How to Write a Financial Narrative Using AI Leaders Actually Read

What leadership actually reads

A financial narrative is the short written explanation that sits beside your numbers and answers three questions: what changed, why it changed, and what happens next. Anything that only answers the first question is a data recap, and executives skim past it because the table already told them that.

Everyone who sets out to write a financial narrative using AI is really trying to close the gap between those two things, and the gap is almost always the middle question. Most finance business partners will tell you the same thing: reports don’t get rejected because a figure is wrong. They get rejected because nobody explained the operational reason behind it, so the reader can’t act on anything.

Read those twice. The second version contains no extra data. It contains one extra fact — the reason — that you already knew and the spreadsheet never held. That is the whole problem in miniature, and it is what the rest of this guide is built around.

Your data before your prompt

Before any of this touches a real month-end file, settle the account question. This is where finance teams get themselves into genuine trouble, and the advice circulating online is mostly wrong.

The common belief is that a free plan trains on your data and a paid plan does not. That is not how it works. According to OpenAI’s own data controls documentation, personal Free, Plus and Pro accounts all have data sharing switched on by default, with an opt-out under Settings, and it is the Business, Enterprise, Edu and API tiers that do not train on your content by default. The dividing line is a personal account versus a work account — not the price you pay.

The Microsoft side works differently but lands somewhere similar. Microsoft’s Copilot privacy documentation states that prompts, responses and data accessed through Microsoft Graph are not used to train its foundation models, and that Copilot only surfaces organisational data the individual user already has permission to open. That is a licensed work product, though — it is not the free chat sitting in your browser.

Whatever account you are on, the practical rule below covers you. It is worth reading alongside our fuller guide on what changes when you use AI on a work account.

Anonymising takes two minutes

You rarely need real names for the AI to write good commentary. Replace cost centre names with generic labels, strip account codes, and drop any column the narrative will not reference. The variance amounts and percentages carry all the meaning; the identifiers carry all the risk.

The Driver Column Method explained

Here is the change that fixes the blank-commentary problem for good, and it happens before you write a financial narrative using AI at all. Before your variance table reaches any AI tool, you add one extra column — Driver — holding the operational cause you already know for each material line, and you write “Driver unknown” wherever you genuinely don’t know. This is the Driver Column Method we teach at PromptPeakAI.

Two things happen the moment you do this. The model stops guessing at causes, because the cause is now sitting in the data it was given. And the lines where you wrote “Driver unknown” become a visible to-do list — the exact questions you need to ask a department head before this goes anywhere near a board pack.

The reason step one matters as much as step two: language models are built to produce fluent text, not to be a calculator. Ask one to compute a variance percentage and draft commentary in the same breath and you have introduced an arithmetic risk you had no reason to take. Your spreadsheet already did that job perfectly. Hand over finished numbers and ask only for sentences.

Write the financial narrative using AI

Four steps, start to finish. First run takes about half an hour because you are building the table shape and saving the prompts. After that it is roughly ten minutes a close.

Filter to material lines only — a threshold like 5% or a fixed currency amount keeps this to eight or ten rows instead of forty. Calculate variance amount and variance percentage as columns, so nothing is left for the model to work out. Then flatten it: no merged cells, no nested headers, one header row. If your source file is messy, our guide to cleaning up a spreadsheet before analysis covers the tidy-up.

Add one column at the end. For each row, write the operational cause in plain words — not accounting language. “Agency rate rose 8% at renewal” beats “unfavourable price variance.” Where you don’t know, write “Driver unknown” and leave it. Resist the urge to fill it with a plausible guess; that guess will end up in the narrative wearing the authority of a fact.

Add one more short column, or a tag inside the Driver text, splitting each cause into timing (it moves between periods and washes out) or structural (it is the new baseline and repeats). This single distinction is what turns commentary into something leadership can plan against, and no model can infer it from a number.

A $62k overspend that is entirely timing is a non-event. The same $62k that is structural changes next year’s budget. Identical figure, opposite decision.

Now the prompt. Note what it does not ask for: no calculations, no root-cause analysis, no interpretation beyond what the table states. It is a writing brief, and the constraint block at the bottom does most of the work — the same idea covered in our guide to telling AI what not to do.

You are an FP&A manager writing variance commentary for non-finance executives.

Below is a variance table. Every figure in it is final and already calculated.

