Flux analysis with AI: how to check the explanations it writes

What a flux analysis is

A flux analysis compares each account's balance with an earlier period, usually last month or the same month last year, and explains every movement over a threshold you set, such as $1,000 or 10%. It's a month-end review step: a movement nobody can explain is a reason to look closer before the books close.

Flux analysis vs variance analysis. Variance analysis usually compares actual results with a budget or forecast, while a flux analysis compares actuals with actuals. People use the terms loosely, and the checks below work for both.

Where AI helps, and where it goes wrong

AI saves real time here. It can pull the accounts over the threshold out of a trial balance in seconds, group the GL lines behind each movement, and give you a first draft of the commentary instead of a blank page.

It also fails in ways that are easy to miss, because a wrong explanation reads just as well as a right one:

• It explains what it can't see. Given only a trial balance, it will still offer reasons. They sound like the kind of thing that moves that account, and they're guesses.
• It tells a plausible story about a mistake. A misposting moves two accounts at once, and the AI may explain both movements as real business changes. Accountants have described AI writing a convincing reason for what turned out to be a misposting.
• It skips an account. In a long list, an account just over the threshold can go missing, and nothing in the output tells you.
• It gets a figure or a direction wrong. An increase described as a decrease, or a change worked out from the wrong column.

Before you start: give it the evidence and the rules

Most weak flux commentary starts with a weak brief. Give the AI what a person would need to do the job:

• the trial balance for both periods;
• the GL detail for the current period, at least for the accounts over the threshold;
• the threshold;
• last month's commentary, if you want the wording to stay consistent.

The rules matter as much as the files. Ask it to cite the GL lines behind every reason, and to say plainly when the detail doesn't explain a movement rather than suggest a likely cause. Practice files are fine for trying this out; for real client files, see what can safely go into an AI tool.

Attached are the trial balances for September and October, and the October GL detail.

For every account that moved by $1,000 or more between September and October:
1. Give the September balance, the October balance, and the change with its direction.
2. Explain the change using only the GL detail. Cite the date, description and amount of each line you rely on.
3. If the GL detail doesn't explain the change, write "Not explained by the GL detail" and don't suggest a reason.

Before the commentary, show a table of every account and its change, largest first.

Five checks before the commentary goes anywhere

1. Coverage: is every account over the threshold there?

Work out the list yourself rather than asking the AI whether it missed any. With last month's balances in column B and this month's in column C, this formula marks each account that needs a comment:

Fill it down, count the TRUEs with COUNTIF, and compare the count with the number of explanations. Asking for the full list sorted largest first, as in the prompt above, also shows you the accounts just under the line.

2. Figures: do the numbers match the trial balance?

Tick every balance and change back to the trial balance. Do the arithmetic in the spreadsheet, not in the chat, and check the direction of each change as well as its size.

3. Support: does each cited line say what the explanation says?

Open the GL line it cites and read the amount, the date and the description. The most common miss is a real line that doesn't support the claim, such as a refund cited as new revenue, or a line from another month.

4. The misposting test: do two movements mirror each other?

If one account rose by about the amount another fell, check whether an entry went to the wrong account before you accept two separate business reasons. A movement with no GL lines behind it can also mean a journal entry outside the detail you gave the AI.

5. Wording: is it as certain as the evidence?

"Due to higher sales volume" needs evidence of volume. Cut any cause the files don't show, and replace it with what is supported ("three invoices to new customers, totaling $8,400") or with "not yet explained, asked AP".

Before that goes into the file, you'd want a GL line or a lease that shows it. If the only new October line is a $2,500 payment described as "Rent - Sept late", the lease is invented, and the movement is September's rent paid in October.

Correct it with specifics

When an explanation fails a check, tell the AI exactly what's wrong and what to use instead. "Are you sure?" or "check again" usually gets you another guess, sometimes the same one.

The Rent explanation isn't supported. The only new October rent line is "Rent - Sept late", $2,500, dated October 3. Rewrite the explanation as a timing difference: September's rent paid in October. Don't mention a lease.

Then check the rewritten line the same way as the rest. A correction can introduce a new error.

Save the prompt for next month

Once a prompt works, save it with the rules built in, so next month starts from a tested brief instead of a blank page. Put placeholders where the details change, so it works with only those filled in.

Attached: trial balances for [last month] and [this month], and the GL detail for [this month].

For every account that moved by [threshold] or more:
- Show both balances and the change, with its direction.
- Explain the change using only the GL detail, citing the date, description and amount of each line.
- If the detail doesn't explain the change, write "Not explained by the GL detail." Don't guess.

Before the commentary, list every account with its change, largest first.

Keep a note of what the AI got wrong each month. If the same kind of mistake keeps coming up, add a rule for it to the prompt.

Who signs off

The AI drafts and a person owns the result. Record who checked the commentary and what they changed, as you would for any other workpaper. If a reviewer asks how you know an explanation is right, the answer should be the GL line, not the AI.