Guides / Guide
Prompt templates that turn a variance spreadsheet into a narrative someone will actually read
Background
Every budget pack has the same problem: the numbers are in the spreadsheet, and the two sentences someone actually reads are in a commentary column that got written in the last ten minutes before the deadline. "Marketing overspent due to higher than expected costs" is technically a sentence about the variance, and it tells a reader precisely nothing they couldn't already see from the number itself.
Ask an AI tool to "write a commentary for this variance" from the raw figures alone and you get the same problem in fluent prose - it can restate that spend was £12,000 over budget, but it has no way of knowing why, because that reason lives in someone's head, not in the spreadsheet. The model will sometimes fill the gap with a plausible-sounding cause it invented, which is worse than an honest blank, because a wrong reason that reads confidently is the kind of thing that ends up repeated in a board pack.
The fix, as with most of these templates, is to separate what the model is good at (structuring and writing clearly) from what only you know (the actual cause and what it means going forward), and build a prompt that makes it ask for the second part rather than guess at it.
Setup
The shape every template below follows:
- Role - a finance business partner writing for people who don't want to open the spreadsheet, not an accountant writing for auditors.
- Context - what this variance report feeds (a monthly pack, a board paper, a department head).
- Inputs - the numbers, and critically, the real cause, in your own words.
- Task - a fixed narrative structure: what moved, why, what it means, what happens next.
- Constraints - no invented causes, no jargon a non-finance reader would have to look up.
The single-line-item variance
For the everyday "this cost centre is over/under, explain it" case:
You're writing a one-paragraph variance commentary for a management
report, aimed at a reader who isn't a finance specialist.
Inputs:
- Line item and period: [e.g. MARKETING - SEPTEMBER]
- Budget vs actual: [FIGURES]
- The real reason it moved, in your own words - however rough:
[DESCRIBE - e.g. "WE BROUGHT A CAMPAIGN FORWARD FROM OCTOBER"]
- Is this a timing difference (will reverse), a one-off, or a genuine
run-rate change: [STATE WHICH]
- What, if anything, needs to happen as a result: [DESCRIBE OR "NOTHING,
FLAGGING FOR AWARENESS"]
Task: one tight paragraph - what moved and by how much, why (using my
reason, not an invented one), whether it's timing/one-off/run-rate, and
the action or "no action needed" line.
Constraints: no cause I didn't give you. No finance jargon that needs
translating for a non-finance reader ("phasing", "accrual reversal") without
a plain-English gloss in brackets the first time it's used. One paragraph,
not three.
The full department pack
For rolling several line items into one coherent section rather than a wall of separate paragraphs:
You're writing the variance section of a monthly management pack covering
a whole department, from several individual line items.
Inputs:
- Department and period: [NAME, PERIOD]
- Each line item, with budget/actual and the real cause in your own words:
[LIST - e.g. "STAFF COSTS: £8K UNDER, UNFILLED VACANCY, NOT YET
RECRUITED. TRAVEL: £3K OVER, TEAM OFFSITE BROUGHT FORWARD."]
- Overall department position: [OVER/UNDER, FIGURE]
Task: lead with the overall position in one sentence, then group the line
items into "worth knowing about" (material, or a change in trend) and "in
line with expectations, noted for completeness" - don't give equal weight
to a £200 rounding difference and a £15,000 genuine overspend.
Constraints: every cause traces to what I gave you. Flag if two line items'
causes look like they might actually be the same underlying thing showing
up twice (e.g. a delayed hire explaining both a staff-cost underspend and a
recruitment-fee overspend) - that's worth saying explicitly, not leaving as
two separate paragraphs that don't obviously connect.
The board-level rollup
For summarising many departmental variances into the handful of lines a board actually has time to read:
You're compressing several departmental variance commentaries into a
board-level summary - three to five sentences total, not a paragraph per
department.
Inputs:
- Overall budget vs actual position: [FIGURES]
- Each department's variance and cause, already written: [PASTE THE
DEPARTMENT-LEVEL COMMENTARY FROM ABOVE]
- Anything that changes the full-year forecast, specifically: [DESCRIBE
OR "NO CHANGE TO FULL-YEAR VIEW"]
Task: lead with the overall position and whether it changes the full-year
forecast. Then only the two or three variances that are genuinely material
at board level - by size or because they signal a trend, not because they
happen to have an interesting story. Everything else gets one sentence
("all other departments within 2% of budget") rather than individual
mention.
Constraints: don't let an interesting-sounding but immaterial variance
crowd out a boring but large one. If you're unsure what counts as material
at this size of organisation, ask me for the threshold rather than guessing.Edge Cases & BCP
- The model will invent a plausible cause if you let it, and it will sound completely reasonable. "Likely due to seasonal demand" reads fine in a board pack and is worthless if nobody actually checked that's true. Never accept a narrative that states a cause you didn't explicitly provide - re-read every draft against your own inputs before it goes anywhere.
- A variance that's actually a timing difference, written up as if it's a permanent change, misleads a reader into the wrong conclusion about the full-year position. Always make the model state the timing/one-off/run-rate distinction explicitly rather than leaving it implied.
- Don't paste real supplier names, individual salaries, or other commercially or personally sensitive figures into a general-purpose AI tool unless you've checked what your organisation's data policy actually allows for that tool - aggregate line-item figures are usually fine; a named individual's pay rarely is.
- A well-written narrative makes a genuinely concerning variance sound managed and under control, purely because fluent prose reads as competent. Read the commentary and ask "does this actually tell the reader something needs to change," not just "does this read well."