Data provenance is the record of where a number came from and what happened to it on the way to your report. If you cannot trace a figure back to its origin, you cannot defend it when someone asks.

Most small teams do not lose trust because of one big error. They lose it because nobody can answer a simple question: "Where did this number come from?"

Why provenance matters more with AI in the mix

A spreadsheet that passes through three people and two exports is already hard to trace. Add an AI tool that summarises, cleans or fills gaps, and the trail gets thinner. The output looks tidy. The origin is invisible.

Provenance gives you a way to check the work. It lets a reviewer see which parts came from a source and which parts were produced or changed by a tool.

What to record for every figure

You do not need special software. A simple log beside the report is enough. For each key number, note four things.

The original source

Write down who published or owned the data. Record the file name, the system it was exported from, or the page it was read from. A link or path is better than a description.

The date and version

Record when the data was pulled and which version you used. A sales export from the first of the month is not the same as one from the end of it. If the source updates, your number may drift without warning.

Every change made

List each step that altered the data. Removing duplicates counts. So does converting currencies, filtering rows, or letting a tool fill blanks. Name the person or tool that did it.

Who checked it

Add the name of the person who reviewed the result and the date. This turns a loose habit into a repeatable control. Our page on data verification explains how that review step works in practice.

A worked example

Say a retail team reports that returns rose 12 percent last quarter. The analyst built the figure from a shop export and a courier file.

A weak record says "returns data, Q3". A useful record says this:

  • Source one: the shop order export, pulled on a named date, file saved in the shared folder.
  • Source two: the courier returns file, supplied by the courier's account manager.
  • Changes: 41 duplicate order lines removed, refunds matched to orders by order number.
  • Checked by: a second analyst, who recounted a sample of 20 orders by hand.

Now a manager can challenge any step. If the courier file turns out to be incomplete, the team knows exactly which number to revisit.

Treat AI output as a claim, not a source

When an AI tool gives you a figure, a summary or a cleaned table, treat it as a claim that needs support. It is not a source.

Ask three questions before the output goes into a report:

  1. Can I find the original data behind this?
  2. Did the tool change or invent anything, such as filling empty cells?
  3. Has a person compared a sample of the output against the original?

If the answer to any of these is no, mark the figure as unverified and keep it out of decisions until it is checked. Good data validation catches format and range problems. Provenance tells you whether the number was real to begin with.

Build provenance into your routine

Start small. Pick the five figures your team quotes most often and write the provenance record for each. Keep it in the same folder as the report.

Then add one rule: no number goes into a report without a source line. This costs a few minutes per report and saves hours when someone doubts a result.

Review the log on a schedule. Sources move, files get renamed, and owners change jobs. A record that was accurate in spring may point to nothing by autumn.

Signs your provenance is weak

Watch for these warning signs:

  • Nobody can name the file a chart was built from.
  • The same metric has different values in different decks.
  • A figure was "cleaned" but no one remembers how.
  • An AI summary is pasted in with no link to the underlying data.

Any one of these is a reason to pause and trace the number. For a view of how we assess evidence quality, see the trust index.

Traceable numbers are easier to defend, easier to correct and easier to reuse. That is what a trustworthy report looks like.