How to Check AI Variance Commentary Before It Reaches the Board

A pattern we see often: a finance manager pastes the month's budget vs actual into ChatGPT or Copilot, asks for board commentary, and gets four tidy paragraphs back. Three are right. The fourth puts a revenue shortfall down to "softer market demand", a phrase nobody in the business used. The real cause was two customer go-lives slipping by a month. The paragraph reads well, which is exactly why it nearly goes into the pack.

The problem is rarely that AI writes badly. It writes fluently, and a fluent paragraph with one invented driver is harder to catch than a clumsy one. This guide gives you a method for catching it, built from the way we teach variance commentary on our AI for Finance Leaders course. For the wider picture of where AI fits in planning and reporting, start with our guide to AI in FP&A.

Why AI Variance Commentary Goes Wrong

A language model predicts plausible text. Give it a table with revenue £70k below budget and it will produce a plausible reason, whether or not you supplied one. It has no access to the conversation you had with the sales director on day two of close. Everything it knows about your month is what you pasted in.

That produces four kinds of error, in roughly this order of frequency in the drafts we review:

• Invented drivers. "Market conditions", "seasonality" and "pricing pressure" are the usual suspects. They fill a gap where you gave no reason.
• Sign errors on cost lines. A cost below budget is favourable. Models sometimes call it adverse because the variance is negative.
• Arithmetic on the wrong base. A percentage calculated on actual instead of budget, or a margin movement quoted in pounds when the board reads points.
• Made-up outlook. A confident sentence about next month that nobody in the business has said.

None of these is fixed by a better model. They are fixed by giving the model less room to guess, and by checking the output in a structured way rather than by reading it and feeling reassured.

The Sample Month

The method is easiest to learn on a month where you already know the right answer. Here is the one in our kit. The figures are sample data in £000, with a materiality threshold of £10k.

And here is what the finance team knows about the month, which is the part the AI cannot see unless you give it:

1. Two customer go-lives slipped to next month. They were budgeted at £60k of revenue.
2. Those go-lives include implementation fees, which carry a higher margin than recurring revenue.
3. The rest of the revenue gap is lower volume across existing accounts.
4. Payroll includes two contractors covering vacancies until the roles are filled.
5. The marketing campaign moved to next month. Timing only.
6. Software is new planning-tool licences, below the threshold.

Prompt 1: Check the Variances Before Writing

The first prompt does no writing at all. It asks the model to recalculate every line and rank what is material. This catches a broken formula in your own export before it becomes a broken sentence, and it means the drafting step starts from a list you have agreed with.

You are a UK management accountant reviewing monthly results.

Using only the table below, recalculate every variance and variance % yourself. List any line where your figure differs from the table. Then list the variances above £10k, largest first, stating whether each is favourable or adverse.

Do not explain causes and do not write commentary yet.

[paste the budget vs actual table]

On the sample month you should get back five material lines: revenue, gross profit, cost of sales, marketing and payroll, plus EBITDA. If the model calls cost of sales or marketing adverse, you have found the sign problem before it reached the draft, and you know to watch for it.

Prompt 2: Draft With Only Your Drivers

Now the draft. Two instructions do most of the work. The first restricts the model to the drivers in your notes. The second gives it a way out when a driver is missing, so it writes a visible placeholder instead of a guess. The example line shows it what your house style looks like, which matters more than any adjective about tone. This is the RACEF structure we use for most finance prompts.

Role: you are a UK finance business partner writing the monthly commentary for the board pack.

Action: write commentary on the material variances (over £10k).

Context: [table] [context notes]. Use ONLY the drivers in the context notes. If a material variance has no driver in the notes, write [DRIVER NEEDED] instead of guessing.

Example of the house style: "Payroll £210k vs £195k, £15k adverse: two contractors covering vacancies."

Format: one opening sentence on EBITDA; one short paragraph per material variance (figure, driver, timing or permanent); one sentence on next month. Plain English, under 200 words, figures in £k.

A [DRIVER NEEDED] tag in the output is a good result. It tells you exactly which conversation you still need to have, rather than hiding the gap inside a confident sentence. For more templates in the same structure, covering revenue, cost, price-volume and headcount variances, see our variance commentary prompts.

Prompt 3: Make It Audit Its Own Claims

The third prompt turns the model from writer into reviewer. Checking one specific claim against a quoted source is a narrower task than writing well, and in our experience models do it more reliably. Run it in a fresh chat if you can, so the reviewer is not anchored on the draft it just wrote.

You are reviewing a draft before it goes to the board. Below are the source table, the context notes and the draft.

For every sentence, list each claim it makes (a figure, a driver or a prediction) and mark it:
SUPPORTED (quote the table line or note it comes from)
CALCULATED (show the sum)
UNSUPPORTED (no source)

Then rewrite only the sentences with an UNSUPPORTED claim, using the sources or [DRIVER NEEDED].

[table] [notes] [draft]

The output is a claim register: every number and every reason in the draft, each with its source. That is also the document you want if an auditor or a non-executive director later asks where a sentence came from. Keep it with the month-end file.

The Human Check Before Sign-Off

The prompts reduce errors. They do not remove responsibility. Before the commentary goes anywhere, the person signing it off works through eight checks:

1. Every figure matches the source table, or is a sum you have recalculated.
2. Every driver named is in your context notes.
3. Favourable and adverse are the right way round on cost lines.
4. Timing differences are labelled as timing, with the month they reverse.
5. One-off items are labelled as one-off.
6. Nothing below materiality has been given a paragraph it does not need.
7. Any forward-looking sentence is something the business has actually said.
8. You could defend every sentence if the CFO asked where it came from.

Want to see how the full method works first? Try the free course preview.

Frequently Asked Questions

Variance commentary is one of the workflows we build with finance teams. Our AI for Finance Leaders course teaches it with marked practice, our AI audit and assessment finds where it would save your team the most time, and our AI consulting team can set it up inside your close.