Every Number Has a Source: Making AI Cite the Page Behind Each Figure

An AI assistant hands back a wage total from a client folder. It looks clean, in a tidy sentence with no hedging. It reads the same whether it's right or wrong.

That's the trouble with any finished number, and AI output is full of them. A sterner prompt about accuracy won't fix it. Ask for the page, then open it and check the cited text. Require each extracted figure to include the document, page, line or box, and exact copied text. Every number should carry a source, or a flag that says the source is missing.

This piece covers the AI-specific part. What to ask the tool for. The four ways its citations go wrong. How a Reviewer checks them before anything lands in a workpaper.

Why an AI figure needs a locator more than a typed one

When a preparer keys a figure, there's a person behind it. You can ask where it came from. They open the statement and point. The base habit for that, a stable document identifier plus a specific locator, is covered in the source-citation habit. This piece assumes you already have it.

An AI tool changes two things. First, its tone doesn't move with its certainty. A figure it found printed on one page and a figure it pieced together from two pages arrive in the same calm voice. Second, there's nobody to ask later. Ask the tool again next week and you may get a fresh answer, not a record of how it reached the old one. Record the source trail when the figure is produced.

So the locator does more work here. With a typed figure, the citation backs up a person. With an AI figure, the citation is the only witness.

Folder access raises the stakes. Anthropic's Cowork help page says: "On desktop, Claude can read from and write to your local files without manual uploads or downloads." ChatGPT's quickstart page says: "ChatGPT can read and modify files in the folder you choose." A tool that can open a whole client folder can also pull from the wrong file in it.

What to ask for

One citation row per figure. Not a paragraph of explanation, and not a list of sources at the bottom. A row, sitting next to the figure, with five fields:

• Document name, as it appears in the folder.
• Page, counted in the file, not the printed page label if the two differ.
• Line, box, or row label on that page.
• The exact words or digits copied from the page, character for character.
• Status: "found", "combined" (the figure uses arithmetic, shown on its own line), or "could not find it".

The last field carries the most weight. If the prompt doesn't name an acceptable way to say "not here," the request reads as an order to produce a number. Give the tool permission to come back empty, and say plainly that a blank is a correct answer.

A pattern you can adapt:

For each figure you report for Jane Doe and Doe Holdings for tax year [tax year], give one row. Use these fields: document name, page number in the file, line or box label, the exact text or digits as they appear on that page, and status. Status is one of: found, combined, could not find it. Copy the text exactly. Do not round, reformat, or summarize it. Use one source per row. If a figure needs numbers from two places, give each source its own row and show the arithmetic on a separate line. If you cannot find a figure, set status to "could not find it" and leave the value blank. Do not estimate. Do not use any file for a different client or a different tax year than the one named here.

Two details in that pattern earn their place. Specify the page number in the file, since the printed folio on a scanned page often won't match what your PDF viewer shows, then confirm that the tool used that numbering. The scope line names the client and period; check each cited document against both.

How AI citations go wrong

Four failure modes are worth naming, because each one survives a quick glance.

An invented page number. The figure is real and the document is right, but the page is not. The citation looks complete, so nobody opens it. The tell shows up the moment someone does: the digits aren't on that page.

A paraphrase shown as a quote. The exact-text field holds something close. A label gets reworded. A total loses its commas or its cents. Close isn't exact. If the copied text doesn't match the page character for character, you can't tell whether the tool read the page or rebuilt the number from context.

Two figures merged into one. The tool adds a corrected amount to the original, or sums two boxes, and cites only one of them. The number may even be right. The citation still points to a page that doesn't show it. Arithmetic belongs on its own line, with every input cited separately.

The right number from the wrong document. Client folders collect near-duplicates. A prior-year copy of the same statement. A draft beside the final. A file the client sent twice under two names. The tool cites a real page with the real digits, in a file that was never in scope. This one is the hardest to catch, because the quote check passes.

The check before the figure enters the workpaper

Citation rows only help if someone opens the pages. The check is short, and it runs the same way every time.

1. Open the cited page in the cited document. Not a similar file. The one named in the row.
2. Compare the copied text to the page, character by character. Digits, commas, decimal places, and the label.
3. Confirm the document is in scope. Right client, right period, final version. This is what catches the wrong-document failure.
4. Check every figure that carries judgment, and a sample of the rest. Judgment figures are anything the tool combined, allocated, or interpreted, plus anything that drives a position. For figures copied straight off a form, check a sample, and set its size before you start, not after the first rows look clean. If one in the sample fails, widen the check.
5. Send every "could not find it" to the Preparer. Don't ask the tool to try harder and fill the gap. A blank is information. A filled blank with no source is a guess dressed as a figure.
6. Record who checked, and when. Initials and a date beside each row, or beside the batch.

The Reviewer trusts the page, the same standard they'd hold a junior preparer to. More on the Reviewer's side of the file is at for reviewers.

Where the trail lives

Chat history is not a workpaper. Keep the citation rows with the workpaper so the next Reviewer has the source trail.

Copy the citation rows out of the chat and into the workpaper, beside the figures they support. Save the prompt you used alongside them, so the next person can see what the tool was told to do. Keep the checker's initials on the rows. The AI-specific test: someone who never saw the chat can read the prompt, read the rows, and open every cited page.

Where TracePrep fits

In TracePrep's Reasonable Compensation Study, every figure traces to its source, occupation, geography, and period, at field level. A Reviewer signs off on each Study, and the audit log records who prepared it, who changed what, and who approved it. Finalized evidence stays with the Firm.

This is general information, not tax or legal advice. Your licensed professional and your Firm's Reviewer are responsible for every return and every conclusion. Keep client data inside the systems your Firm has approved. As general guidance only, not advice: if you use an AI tool with client data, firms generally look for a business or commercial plan whose terms cover how that data is handled; read the vendor's current terms before you do.

Join the waitlist