Section 1
Analysis finds the claim, narrative carries it
The division of labour is clean once you name it. Analysis tells you what is true, at what confidence, for which segment. Narrative tells a person why it matters and what to do. Founders get into trouble by asking one to do the other job. A dashboard handed to a buyer does not persuade, because they have no idea which of the forty numbers deserves their attention. A story with no verifiable number behind it does not persuade a technical buyer either, because they have learned that a good anecdote costs nothing to produce. The working sequence is analysis first, mechanism second, story third, with the underlying query kept so the number can be reproduced later. Impact organisations face the same evidential burden, discussed in [Storytelling for Social Impact Entrepreneurs](/blog/storytelling-for-social-impact-entrepreneurs).
Section 2
Averages hide the story
An aggregate improvement figure is usually the least interesting output of any analysis, because it blends groups that behaved in opposite directions. Split the same population by segment, by cohort month, or by which feature they adopted first, and the picture separates: one group improved substantially, another did not move, and a third got worse. That separation is where the story lives, because it identifies who the product is actually for and under what conditions. It also protects you from the awkward moment when a prospect from the flat segment asks why they are not seeing the average you advertised.
Section 3
From query to claim
Write the question, run it, record the query and the date, check the segment sizes, then state the claim in one sentence with its conditions attached. Once the claim exists, choosing how to show it matters, and [Visual Storytelling: Using Images and Videos to Enhance Your Message](/blog/visual-storytelling-using-images-and-videos-to-enhance-your-message) covers the presentation layer.
Section 4
Using models without borrowing their confidence
Language models are genuinely useful in this pipeline for a narrow set of jobs: reshaping a finding into several draft framings, summarising a long transcript so you can find the quotation you half-remember, and pushing back on a draft you have read too many times to see clearly. They are unsafe for supplying facts. A drafting tool will produce a plausible number, a plausible client name, and a plausible date with exactly the same fluency it produces the true ones, and nothing in the output marks the difference. Fabricated specifics are the failure that ends a company's credibility rather than damaging it. The control is procedural. Every figure in a published story traces to a query you can rerun, every quotation to a transcript, every client name to a written approval. If a sentence cannot be traced, it comes out.
Section 5
Three ways this fails
The window problem comes first. Choose the start date after seeing the chart and you can produce almost any trend you want. Fix the period before you look, and state it in the story. Second is voice flattening. Run enough drafts through a general-purpose model and everything converges on the same competent, unmemorable register, which defeats the point of telling a story at all. Third is false precision. Reporting a change to two decimal places on a sample of nine customers signals to any numerate reader that you do not understand your own data. Give the sample size and round honestly. Model outputs used for live decisions carry the same verification burden, which is the subject of [Using AI for Real-Time Fraud Detection](/blog/using-ai-for-real-time-fraud-detection).
Section 6
What the evidence shows
The case for pairing data with narrative is well documented. In the classroom experiment made famous by Made to Stick, only 5 percent of listeners remembered an individual statistic after a round of short presentations, while 63 percent remembered the stories (Heath & Heath, 2007). A number on its own does not survive the week. Neuroscience offers a mechanism. Character-driven stories trigger oxytocin synthesis, and the amount released predicts how willing people are to act on the message afterwards (Zak, Harvard Business Review, 2014). Brain-imaging work goes further: during effective storytelling a listener's brain activity becomes coupled with the speaker's, and stronger coupling accompanies better comprehension (Stephens, Silbert & Hasson, PNAS, 2010). The commercial evidence points the same way. Buyers who judged a supplier's information genuinely helpful were 2.8 times more likely to experience purchase ease and three times more likely to buy a bigger deal with less regret (Gartner, 2019). Across the IPA Databank of more than 1,000 campaign cases, emotional campaigns were nearly twice as likely as rational ones to report very large profit gains, 31 percent against 16 percent (Binet & Field, IPA, 2013). Use analytics to find the insight. Use structure to make it land.