The document discusses the challenges of building chatbots using natural language. It notes that natural language has insane variability and long tail distributions, meaning scripts do not work. It also discusses the high dimensionality of natural language and techniques for reducing this, such as preprocessing, understanding, and detecting intents. However, it notes there are long tails of intents across business verticals. An ideal chatbot is not a silver bullet and what businesses expect from bots does not always match reality. The best approaches involve statistical analysis of conversations, understanding intent distributions, semi-supervised learning from conversations, and reinforcement learning with an agent-in-the-loop.