Industrial chatbots demand stable support infrastructure, simplicity of integration into production and compliance with corporal data protection policies. This frequently leads to the fact that the focus is shifted away from the actual performance of the NLP components. While recent advances of academic research on dialogue systems such as sequence-to-sequence systems, multitask neural NLU classifiers etc. are very inspiring, leading industrial solutions still offer blackbox NLUs which seem to stagnate in their overall performance. It is hard to assess their innovation degree or improve their performance in any other than data-based manner. This is definitely one of the factors why performance of the industrial chatbots can be subbar and may lead to negative customer feedback. As the AI integration became an indispensable component of success for many ventures, many non-IT companies create their own departments of AI and NLP specialists that are willing to contribute to the improvement of NLU algorithms. They are not satisfied with all-inclusive blackbox solutions and redirect their attention towards more transparent or even open-source frameworks that run on premise. While transparency of an NLU solution and its high performance is definitely important, the decisive factor is still whether or not the solution can be easily integrated and can run flawlessly in production. While choosing a chatbot solution, it has proven to be very difficult to find an optimal trade-off between transparency and stability. In this talk, we present our experience with an open-source NLU solution and discuss its advantages and limitations.