
Most marketing, research, and product teams sit on a growing pile of numbers that never make it into a video. Dashboards get screenshotted, charts get animated, and the result is a video that reports instead of persuades. Data storytelling reverses that order: the numbers are chosen after the narrative question is defined, not bolted on at the end.
The practical consequence is a different production sequence. Instead of writing a script and then hunting for a statistic to decorate it, you start with a claim you can defend, assemble the smallest set of evidence that proves it, and only then design the visuals. It is the same discipline a documentary editor applies when cutting a three-hour interview into four minutes: the story is a selection, not a summary.
AI changes the economics of that selection. Drafting alternate narrative frames, testing hooks, generating storyboards, and producing shot lists used to consume days of a small team. Those steps can now run in parallel, which means the scarce resource shifts back to judgment: which claim is worth making, which evidence is honest, and which visual clarifies a number instead of decorating it.
A useful sanity check before you open any tool: if you cannot state your core claim in one sentence that a stranger could repeat back to you, no amount of production polish will rescue the video.
Analysis answers what happened. Narrative answers so what and what now. Most data videos stall at the first question, which is why they feel like narrated slide decks. A three-layer model keeps the script honest:
Consider a subscription product whose churn data shows a spike at day fourteen. The analysis is a chart. The narrative is: most people who leave never reach the moment the product was built for, and that moment currently sits at day twenty-one. Now every chart in the video has a job, and every scene that does not serve the argument gets cut.
The tension layer is what separates a memorable explainer from a forgettable one. Audiences do not remember percentages; they remember a contradiction. If your data contradicts nothing, you either have the wrong dataset or the wrong audience. That is not a failure of the analytics, it is a signal to keep looking before you spend a production budget.
Before scripting, build a small, well-labeled evidence pack. This is the step most teams skip, and it is why AI drafts so often sound generic: the model is asked to invent structure without receiving any.
Start by inventorying what you actually have: analytics exports, survey results, support ticket categories, sales call transcripts, internal benchmarks, published research, competitor teardowns. Then normalize each item into a consistent record. A lightweight schema is enough:
finding: trial users who invite one teammate convert three times more often
source: product analytics cohort export
magnitude: 3.1x
sample: 4200 trials
caveat: correlational, invite behavior may signal intent
visual_potential: funnel comparison, split screen
The caveat field matters more than it looks. Scripts that ignore limitations invite challenges in the comments and undermine the whole argument. Writing the caveat down early means you can decide deliberately whether to state it on screen, in the description, or in a linked appendix.
Next, score candidates before writing a word of narration:
| Criterion | Question to ask | Scale |
|---|---|---|
| Surprise | Does it contradict a common assumption? | 1-5 |
| Stakes | Does it change a decision or budget? | 1-5 |
| Specificity | Is there a concrete number or before/after? | 1-5 |
| Verifiability | Can you cite the source and sample size? | 1-5 |
| Visualizability | Can a viewer grasp it in three seconds? | 1-5 |
Pick three to five findings with the highest totals. A video built on one well-evidenced claim outperforms one that races through twelve. The rest of the evidence pack becomes b-roll for the description, a carousel, or a follow-up video rather than clutter in the main narrative.
Force the whole project through a single sentence: for this audience, the data shows X, therefore they should do Y. If two claims compete, you have two videos. Keep this sentence at the top of the document so every later decision can be tested against it.
A beat sheet is eight to twelve lines describing what the viewer knows after each beat. Not shots, not lines, just knowledge states. Start with the assumption the audience holds, introduce the contradiction, show the evidence, then land the resolution. Beats that only restate a previous beat get deleted here, where deletion is cheap.
Feed the beat sheet, the evidence pack, the audience description, and a tone reference into a script assistant. Ask for three versions: one explanatory, one skeptical, one conversational. Do not paste any of them directly into production. Instead, mark the strongest sentence of each version and rebuild the script around those lines. AI is excellent at producing options and mediocre at producing your voice.
Ask the assistant to convert the approved script into a shot list with columns for timecode, on-screen text, visual description, data asset, and audio. Generating the shot list as structured data rather than prose means you can paste it into a spreadsheet, reorder it, and hand it to an editor or an image generator without retyping.
Read the script aloud while checking every number against the evidence pack. Look for drift: a claim that was hedged in the data but stated absolutely in the narration. Then check visual continuity, meaning the chart colors, on-screen typography, and framing rules stay identical from shot to shot. This pass takes twenty minutes and prevents the most expensive kind of rework.
A data video pipeline has five working stages, and no single tool covers all of them well. Pick deliberately at each stage.
Exploration. Spreadsheets or a BI tool where you can filter, segment, and sanity-check before anything creative happens. Priority: correct numbers and reproducible queries.
