Automating Variance Analysis & Flux Commentary with AI

By Christian Matumona, Kyootek AI Financial Controller

Flux commentary is the part of the close that punishes competence. The better you understand your numbers, the more lines you feel obligated to explain properly — and explaining twenty-two variances in board-ready prose takes hours that have nothing to do with whether you understood the business. This is the highest-leverage place to bring AI into a controller's workflow, and it's also the one most controllers get wrong on the first attempt, usually by giving the model too little structure and too much trust.

This article is the deep dive promised in the Complete Guide: a concrete before/after, the exact inputs that work, and a prompt you can run this period.

What flux commentary actually costs you

For a mid-complexity P&L with 30-50 lines that move enough to require explanation, the manual workflow looks like this: pull the variance report, sort by materiality, research the ones you don't already know the answer to, write one to two sentences per line in a consistent tone, then reformat for whichever audience — CFO, board, lender — needs to read it next.

The research time doesn't change — AI doesn't know why the freight line moved unless you tell it. What disappears is the drafting time: the part where you already know the answer and are simply turning it into a complete, well-formed sentence, twenty-two times, in a consistent register.

The three inputs that determine output quality

AI commentary quality is almost entirely a function of what you give it. Three inputs matter most.

The variance table itself. Account name, budget, actual, variance in dollars and percent, prior period for trend context. A clean table, not a screenshot.

Reason codes. Even a two-word fragment per line — "vendor timing," "headcount delay," "one-off legal fee" — turns a generic-sounding draft into a specific, accurate one. This is the single highest-impact addition you can make to the prompt.

A tone and length example. Paste two or three sentences of commentary you've written previously. AI will match the register, the level of formality, and roughly the sentence length — which matters more for board-pack consistency than people expect.

Skip any of the three and you'll get commentary that reads fluently but says less than it should — accurate in direction, vague in substance.

The prompt

You are drafting variance commentary for a management accounts package.

Variance table (Account | Budget | Actual | Variance $ | Variance % | Reason code):
[PASTE YOUR TABLE HERE]

Tone and length example (match this style):
[PASTE 2-3 SENTENCES OF YOUR PRIOR COMMENTARY]

Write one commentary line per account that has a variance above [YOUR MATERIALITY THRESHOLD]. Each line should:
- State the direction and magnitude in plain terms
- Explain the cause using the reason code provided — do not invent a cause not implied by the reason code
- Match the tone and sentence length of the example provided
- Flag with [VERIFY] any line where the reason code is missing or unclear

Do not comment on lines below the materiality threshold. Do not editorialize on whether a variance is good or bad — state what happened.

Run it once with two or three months of historical data before relying on it live. Compare the output line by line against what you actually wrote that period. The gaps you find — usually around lines with thin reason codes — tell you exactly what to tighten before the next close.

Where the human checkpoint sits

Every line, not a sample. AI gets the direction and the stated cause right almost every time; what it cannot do is know the context you didn't write down — the customer who moved their renewal date, the policy your CFO mentioned verbally last week, the fact that the "one-off legal fee" was actually the second one-off this year and worth flagging as a trend. That judgment is the part of flux commentary that was never really about typing speed, and it stays entirely with you.

For the governance layer around using AI in commentary — what to log, what never to paste into a model, how to document review — see Building a Claude System for Your Finance Team and the guardrails section of the Month-End Close Playbook.

Calibrating it to your business

The prompt above works on the first try. It gets meaningfully better after two or three cycles of calibration: add your specific materiality threshold, add the two or three reason codes your business uses most often with a one-line definition of each, and add an example of a variance your model got wrong so it learns the boundary. Most controllers reach a stable, reliable version within a single quarter.

That stabilization — from manually drafting every line to reviewing a calibrated first draft — is a small, concrete example of what it means to practice as an AI-Augmented Financial Controller: the judgment stays yours, the assembly doesn't have to.