How to Write a Data Narrative Using AI From Raw Numbers

Why Raw Numbers Get Ignored

A metric on its own carries no instruction. “Churn was 4.1% in Q3” tells a reader something happened, but it does not tell them whether to worry, celebrate, or do nothing at all. That interpretive work still has to happen somewhere, and if you don’t do it, your reader does — usually badly, and usually in a meeting where you’re not present to correct them.

Most senior people will tell you the same thing about reports they receive: the ones they remember are the ones that ended in a recommendation. The numbers were the evidence, not the point.

Same underlying data. The second version tells a reader what to do, and it took the writer maybe four extra minutes. That gap — four minutes of interpretation — is where nearly all the perceived value of a report lives.

Ask any operations lead why they stopped reading a weekly dashboard and you get a version of the same answer: it never told them anything had to change. A finance manager will say it differently, but it amounts to the same complaint. Numbers without a verdict become background noise, and background noise is what people learn to skip.

What the Count-Then-Narrate Method Is

The Count-Then-Narrate Method is a two-phase discipline: you make the AI finish and display every calculation before it writes a word of prose, then you make it narrate the verified figures in a fixed three-part shape. It’s the approach we teach at PromptPeakAI for any task where wrong numbers would be more embarrassing than no numbers.

• Count. The tool lists every metric, calculates each change, and displays its working. No commentary, no conclusions, no prose of any kind.
• Narrate. A second, separate instruction turns those settled figures into the key finding, the supporting data, and the implication.

The order matters more than it sounds. Language models generate text by predicting what comes next, which is a poor way to compute a percentage change. When arithmetic and prose happen in the same breath, the prose tends to win and the arithmetic drifts. Separating them removes the failure mode almost entirely.

Most guides to write a data narrative using AI hand you one long prompt and hope for the best. That single-prompt approach is exactly what produces confident, well-written reports with a wrong number buried in paragraph two — the worst possible combination, because polish makes the error harder to spot.

Stage four is not optional padding. It’s the reason you can put your name on the output.

The Three Planks of a Narrative

Every data narrative you write should stand on the same three planks, in the same order. The fixed order is what makes the output usable — a reader learns the shape after one report and can then skim any future one in fifteen seconds.

You’ll notice a fourth row. In practice the implication and the next step are often written as one sentence, which is why the framework is named for three planks — but separating them while drafting stops the AI from producing a vague “this should be monitored closely” ending. Vagueness at the end is the most common failure in AI-written reports.

Prepare the Spreadsheet First

Roughly half the complaints about AI misreading data are actually complaints about export formatting. Merged cells, three stacked header rows, currency symbols glued to figures, and blank spacer columns all confuse the parser before any analysis begins.

Clean the file

• One header row at the top. Delete decorative title rows above it.
• Unmerge every merged cell — merged headers are the single biggest cause of misread columns.
• Strip currency symbols and thousands separators so 1200 stays a number.
• Delete blank spacer columns and any footer totals row.
• Save as CSV rather than a formatted workbook when the file is large.

If your export is genuinely messy, it’s often faster to have the AI do this pass for you first — our guide to cleaning messy data using AI covers that as a standalone job before any narrative work starts.

Redact before you upload

The privacy question is the one that stops most people, and the honest answer is narrower than the panic suggests. The real line is not free versus paid — it is a consumer account versus a company-managed business account. On a business or enterprise account your organisation controls retention and training; on a personal account you’re relying on your own settings.

Anonymising costs almost nothing in analytical value. Replace names with “Client A” and “Client B” using find-and-replace, delete the identifier column entirely, and the narrative reads exactly the same. Our fuller breakdown of whether ChatGPT is safe to use at work walks through the account-type distinction in detail.

How to Write a Data Narrative Using AI

Here is the run itself — four steps to write a data narrative using AI from a file you already have. Budget roughly twenty-five minutes the first time and about five once the prompts are saved. Everything below assumes a cleaned export and a tool you are allowed to use with this data.

Upload the cleaned export, or paste the table directly if it’s small — under about a hundred rows pastes fine as plain text. Then set the context before asking for anything. The tool needs to know who reads this and what decision it feeds.

This is the step everyone skips and the one that prevents the embarrassing error. You are explicitly forbidding prose until the calculations are done and visible.

You are a business analyst. I have attached [monthly sales export / campaign report / survey summary].

Before writing any narrative text, do this and nothing else:
1. List every metric in the file with its value for the current period and the prior period.
2. Calculate the absolute change and the percentage change for each one. Show the calculation.
3. Flag the three largest movements by percentage change.
4. Flag anything that looks like a data error rather than a real movement.

Rules:
- Use ONLY numbers present in the attached file.
- If a figure is missing, write "not provided" rather than estimating it.
- Do not write any summary, commentary, or conclusions yet.

Stop after step 4 and wait for my next instruction.

Only now does the writing start. The second prompt references the figures already on screen, which is why the numbers in the narrative match the numbers in your spreadsheet.

Using ONLY the figures you calculated above, write a data narrative for [my head of department / the executive team / a client].

