Open Source Chatbot that handles sensitive information on premise

Situation

A large government agency needed a solution to service their agricultural customers outside of their busy office hours. The agency preferred an open-source solution to limit vendor dependence and to facilitate synergy with other government agencies and departments. The solution needed to be able to handle sensitive data in the future, hence no information should be sent to (foreign) vendors; the components of the chatbot that handle sensitive information should be able to run completely on premise.

Approach

We developed a chatbot that combines custom and open-source components. The chatbot is able to detect and correct spelling mistakes specific to the complex agricultural context. The parts of the chatbot that handle sensitive information can be run completely on premise should this be required. Customer care employees with no previous experience were trained to monitor user conversations and improve the language model.

Results

The chatbot stands as one of the early examples of a successful AI project within the Dutch government. Customer-care employees are actively improving and monitoring the chatbot with little to no assistance.

Case Studies