Pulumi's development of the AI assistant, Pulumi Copilot, for cloud infrastructure highlights the challenges and insights gained from integrating large language models (LLMs) into software engineering. A significant challenge arose when Copilot suggested a non-existent command, revealing the need to balance prompt engineering with traditional coding. This experience underscored the importance of using LLMs for natural language tasks while relying on traditional code for structured data transformations to reduce costs and improve efficiency. Pulumi refined Copilot by breaking it into modular "skills" that handle specific tasks, enhancing the AI's ability to interact with users and manage infrastructure queries. Despite the potential for AI-generated outputs to appear accurate while containing errors, Pulumi developed rigorous testing methods, including using LLMs as evaluators, to ensure response accuracy. Interestingly, AI hallucinations sometimes revealed user expectations, offering unexpected insights into product development. These lessons led to the launch of the Pulumi Copilot REST API, allowing users to integrate Copilot's capabilities into their own tools, with a focus on continually learning from user interactions to refine AI systems.