Write three short paragraphs:
1. The headline — net effect on the total, in one or two sentences.
2. The drivers — group the material variances by cause, and state for each whether it is timing or structural.
3. What it means next period — the implication that follows from paragraph 2.

Rules:
- Use only what appears in the table. Do not calculate anything.
- Where the Driver column says "Driver unknown", write "cause not yet confirmed" and do not suggest a reason.
- Plain business English. No accounting jargon, no words like impressive, robust, strong or challenging.
- Under 250 words total.

TABLE:
[paste your table here]

Three prompts for common finance moments

The core prompt handles a standard budget-versus-actual table. Three variations cover most of what else lands on a finance calendar. Each assumes you have already done the Driver column work, and each is written to be saved and reused rather than retyped — a habit worth building alongside the rest of our free workplace AI templates.

You are writing the opening summary of a quarterly board pack.

Below are the quarter's headline figures and the confirmed drivers behind them.

Write 200 words, three paragraphs:
- Performance against plan, and the single biggest reason for the gap.
- Cost control: what held, what did not, and which of those repeats.
- The one risk the board should be watching next quarter.

Rules:
- Every claim must map to a line in the input. Do not add context I have not given you.
- Short declarative sentences. No hedging, no adjectives of enthusiasm.
- If a driver is marked unconfirmed, say so plainly rather than omitting the line.

INPUT:
[paste figures and drivers here]

Rewrite the accounting explanation below for a department head with no finance background.

Two sentences maximum. Tell them what happened to their budget and why, in operational terms. Do not use the words accrual, amortisation, true-up, deferral or capitalisation.

Then add one line: "What this means for you: ..."

ACCOUNTING NOTE:
[paste the note here]

Below is a variance narrative I am about to send to my CFO.

Do not rewrite it. Instead, list the three questions a sceptical CFO is most likely to ask about it, and for each one state which specific sentence prompts the question.

Be blunt. I would rather find the weak spot now.

NARRATIVE:
[paste your draft here]

That third one is the highest-value prompt in this article, and almost nobody uses it. Turning the model into a critic of the draft it just wrote catches soft claims and unsupported causes far more reliably than reading your own work at 9pm.

Where AI gets financial commentary wrong

Time to be straight about the limits of writing a financial narrative using AI. There is a real productivity gain here — a narrative that used to take two hours can reasonably take twenty minutes once the table shape is settled — but that gain comes from the drafting, not the thinking. Four things stay yours.

There is a second, quieter risk. A well-written wrong explanation is far more dangerous than an obviously thin one, because it survives review. Fluent prose reads as considered prose, and a plausible invented driver can travel a long way through an organisation before anyone tests it. Our guide to why AI tools state false things confidently goes deeper on the mechanism.

Which tool you use matters less than most comparison articles suggest. The differences that actually show up in this workflow are narrow.

If you live inside spreadsheets all day, running this natively is the bigger win — our walkthrough of using Copilot inside Excel shows how the same steps work without leaving the file.

Check it before the CFO does

A financial narrative using AI is a draft until you have checked it. Three gates, in order. Each one takes under a minute and together they catch nearly everything that would embarrass you in a review meeting.

Gate two is the one people skip and the one that matters most. Take each stated cause in the draft and physically point at the Driver cell it came from. If you cannot find one, the model supplied it — delete the sentence rather than trying to justify it. The same tracing discipline shows up in our guide to verifying AI output before you put your name on it.

The second version looks weaker on the page. It is stronger in the room, because nobody can knock it over with one question. Finance credibility is built on that distinction.

When the output still comes out wrong

Most of these trace to the same root: something the model needed was missing, so it filled the space. The Driver Column Method works because it leaves fewer spaces to fill.

Where this fits in your reporting

Variance commentary is one layer of a reporting stack, and the same Driver principle holds at every level of it — the model writes from what you put in front of it, and the “why” is always something you carried in.

• Above it: the one-page executive summary that compresses the whole pack, and the board report that wraps both.
• Beside it: the analysis written for non-finance stakeholders, where the translation problem is the whole job.
• Below it: the raw numbers, and the wider discipline of turning any dataset into business language.

Try the whole thing once on a month you have already closed and reported, rather than on live figures under deadline. Write a financial narrative using AI for a period where you already know the right answer, then compare the draft against what you actually sent. That single comparison will tell you more than any guide about which part of your close was drafting and which part was judgment — and judgment is the part worth protecting.