In this paper we propose a simple but powerful method of extracting key client requests from bank chat logs. Many companies nowadays are interested in building a chat bot to optimize their business, and are ready to provide chat bot developers with large amounts of data, but such data often need special preparation to be successfully used for a chat bot system. We propose a method of data preparation which includes not only data cleaning but also data mining: we extract key notions from chat logs and retrieve typical client requests in generalized form. The method uses simple metrics as well as word embeddings and additional semantic resources to extract typical client requests from client-manager chat logs.