How to check AI-drafted variance commentary: tie every figure to a cell

The errors in AI-drafted variance commentary are small and specific: revenue rounded the wrong way, a percentage worked from the wrong column, a profit figure taken from last month's version of the file. You read straight past each one, because the sentence around it is fluent. Finance reviewers call these confident wrong answers, and they are what makes an AI draft slow to review.

To review faster, borrow an auditor's habit: trace each number in the prose to its source before you judge the sentence.

Check AI-drafted variance commentary with a tie-out table: one row for every figure in the prose, the cell or formula it comes from, the cell's value, and whether the figure matches at the rounding stated in the file. Ask the model to build the table itself, then check the table rather than the prose. If more than two figures fail, stop: fix the inputs and redraft instead of patching the words.

A wrong figure costs more than a weak sentence

You can fix a weak sentence in a minute. To find a wrong figure you recompute every number in the draft, and checking the draft then takes as long as writing it.

In October 2025, CFO Dive reported that Deloitte had refunded more than A$97,000 of a contract worth about A$440,000 to Australia's Department of Employment and Workplace Relations, after its report was found to contain a fabricated quote attributed to a federal court and references to research papers that did not exist. The department said Deloitte had confirmed that some footnotes and references were incorrect; the report's review, whatever it involved, did not catch them before delivery.

With a tie-out, the reviewer stops asking whether the draft looks right and checks it against a pass mark.

The tie-out table: figure, cell, formula, tied

The table has 4 columns:

1. Figure, word for word as the draft has it: "about £160,300", "up 21% on July".
2. Cell or formula, the source in the file: B4, or (B4 - C4) / C4 for a derived figure.
3. Value, what the cell or formula returns, unrounded.
4. Tied, yes or no, at the rounding rule set in advance.

Every number in the prose gets a row, including counts, percentages, and differences. A figure with no source cell is a guess, and it goes on the list of questions.

A worked month, redacted

The example below extends the sample on The Setup page, redacted from real files: names, entities, and every figure changed. That page shows the commentary and 6 rows of its tie-out. This article adds the hotel's full August profit and loss account, with July and the August budget for comparison, and all 10 rows.

OTA is online travel agent, the booking sites that take a commission on each reservation. Totals and profit foot in all three columns.

The draft commentary:

Revenue was about £160,300, up 21% on July and 11% above budget, on 702 room-nights, up 38 on July. Profit was about £38,900, about £3,400 above budget. OTA commission was about £25,100, 19.5% of room revenue against 14.3% in July. Reason not in the figures: the split by channel is not in this file.

And its tie-out:

10 figures, 10 tied. The reviewer now checks 10 cell references and one rounding rule, instead of recomputing a page of prose.

Notice the last sentence of the draft. OTA commission rose from 14.3% to 19.5% of room revenue, and the file does not say why. A good draft says so in plain words rather than inventing a cause such as "a shift towards online bookings". The tie-out cannot catch an invented cause, so the instruction not to infer one has to be in the prompt.

Rounding rules that make "tied" mean something

You can check "about £160,300" once "about" has a definition. Set the rounding in the file before the model drafts a word, and the tie-out becomes mechanical:

• Money to the nearest £100. A figure ties if the cell is within £50 of it. 160,330.25 ties to "about £160,300"; 160,360 would not.
• Percentages to the nearest whole per cent. A figure ties if the computed value is within half a percentage point. 21.25% ties to "21%".
• Ratios to one decimal place, where the movement is small enough that a whole per cent would hide it. 14.28% ties to "14.3%".
• Counts exactly. 702 room-nights is 702.

Pick one rule per kind of figure and hold it. The same value can tie or fail depending on the rule: 10.57% is "11%" at whole per cent and "10.6%" at one decimal. If the draft says "10.6% above budget" in a file set to whole per cent, the tie-out should flag it, because the reader of the pack should see one convention throughout.

The sabotage test: change one cell, watch the check fail

Before you trust the routine, break the input on purpose and confirm the check notices.