Structure it in exactly three parts, labelled:

KEY FINDING: One sentence naming the single most important change and its percentage.

SUPPORTING DATA: Two or three sentences giving the specific figures that prove it, including the largest contributor and the largest drag.

IMPLICATION: Two sentences — what this means for [revenue / headcount / margin / retention], and one specific action for next month.

Constraints:
- Plain business English. No jargon, no hedging, no "it is worth noting".
- Do not introduce any number that is not in your calculations above.
- Maximum 150 words total.
- If the data does not support a clear recommendation, say so instead of inventing one.

The same verified figures usually need to go to two audiences at different altitudes. Rather than rewriting, ask for a translation — the numbers are already locked, so this pass only changes register.

Rewrite the narrative above for a different reader.

Original was written for: [executive audience].
New version is for: [the delivery team who own the day-to-day work].

Change only the emphasis and vocabulary:
- Keep every figure identical.
- Replace strategic framing with operational detail.
- End with what this team specifically should change in their next sprint or cycle.
- Keep it under 200 words.

Save both prompts somewhere you can reach them next month — that’s most of the time saving. If you’d rather start from ready-made versions, the free AI work templates library includes reporting prompts built on this structure.

Where the Arithmetic Actually Happens

This is the part competing guides skip, and it’s the one that decides whether your narrative survives contact with someone who checks. There are two very different things that can happen when you hand a tool some numbers, and they look identical from the outside.

When a table is pasted as plain text, the model reads it as language and estimates the maths the way it estimates everything else — plausibly, and sometimes wrongly. When a file is uploaded to a tool that can run code, the calculation is genuinely executed. OpenAI’s documentation on data analysis with ChatGPT describes it writing and running Python in a notebook environment for some data tasks, and advises reviewing the generated code and its assumptions before relying on the result.

Free tiers do run data analysis — OpenAI’s own free-tier documentation describes file uploads and data analysis as available, with stricter limits than paid plans. So the honest position is that free works and then runs out, rather than free being locked out. For occasional small tables, don’t pay. For a monthly reporting habit with real file sizes, a paid tier around twenty dollars a month stops being a luxury; our ChatGPT free versus Plus breakdown covers where the line falls.

Prompting Rules That Change the Output

Two prompts can request the same thing and produce wildly different quality. The difference is almost always constraints — what you forbid matters more than what you request.

The second version wins because it removes every decision the model would otherwise make badly — length, structure, scope, and whether to hedge. Banned-word lists and explicit prohibitions do more work here than any amount of polite phrasing, which is why negative constraints deserve a place in every reporting prompt you save.

Audit It Before You Send

Three checks, ninety seconds, and they catch nearly everything that would otherwise surface in a meeting. Run them every time you write a data narrative using AI, including the months when the output looks obviously fine — those are the months a wrong figure travels furthest.

The second gate is the one that catches invented figures. Take every number in the finished narrative and find it in your source file — not approximately, exactly. A fabricated number usually looks completely reasonable, which is precisely why it survives a casual read. The same verification instinct applies whenever you fact-check AI research output before attaching your name to it.

Audit the narrative you just wrote against the source data.

Produce a table with these columns:
Claim | Figure used | Present in the source file? | Notes

List every numeric claim in the narrative as a separate row.
Mark anything you cannot trace directly to the attached file as UNVERIFIED.
Do not correct or rewrite anything — just report.

Then list any statement that implies causation where the data only shows correlation.

That last instruction earns its place. “Leads rose because we increased spend” is a causal claim your spreadsheet almost certainly cannot support, and it’s the kind of sentence that gets challenged by the one person in the room who reads carefully.

When the Narrative Goes Wrong

The failures repeat, which is good news — a fixed set of symptoms means a fixed set of fixes.

The last row deserves a note. When a large export gets truncated, the tool rarely announces it — the narrative simply describes a subset of your data as though it were everything. If your file is long, ask directly how many rows were read before you trust anything downstream. Spotting that kind of silent gap is a close cousin of the work involved in finding patterns in business data, where the same truncation quietly changes conclusions.

What AI Genuinely Cannot Do Here

Being straight about the limits is what makes the rest trustworthy.

There’s also the question people ask quietly: will anyone be able to tell? Not from a well-audited narrative — the figures are yours, the context is yours, and the recommendation is yours. What gives AI-written reports away is the hedging, the passive voice, and the endings that recommend nothing. Fix those and the question stops coming up.

One narrative rarely stays one narrative. Once the figures are verified, the same three planks travel:

• Into the executive summary at the top of a longer document, almost unchanged.
• Into a finance-specific version, which our guide to writing a financial narrative with AI covers on its own terms.
• Into a visual, if the finding lands better as a picture — the same cleaned file will produce the chart too.

Whichever direction it goes, the structure carries the weight. As the team at Storytelling with Data argue, the goal was never to use AI more — it was to say something clearer, and decoration does not rescue a point that was never made.

So run it once on last month’s real export. The first attempt to write a data narrative using AI takes twenty-five minutes and feels slower than doing it by hand. The second takes five. By the third you will wonder how you used to spend an afternoon on it.