Writing Variance Commentary With AI

Variance commentary — the written explanation of why actual results differ from budget, forecast or prior period — is one of the most time-consuming parts of the monthly reporting cycle, and one of the tasks where AI genuinely helps. Turning a set of checked numbers and driver notes into clear, consistent prose is exactly the kind of writing a large language model does well. The catch is equally clear: AI is very good at producing explanations that sound right. If it is asked to explain a variance it has no evidence for, it will often supply a plausible cause anyway, and plausible-but-wrong commentary is worse than none.

This guide is a practical method for writing variance commentary with AI without that risk. It covers what to feed the tool, how to structure the prompt, a fully worked illustrative example (a variance table that adds up, the AI’s draft and the reviewer’s edits), and the controls that keep the commentary honest. It assumes the underlying analysis is already sound — if that is the weak point, start with our guide to variance analysis that drives decisions, because AI can only articulate an analysis, not replace one.

What AI Should and Should Not Do in Variance Commentary

The simplest way to use AI safely here is to split the job in two. The analysis — establishing what drove each variance — belongs to the finance professional, working from the ledger, operational data and conversations with budget holders. The articulation — turning those findings into readable, consistently formatted commentary at the right length for the audience — is where AI saves time.

• Good uses: drafting commentary from a checked variance table and driver notes; applying a house style and materiality threshold consistently; shortening a long draft for a board summary; flagging lines where the notes do not explain the variance; checking that every number quoted in the text matches the table.
• Poor uses: asking the tool to work out why a cost line moved; letting it calculate the variances or splits it then writes about; accepting phrases such as “due to timing” or “driven by market conditions” that nobody has evidenced.

Held to that split, AI typically cuts drafting time substantially on a routine month while the quality of the commentary goes up, because the discipline of writing driver notes for the tool forces clearer thinking. For the wider picture of where AI is safe in finance, see where AI helps and where it is dangerous.

The Inputs: What to Give the AI

The quality of AI-drafted commentary depends almost entirely on the inputs. Three are essential, and a fourth makes the output much better.

1. A checked variance table

Give the tool the final numbers, not raw ledger exports: budget (or forecast or prior period), actual and variance for each line, with subtotals. Do the arithmetic in your spreadsheet or reporting tool and confirm it foots before anything goes into the prompt. State the sign convention explicitly — for example “positive = favourable to profit” — because language models handle signs inconsistently if left to infer them. Where a price/volume or rate/efficiency split matters, calculate it yourself and include it as extra columns.

2. Driver notes

For each material variance, a line or two of evidenced explanation, ideally quantified: “Customer A deferred an order of about 2,000 units into next month — confirmed by the Sales Director.” These notes come from the finance team’s own investigation and from budget holders. They are the substance; the AI’s job is to express them. Where you do not yet know why a line moved, say so in the notes — that is far better than leaving a gap the tool will fill.

3. Context notes

Short background the reader will need: one-off events, changes in the budget basis, known timing differences, whether a variance is expected to reverse, and any commentary already given last month that should be followed up. This lets the draft distinguish timing from permanent variances and keeps the story consistent month to month.

4. Audience and house style

Who reads it (board, executive team, budget holders), the length, the materiality threshold for comment, and a previous month’s approved commentary as a style example. Commentary for a board pack should lead with the overall message; commentary for budget holders can go line by line. Our guide to management reporting that gets read covers what each audience actually wants.

Before pasting anything, check that the tool is approved for the data involved. Management accounts are usually confidential and, for listed companies, may be price-sensitive. Use only tools your AI usage policy permits for that class of data, and remove names of individuals where they add nothing; our guide to AI data security in finance explains why. Where commentary touches personal data — payroll variances by named employee, for instance — UK GDPR applies, and the Information Commissioner’s Office has published guidance on AI and data protection. If your IT team is connecting an assistant directly to ledger or reporting data, the National Cyber Security Centre’s guidelines for secure AI system development are useful background.

A Prompt Structure That Works

A good variance commentary prompt has six parts: role, task, inputs, rules, format and a self-check. The rules section does most of the work, because it is where you forbid invented explanations and tell the tool how to handle gaps. The template below can be saved and reused each month; for the general principles behind it, see prompt engineering for finance, and for more templates our finance prompt library.

The self-check is worth keeping. It does not make the output trustworthy on its own, but it gives the reviewer a fast map of what to verify and tends to surface any drift from the inputs.

Worked Example: From Variance Table to Reviewed Commentary

The following example is illustrative — a fictional distribution business, with figures constructed for this guide — but it follows exactly the process above. The business budgets to sell 40,000 units in the month at an average £30 a unit, with a standard cost of £19.50 a unit.

Step 1: The variance table

The table foots: overheads total £295k budget and £302k actual, a £7k adverse variance, so EBITDA is £12k adverse at gross profit plus £7k adverse in overheads, giving £19k adverse. The finance team has also split the two largest lines before prompting:

• Revenue (£66k adverse): all volume. 37,800 units were sold against 40,000 budgeted, a shortfall of 2,200 units at £30 = £66k. Average price was on budget.
• Cost of sales (£54k favourable): £42.9k from lower volume (2,200 units at £19.50) and £11.1k from a lower actual cost per unit (budget cost of actual volume £737.1k against actual £726k). Gross margin was 36.0% against 35.0% budget.