In the sample, change one cell: food and beverage for August (B3) from 31,870.25 to 22,870.25, as if the figure had been keyed £9,000 low. Total revenue (B4) and profit (B9) recalculate. Run the same tie-out against the same draft:

5 of 10 fail. The other 5 (room-nights, the change in room-nights, OTA commission, and both commission ratios) still tie, because none of them depends on food and beverage. The failures land on the 5 figures the changed cell feeds, and nowhere else, which shows the check is tracking the source.

On The Setup page, the kit's rehearsal runs report the same behaviour at a larger scale: when three cells in the source file were changed, 9 figures failed to tie, and Claude stopped and asked instead of saving the draft.

The stop rule

Decide in advance what happens when figures fail:

• One or two fail: keep the draft, list each failure under "Questions for you" with the draft value against the cell value, and resolve them one by one. One failure tends to be a typo or a rounding call.
• More than two fail: stop. Do not save the draft and do not patch the prose. Several failures at once point to inputs that moved, or to a draft written against a different version of the file. Fix the file, then redraft.

The easy fix for a failing figure is to rewrite the number until it matches the cell. That hides the discrepancy the reviewer needed to see, so the prompt forbids it.

The prompt, with the tie-out instruction

Copy this and fill in the brackets:

You are drafting month-end commentary for [entity], [month].
Source file: [file name], sheet [sheet name]. Use only figures from this file.

Rounding: money to the nearest £100, percentages to the nearest whole per
cent, ratios to one decimal place, counts exactly.
Materiality: comment on any line that moves by more than the lower of
£[amount] or [percentage]% against budget or the prior month.

1. Draft the commentary in two short paragraphs: revenue first, then costs
   and profit.
2. Then build a tie-out table for checking, not for the reader. One row for
   every figure in your draft: the figure as written, the cell or formula it
   comes from, the unrounded value, and "tied" or "fails" at the rounding
   above.
3. If a figure fails, list it under "Questions for you" with the draft value
   against the cell value. Do not change the figure to make it tie.
4. If more than two figures fail, stop. Do not save the draft. Tell me which
   figures failed and ask me to check the source file.
5. Where the file does not explain a movement, write "Reason not in the
   figures" and say what data would explain it. Do not infer a cause.

Run it in a tool that reads the workbook itself, such as Claude for Excel, ChatGPT for Excel, or a desktop agent with the file in its folder, so the model reads cells rather than a pasted copy that may already be out of date.

The limits of a tie-out

The tie-out proves that the figures match the file. It does not prove the commentary is right. Three checks stay with the controller:

• Drivers. Is the stated cause supported by the numbers, or is it a plausible story? "Reason not in the figures" is an acceptable answer; an invented reason is not.
• Completeness. Did the draft skip a material movement? Compare the lines over the materiality threshold with the lines the draft mentions.
• The file itself. Is this the final version, after late journals? A perfect tie-out to the wrong file is still wrong.

These sit inside the wider routine in AI for month-end close, and they map onto Gate 4 of the four-gate governance model: verify before action.

The controller's part

The model drafts and checks its own arithmetic. The controller decides what the movement means, whether the explanation is good enough for the board, and whether the pack goes out. That signature is the control, and a tie-out table is the evidence behind it.

If you want this routine set up on your own files, The Setup is a Claude folder that drafts your monthly commentary from your management accounts and runs this tie-out every month. $97, paid once, with a 30-day refund.

Sources

All checked on 22 October 2026.

• The redacted sample, its commentary, and 6 of its 10 tie-out rows: The Setup page, https://aielitethinkers.com/setup (the full profit and loss account and the remaining rows are worked here from the same redacted file)
• CFO Dive, Deloitte refunds over $60K for report with AI errors, Australian government says (Alexei Alexis, 21 October 2025): https://www.cfodive.com/news/deloitte-refunds-60k-report-ai-errors-australian-government-accounting/803321/
• Anthropic, Claude for Microsoft 365: https://claude.com/claude-for-microsoft-365
• OpenAI, Introducing ChatGPT for Excel: https://openai.com/index/chatgpt-for-excel/

Start with The AI Playbook. Free, 45 minutes, written for senior professionals in finance, property, and hospitality.