Chatbots are text-based conversational agents. Natural Language Understanding (NLU) models are used to extract meaning and intention from user messages sent to chatbots. The user experience of chatbots largely depends on the performance of the NLU model, which itself largely depends on the initial dataset the model is trained with. The training data should cover the diversity of real user requests the chatbot will receive. Obtaining such data is a challenging task even for big corporations. We introduce a generic approach to generate training data with the help of crowd workers, we discuss the approach workflow and the design of crowdsourcing tasks assuring high quality. We evaluate the approach by running an experiment collecting data for 9 different intents. We use the collected training data to train a natural language understanding model. We analyse the performance of the model under different training set sizes for each intent. We provide recommendations on selecting an optimal confidence threshold for predicting intents, based on the cost model of incorrect and unknown predictions.