Conversational Local Analytics

Large language models are often presented as tools for "chatting with data": users upload a dataset, ask questions in natural language, and receive summaries, charts, or insights. This model is useful for small or moderately sized files, but it becomes inadequate when datasets are large, sensitive, messy, or methodologically complex. A one-million-record CRM database, a large survey archive, or a scientific dataset cannot always be conveniently or responsibly uploaded into a conversational interface. This position paper proposes a different paradigm: Conversational Local Analytics. In this approach, the AI system does not need to contain the data. Instead, it supports the user in designing, generating, debugging, and iterating local analytical scripts that run on the user's own machine or infrastructure. The raw data remain local. The conversation contains the problem framing, the methodological discussion, the code, the selected outputs, and the next analytical decisions. The key shift is not merely technical. Conversational Local Analytics replaces the idea of a monolithic analytics platform with an iterative dialogue between human judgment and script-mediated computation. It can make advanced data work more flexible, scalable, educational, and accessible, while preserving reproducibility and a meaningful degree of user control.