Intelligent Agents/chat-bots have become a hot topic in industry. Amazon, Apple, Google, Facebook and Microsoft have all invested heavily in the area. Many start-ups work on different perspective of the space as well, ranging from language understanding techniques to solutions for specific tasks (e.g., appointment scheduling). However, it is still very costly to introduce a new experience to an agent/bot. A major issue here is that language understanding and conversation management modeling are often performed in a domain-specific fashion – either with data-driven statistical modeling or with semantic grammar authoring – the former requires a large amount of labeled training data; the latter needs the combined expertise in linguistics and domain knowledge. In this talk, we formulate the domain scaling as a training data demand-supply problem, and introduce some preliminary investigations and experiment results on this problem.