Why LLMs struggle with analytics

AI agents with large language models (LLMs) have the potential to transform data analytics by providing intelligent responses to business queries, but they face significant challenges in handling analytical data. LLMs, trained to predict sequences in narrative texts, struggle with interpreting the multidimensional relationships in tabular data, leading to difficulties in generating accurate SQL queries and managing complex datasets. These models often underperform with long-running queries, exhausting context windows, and failing to grasp the specific data architectures of large organizations. Tinybird's implementation of an MCP server and the development of the Explorations UI demonstrate a solution by emphasizing the importance of context—both static and dynamic—to help LLMs map user intent to data more effectively. By documenting resources and employing semantic models, organizations can improve LLMs' understanding of data structures, allowing for more meaningful insights rather than just syntactically correct queries. The future of agentic analytics will belong to companies that can bridge the gap between human language and the intricate data environments, thus democratizing analytical thinking and making it quickly accessible to all users.