For precise questions over tabular numeric data, convert the workbook data into a structured relational/table representation and expose controlled query functionality to the AI application through function calling. This preserves columns such as quarter, region, revenue, product, and fiscal period as typed fields that can be filtered, aggregated, and compared deterministically. Microsoft demonstrates natural-language-to-SQL patterns in which a user ' s question is translated into a query against structured PostgreSQL data, while Foundry function calling allows a model or agent to invoke application-defined query tools with structured parameters.
This is materially more reliable than embedding spreadsheet rows as unstructured text. Vector search is optimized for semantic similarity , not exact arithmetic, grouping, filtering, or aggregation. A query such as " Q3 revenue for the West region " is fundamentally a structured lookup or aggregation problem. SQL can express those constraints explicitly and return the authoritative numeric result.
Microsoft Foundry also supports agents working with structured data and provides examples where agent tools analyze numeric files and return computed values.
Relying on parametric model knowledge cannot provide current organization-specific financial figures. Azure AI Vision could OCR an image of a spreadsheet, but converting an already structured workbook into an image would destroy useful schema and increase extraction error risk.
Study Guide references: Microsoft Foundry agents; function calling; structured data access; natural- language-to-SQL; grounding AI responses in enterprise data.