Data resource building is one of the main parts in training deep learning-based chatbot. To address this, it is a time-consuming and expensive task, especially in specific domains such as customer service. In this paper, we propose a method to build dataset for customer service chatbot in the mobile phone domain. We applied a "light annotation" and extended it adapting to online service by continuous operations. And we call the approach as PICO (Planning - Implementation - Continuous Operations). Actual use indicates our approach can reduce costs and secure the quality. The rate and trend of offline evaluation results with the datasets are very close to online sampling analysis results, which further indicates the built datasets are well representative and reliable. We believe the approach can be extended to other consumer service domains.