Script assistance. A writing assistant that accepts long context and structured input. Priority: does it respect the evidence pack, and can it output variants rather than one definitive draft?
Storyboard and keyframe generation. Image generation with reference images, seeds, and keyframe control. Priority: consistency, not novelty. A model that produces beautiful but inconsistent frames costs more than it saves.
Voice and narration. Synthetic or human. Priority: pronunciation of numbers, units, and product names. Test the script with the actual voice before finalizing word choices.
Assembly. A timeline editor or template system. Priority: how quickly you can version the same video for different audiences.
When comparing options, weigh four criteria: how much control you get over individual shots, whether you can reuse seeds and style references, how cheap an iteration is, and how the tool handles structured input such as JSON or CSV. Tools that accept structured data will always be easier to integrate into a repeatable pipeline.
Vague prompts produce vague visuals. These four patterns cover most of what a data video needs.
Beat expansion. Give the model one beat and the evidence item, then ask for two visual interpretations and one line of narration for each. You get contrast instead of a single guess.
Beat: the audience assumes churn is driven by price.
Evidence: cancellation survey, top reason is setup complexity, 41 percent.
Produce two visual directions and one narration line each.
Tone: calm, specific, no hype. No statistics in the voiceover that are not on screen.
Shot list conversion. Request a table, not prose, and specify the columns explicitly. Models comply far better when the output shape is defined.
Visual metaphor. Ask for metaphors that can be rendered without text: a funnel narrowing, a queue lengthening, a dial turning. Text-heavy generative visuals age badly and are hard to localize.
Continuity anchor. Once you approve a look, write a short style block describing palette, lighting, lens, and texture, and paste it into every subsequent prompt. This single habit removes most of the visual drift that makes AI-assisted videos feel stitched together.
One hard rule: never let an image generator draw your numbers. Charts, axes, and labels belong to real charting software or your motion graphics tool. Generate the texture and the environment, composite the accurate data on top.
Consistency in a data video comes from constraints, not from model quality. Set them early:
Also decide early how the data will be shown: animated bar race, static annotated chart, or a single large number with a supporting caption. Mixing all three in one short video creates cognitive load and makes the argument harder to follow, even when each individual shot is well made.
Production quality is easy to judge and mostly irrelevant. What matters is whether the narrative landed.
Track the usual distribution metrics, but interpret them relative to the claim: three-second view rate tells you whether the hook framed a real tension; average view duration relative to total length tells you whether the middle earned its place; saves and shares tell you whether the insight felt worth keeping. Comments are the richest signal of all. When viewers quote a specific number back at you, the data storytelling worked. When they argue about production value, it did not.
Add a lightweight comprehension check before publishing. Show the draft to five people who match the target audience and ask two questions: what was the main point, and what would you do differently now? If the answers diverge, the script needs another pass, not another render.
Do I need an analyst to make a data video?
No, but you need someone who can validate numbers. A marketer who can write a correct query and read a cohort chart is sufficient for most explainers. What you cannot skip is a second pair of eyes on every figure that appears on screen.
How long should a data-driven video be?
As long as the argument requires and no longer. One claim with strong evidence fits comfortably in sixty to ninety seconds. Multi-finding reports work better as a series of short videos than one long one, because each episode gets its own hook and its own resolution.
Can AI generate the charts too?
It can generate images that look like charts, which is exactly the problem: invented axes and impossible trends. Build charts with real data in real charting tools, export them as assets, and use generative visuals only for backgrounds, textures, and illustrative scenes.
What if my dataset is small or inconclusive?
Small samples can still make good videos if you are honest about the size and treat the finding as a hypothesis rather than a verdict. A narrative built around an open question, with a clear call for more data, often performs better than a shaky claim presented as fact.
How do I keep the script accurate after the data updates?
Separate the evidence pack from the script. When numbers change, update the pack first, then run a diff against the script and the shot list. Keep the caveat field populated so you know which claims are time-sensitive and which are structural.
Should I show methodology on screen?
A single lower-third line naming the source and sample is usually enough for a short video. Longer formats can dedicate a scene to method, placed after the resolution rather than before it. Full methodology belongs in a linked document for the small percentage of viewers who want to verify it.

Alexander
Come progettare thumbnail che fermano lo scroll: strumenti, workflow, criteri di test e ottimizzazione per TikTok e YouTube, con esempi pratici.
Guía práctica para descargar, elegir y reutilizar portadas de Reels de Instagram sin perder calidad, con flujos de trabajo, errores y consejos de diseño.
Build a repeatable workflow for AI voiceover and generated music: script prep, voice selection, mixing levels, loudness targets, and QA checks.