Katarina Šupe's article explores the use of large language models (LLMs) in graph databases, specifically focusing on their application in Memgraph, an in-memory graph database that uses openCypher. The piece compares two approaches for enabling natural language interfaces: direct Cypher generation by LLMs and tool invocation, where predefined tools are used by LLMs to interact with the database. While direct Cypher generation offers flexibility, it poses challenges such as fragility, bias, and maintenance issues, especially as the application scales. In contrast, the tool invocation method provides a structured and reliable interface, enhancing control, safety, and performance in dynamic environments like Memgraph. This approach allows developers to optimize queries and maintain control over performance without retraining models, leading to more stable AI systems. The article introduces the AI Toolkit for Memgraph, which offers ready-to-use tools to facilitate building reliable AI agents, and previews upcoming features in Memgraph Lab that will support tool invocation for more reliable natural language querying.