LLMs are built mainly for natural language, not for directly understanding spreadsheets. When you pass a table as text (CSV, Excel copy-paste, etc.), the model just sees a stream of tokens — it doesn’t “know” rows and columns. This often breaks the structure and leads to wrong (or hallucinated) answers.

Better approaches:

• Clean and normalize tables before sending (consistent headers, no merged cells).

• Convert to JSON or key-value pairs so relationships are explicit.

• Split large tables into smaller, labeled chunks.

• For complex data, query it with SQL/Pandas first, then give the summarized result to the LLM.

General Principle:
Think of it this way:

• LLMs = language + reasoning layer

• Databases/spreadsheets = structured data layer

Best practice is to let each system do what it’s best at, and connect them.