harshaA natural-language interface makes a large dataset easier to approach. It also makes it easier to hide the steps between a question and a number.
My work on a departmental Data Science Agent includes large Excel and PDF inputs, retrieval, and analysis workflows. The product challenge is to let a user ask a broad question while preserving enough detail to verify the result.
Make the transformation visible
For an illustrative question such as “which department changed most this quarter?”, the answer depends on definitions. Which records count? What defines the quarter? Is the comparison absolute or relative? How are missing values handled?
A useful response should reveal those choices. It can include the selected dataset version, relevant columns, filters, aggregation, and a compact result table. The user should not have to infer the calculation from a fluent paragraph.
If a question is ambiguous, clarification can be more valuable than immediate computation.
Distinguish search from calculation
Retrieval can find a relevant sheet, definition, or explanatory passage. Arithmetic still needs a controlled computational path. A language model's summary should not become the only record of how a numerical result was obtained.
I would retain an inspectable transformation or query where appropriate, with boundaries on execution and access. The answer can then describe a result computed by a defined process rather than presenting an unsupported estimate.
This distinction is particularly useful when evidence spans both narrative documents and tables.
Keep the evidence small enough to inspect
Showing every row is not the same as making the work understandable. A concise table, row count, and explanation of exclusions may be more useful than a giant dump.
The interface should let the user drill down when necessary while keeping the main answer readable. Sensitive records still need the same permission boundaries as any other data-access workflow.
The goal is not to make every user inspect every operation. It is to make inspection possible when the result matters. A data agent becomes more useful when it can answer both “what is the result?” and “how did you get there?”