Step 2: Driver and context notes

• Revenue: Customer A deferred an order of about 2,000 units to next month (confirmed by Sales Director; order is booked). Remaining c.200 units: no specific cause identified.
• Cost of sales: new freight contract from the 1st of the month reduced carriage costs by about £8k. Remaining c.£3k of the unit-cost saving not yet analysed.
• Staff costs: agency cover for two warehouse vacancies cost £9k. Both posts are being recruited. Remaining £3k not yet explained.
• Marketing: trade show (£9k) moved from this month to next. Timing only.
• Other overheads: no note yet.
• Materiality for comment: £5k or 5% of budget. Audience: executive team. Maximum 200 words.

Step 3: The AI draft

Run through the prompt template, a well-behaved assistant produces something like this:

Step 4: The reviewer’s edits

The draft is good, but a careful reviewer still makes changes. This is the step that cannot be skipped.

The first edit is the important lesson. Even with an explicit rule against inferring causes, the tool added a generic explanation for the unexplained part of a variance. That is normal behaviour for language models, and it is why review against the driver notes — sentence by sentence for anything material — is a control, not a courtesy. It also shows the review adding value the tool cannot: the margin effect of the deferred order and the reversal next month are the two points the executive team most needs.

Controls: Keeping AI Commentary Honest

The worked example points to a short set of controls that should be written into the monthly process, not left to individual habit. They fit naturally within a wider human-in-the-loop control framework.

• No explanation without evidence. Every causal statement in the final commentary must trace to a driver or context note. If it cannot, it is removed or replaced with an open item.
• Mark gaps visibly. Use a consistent marker such as [EXPLANATION NEEDED] in drafts, and either close or explicitly report every one before issue. Unexplained is an acceptable answer; invented is not.
• Numbers come from the table, not the tool. Do calculations in the spreadsheet or reporting system. Tie every figure in the text back to the table before sign-off — the self-check list makes this quick.
• Timing versus permanent is a judgement. Only the finance team decides whether a variance will reverse. Do not let the draft label anything as timing unless the notes say so.
• Named reviewer and sign-off. The person who signs the management accounts owns the commentary, whoever or whatever drafted it. Keep the prompt, inputs and approved output on file for the month.
• Approved tools only. Use tools permitted for confidential financial data under your policy, and note in the file which tool was used.

Auditors are starting to ask how AI is used in reporting processes, even where management commentary is not itself audited. A simple record of prompts, inputs and review is usually enough to answer them; our guide to what auditors ask about AI covers the questions. ICAEW’s generative AI guide is also a useful reference on the professional responsibilities that sit with the accountant, not the tool.

Common Mistakes

• Pasting a trial balance and asking “explain the variances”. This invites the tool to guess. Give it a checked table and notes instead.
• Letting the tool do arithmetic. Language models can misread signs, transpose figures or round inconsistently. Calculate first, then prompt.
• Commentary that restates the table. “Revenue was £66k below budget” tells the reader nothing the table did not. Push for driver, impact and action.
• Generic filler. “Market conditions”, “phasing” and “one-off items” without specifics are signs the commentary has not been reviewed against evidence.
• Losing the thread month to month. Include last month’s commentary so the draft follows up on items flagged as timing or under investigation.
• Skipping review because the draft reads well. Fluency is not accuracy. Polished prose makes errors harder, not easier, to spot.

Where This Fits in the Reporting Cycle

Variance commentary is one part of a wider reporting narrative. The same method — checked numbers, evidenced notes, a structured prompt, documented review — extends to the full management accounts narrative and board papers; see our guides to AI for management reporting narrative and AI for board pack drafting. Upstream, a faster, cleaner close leaves more time for the investigation that produces good driver notes; our guide to AI for the month-end close covers that end of the process.

The skill that matters most is unchanged: understanding the business well enough to know why the numbers moved. That is why employers still value strong management accountants and business partners who talk to budget holders and dig into operational data. AI changes how long the writing takes, not who is responsible for the explanation. Our guide to management accountant AI automation looks at how the role is shifting more broadly.

Frequently Asked Questions

Can AI write variance commentary on its own?

It can write fluent commentary, but it cannot know why a variance happened unless you tell it. Without driver notes it will tend to supply plausible causes. Use it to articulate explanations your team has evidenced, and review every causal statement.

Which AI tool is best for variance commentary?

Any capable general-purpose assistant your organisation has approved for confidential data will do the drafting well; the inputs and controls matter far more than the tool. Some planning platforms also generate narrative inside the system. Our comparison of Claude, Copilot and ChatGPT for finance and AI tool landscape for finance cover the options.

How do I stop AI inventing explanations?

Give it only checked figures and evidenced notes, instruct it explicitly not to infer causes, require a visible marker such as [EXPLANATION NEEDED] for gaps, and ask for a self-check listing unsupported statements. Then review against the notes anyway — rules reduce invented content but do not eliminate it.

Should the AI calculate the variances?

No. Calculate variances and any price/volume or rate/efficiency splits in your spreadsheet or reporting tool, confirm the table foots, and give the AI the results. Ask it to quote figures exactly and to list each one it used, so you can tie them back.

Is it safe to put management accounts into an AI tool?

Only into tools your organisation has approved for confidential financial data, with appropriate contractual terms. Remove personal data that is not needed. If you are a listed company, treat unpublished results as inside information and follow your policy strictly.

How much time does it save?

It varies with the business and how good the driver notes are. The drafting itself becomes much quicker; the investigation and review do not. Teams usually find the overall gain is real but moderate, and that commentary quality and consistency improve.