition of complex concerns: AI recognizes key issues within complex concerns or complaints.
  
  
• Keyword suggestions: Automatic generation of relevant keywords from the context of the request.
  
  
• Automatic text module selection: After confirming the keywords, the AI selects suitable modules and information.
  
  

Customizable response style: choice between different styles (child-friendly, youthful, formal, friendly).

In phase 2, the application was optimized for local operation in the BKAmt, connected to a locally hosted version of Mixtral 8x7B (version 0.1, 2023) and installed on internal servers.

At the end of the project, a virtual test environment was set up that realistically replicates the BKAmt system. Selected language models were localized to maximize security, reliability and performance. After intensive testing, the BKAmt received the optimized application including the model and complete documentation for direct integration.

• Partial automation of the process including the creation of important letters.
  
  
• Compliance with communication rules: e.g. no formulations that invite questions.
  
  
• Content accuracy & compliance: Answers are contextually accurate and adhere to the linguistic specifications of the text modules.
  
  

The project was convincing:

• Secure, GDPR-compliant implementation
  
  
• Increased productivity thanks to user-friendly interface
  
  
• Rapid development and iteration
  
  
• Clear roadmap for further developments