What provenance means in an AI data analysis

The four parts of provenance

A useful source line for a figure answers four questions:

• Where: which file and which sheet. Workbooks often hold last year's data on a second sheet with the same columns.
• What: which columns were used. "Margin" computed from list price and from net price are different numbers.
• Which rows: the filters applied, in plain words — "Region = West, 2026 only".
• How many: the rows the figure was computed over, against the rows in the file.

Why the row count matters most

Most wrong answers change the number of rows. Totals rows add rows; a missed filter adds rows; a join on the wrong key multiplies rows; blank cells remove rows from an average. A figure that says "computed over 1,180 of 1,240 rows" invites the right question — which 60 were left out? — before anyone relies on it.

Declared versus observed

An AI tool can describe its own method, but a description is a claim. Stronger provenance compares what the analysis says it used with what the code was actually seen reading. If the two disagree, the declared version should not be trusted.

In DataMimi, the columns in the source line are the union of the columns the analysis declared and the columns the code was observed reading, and a column name that does not exist in the file is dropped rather than shown. A row count larger than the data is discarded instead of displayed.

What to ask of any tool

Before trusting a tool with numbers that leave your team, ask whether it can show, for every figure:

• The sheet it read, not just the file
• The exact columns, checked against the file
• The filters, in words a non-analyst can verify
• The rows used out of the rows available

Common questions