Cloud platforms offer a wide range of services, stock-keeping units (SKUs), and pricing models, which complicates the selection of optimal configurations. Although cloud providers and user communities have developed tools and best practices to support this process, existing solutions mainly operate post-deployment and lack architecture recommendations based on user constraints during the design phase. This paper proposes AzureCraftAI, a conversational agent that transforms natural-language requests into cost- and performance-aware Azure architectures and configurations with explicit explanations. AzureCraftAI generates recommendations by extracting data from Azure documentation using a hybrid Retrieval Augmented Generation pipeline, employing structured Chain-of-Thought reasoning, and adapting outputs to user perspectives via dynamic prompt engineering. In addition, it integrates official Azure retail prices to enable transparent comparisons across regions and SKUs. We evaluate AzureCraftAI using an LLM-as-a-Judge protocol with the cross-evaluation strategy. The experimental results reveal that AzureCraftAI outperforms well-known large language models in providing clear, efficient cloud architecture and configuration recommendations across different requirements and evaluation criteria.