Let AI Write What Happened. Only You Can Write Why It Matters.
Let AI Write What Happened. Only You Can Write Why It Matters.
Variance commentary is one of the better uses of AI in a finance function, and one of the easiest to get wrong — because the failure mode isn't an error. It's fluent, accurate text that explains nothing.
The September quarter closed last week, which means a lot of finance teams are now writing the commentary that goes with the first management reporting pack of FY27. It is the task most finance leads would hand over first if they could, and with reason: it is slow, it is repetitive, and the numbers are already sitting in a table.
It is also a good fit for a language model. The inputs are structured, the output is prose, and the shape of the task barely changes from quarter to quarter. Teams that have put AI on it report the obvious benefit — the first draft of a page and a half appears in minutes rather than over an afternoon.
And then a specific thing goes wrong, which is worth naming precisely, because it does not look like a mistake.
The failure mode is restatement, not error
Give a model a variance table and ask it for commentary, and it will give you something like this:
"Employee expenses were $148,000 above budget for the quarter, an unfavourable variance of 6.2%. This was driven by higher than anticipated staffing costs. Management continues to monitor employee expenses closely."
Nothing in that is wrong. The number is right, the percentage is right, the direction is right, and it reads like every set of accounts you have ever seen. It is also worth nothing to the person reading it, because it contains no information that is not already in the table two pages earlier.
"Driven by higher than anticipated staffing costs" is the variance restated as its own cause. "Management continues to monitor" commits to nothing. A reader who wanted to know whether this recurs next quarter, whether it was a decision or a surprise, and whether anything needs to change, learns none of those things.
This is the characteristic output of a model given a table and no drivers. It cannot know why the number moved, so it paraphrases the number — in confident, well-formed sentences that pass a quick read. The error is invisible precisely because the facts are correct.
Which tells you exactly how to use the tool. The model's job is the part it can actually do: say what happened, accurately, in consistent language, with the figures right. Explaining why requires knowledge that exists only in the heads of the people who ran the quarter — and the whole method follows from getting that knowledge into the process.
The method: four inputs, one hard rule
The difference between useful and useless output here is almost entirely about what you put in.
Input one: the variance table itself — actual, budget or reforecast, variance in dollars and percent, for the quarter and year to date. Clean, with consistent account descriptions.
Input two: a driver list. This is the input that does the work and the one teams skip. Before generating anything, spend fifteen minutes writing a short list of what you already know happened: two staff on extended leave across July and August, the agency rate rose in July, a grant acquittal moved into October, the insurance renewal came in above the budgeted figure, a program started a month later than planned. Ten or twelve lines of plain English. If you cannot write any of them, the commentary is not the problem — nobody knows what happened yet, and you need a conversation with operations before you write a word.
Input three: last quarter's commentary, for tone, structure and the house conventions. This is what stops the output reading like it came from somewhere else.
Input four: the materiality threshold — the dollar and percentage level below which a variance does not get a comment. Without it you get a paragraph on every line, including the ones that moved by $600.
Then one rule, and it is not negotiable: every sentence must cite the account and the amount it refers to. No general statements, no summarising sentences without a number attached, no "overall, costs were well controlled". This rule is what makes the draft checkable — you can verify each sentence against the ledger in seconds — and it also makes an invented cause much easier to catch, because a sentence tied to a specific figure has nowhere to hide.
Add one more instruction: where the driver list does not explain a variance, the model must say so explicitly and leave it for you. An output that says "the $31,000 movement in property costs is not explained by the drivers provided" is far more valuable than a plausible guess, and models will produce the guess by default unless told not to.
What the same variance looks like done properly
Same number, with a driver list in the inputs and a human finishing it. The following is illustrative, not from any organisation's accounts:
"Employee expenses were $148,000 (6.2%) above budget for the quarter. Agency cover for extended leave accounts for $121,000 of this, concentrated in two sites across July and August; both staff have now returned. The remaining $27,000 reflects the agency hourly rate increase from 1 July, which was not in the May budget and will recur at approximately $27,000 per quarter for the remainder of FY27. So what: the leave cover is behind us, but the rate increase is permanent and adds about $81,000 across the remaining three quarters. Decision sought: approve an $81,000 reforecast of employee expenses to absorb the rate increase, with the offset identified in the property line."
Everything before the "so what" — the figures and the drivers attached to them — is the model's work. The "so what" and the "decision sought" are not, and could not be — they require knowing what the board can act on and what the organisation intends to do.
That split is the whole method. The model writes what happened. The finance lead writes what it means and what is being asked. If the second half is missing, you have sent the board a longer version of the table.
The check that cannot be skipped
Every number in the generated text gets traced back to the ledger before the pack goes anywhere. Not a sample — every number.
This sounds excessive until you have seen the specific way it fails. A model handling a table of forty accounts will occasionally attach the right figure to the wrong account, or combine two lines into a total that appears nowhere in your ledger, or carry a year-to-date number into a sentence about the quarter. The sentence reads perfectly. The figure exists somewhere in your data. It is simply in the wrong sentence, and no amount of re-reading the prose will catch it, because the prose is not where the error is.
The citation rule above makes this check fast rather than painful: each sentence names one account and one amount, so verification is a mechanical pass down the page against the table. Ten minutes for a page and a half.
This is the same discipline that applies to any AI output a finance function relies on — verify the evidence, not just the conclusion — and quarterly commentary is where it gets skipped most often, because the output is prose and prose feels like it has been read once you have read it.
On what you are pasting in: a variance table plus a driver list is a fairly complete picture of your organisation's financial position, and driver lists tend to contain staff details — who was on leave, who left, which manager made which call. Confirm whether your AI vendor retains customer inputs or uses them to train future models, and check the setting rather than assuming the default; the safest posture is a tool where your data is not retained for training. Write driver lines by role rather than by name, and keep participant, resident and client identifiers out entirely.
This post is general commentary based on publicly available information and does not constitute legal or tax advice. Always seek independent professional advice before acting.
What a reviewer should reject on sight
Worth pinning up next to whoever reviews the pack. Any of these means the draft goes back:
A cause that is the variance restated. "Higher than expected costs", "lower than anticipated revenue", "increased activity" — these describe the number again in words.
A sentence with no figure in it. If it cannot be checked, it should not be there.
"Management continues to monitor." Either something is being done, in which case say what, or nothing is, in which case say that.
A number that is not in the ledger. Including subtotals the model helpfully computed.
No distinction between one-off and recurring. This is the single most useful thing commentary can tell a board, and it is the thing a model cannot infer without being told.
No decision sought, anywhere in the pack. A quarterly report with no decision asked of the board is a historical document.
What this is actually worth
Be clear-eyed about the gain. Done this way, the time saved on a quarterly commentary is perhaps half to two-thirds of the drafting, and the fifteen minutes spent on the driver list is new work that did not exist before. The net saving on one quarter is a couple of hours.
The better return is consistency. The commentary comes out in the same structure every quarter, with the same materiality threshold applied and the same distinction drawn between one-off and recurring, regardless of who was available to write it or how compressed the timetable was. For a small finance team, that reliability is worth more than the hours — and it is the kind of improvement that makes the second quarter's pack better than the first.
Then spend the recovered time on the half the model cannot do. The "so what" and the decision sought are the only parts of the commentary a board actually needs, and they are the parts that get rushed when the clock runs out on a Thursday afternoon.
Does your quarterly commentary tell the board anything the table doesn't?
One-off versus recurring, and the decision being asked for — that is the part worth paying for. PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations, including management reporting packs written to be acted on.
Talk to PFL →
Comments
Post a Comment