From analysis to insight narratives

0:00 Insight narratives

Here's a sentence from a real AI-drafted report: customer satisfaction drove a nine percent increase in repeat orders. Both numbers were real. The word drove was not. Nothing in the analysis showed that one caused the other. In this lecture you'll learn to turn analysis into narratives that drive decisions: the Situation, Insight, Implication, Action structure, prompts with guardrails, the so what test, versions for different audiences, and a checker prompt that catches invented numbers and unsupported causal claims.

0:34 Why it matters

Why does this matter? Because decision-makers rarely read your full analysis. They read the summary. If the summary overclaims, the decision follows the overclaim. AI is an excellent drafting partner here, but it tends toward two failures: bland summaries that list numbers, or confident stories that add causation and even figures that aren't in your tables. Your job is to keep the narrative specific, honest and actionable.

1:03 S-I-I-A

Here's the structure. Situation: the context and the question, for example, we asked whether the SUMMER10 discount is worth continuing. Insight: the key finding with supporting numbers, like discounted orders had lower order values and similar repeat rates. Implication: what it means, like the discount seems to shift timing and margin rather than create loyal customers. Action: the recommendation and next step, like end it in July and run a small holdout if we want stronger evidence. Then one clear line on confidence and the main caveat.

1:41 Guardrails

Now prompting with guardrails. Give the AI the tables and the decision, and ask for a two-hundred-word summary for, say, the marketing director, using Situation, Insight, Implication, Action. Then the rules. Use only numbers from the tables; don't introduce new figures. State the confidence level and the main caveat in one sentence. Avoid causal language unless the tables come from a controlled test. Plain English, no jargon. Without those rules, drafts routinely add plausible numbers, overclaim causation, or bury the recommendation.

2:16 The 'so what' test

Next, the so what test. For every sentence, ask so what? If the answer isn't clear, cut it or connect it to the decision. Traffic from Instagram rose twelve percent, is an observation. Instagram traffic rose twelve percent, but it converts at half the rate of search, so it added little revenue, and we shouldn't shift budget there yet, is a sentence that earns its place. Think of each sentence as paying rent in the reader's attention. If it doesn't pay, it goes.

2:52 Audiences and honesty

Tailor to the audience, but keep facts identical. Executives: one paragraph, decision first. Team leads: key numbers by segment and next steps. Analysts: methods, definitions, limitations and links to code. Ask the AI for each version, then check it didn't change any number while rewriting. And storytelling without spin: include the strongest counter-evidence you found, and explain why your recommendation still holds, or what would change it. Readers trust analysts who show their doubts.

3:24 Example 1: the quarterly note

First example, a simple one. An operations analyst in Dubai asks the AI to draft a quarterly note from four tables. The first draft claims customer satisfaction drove a nine percent increase in repeat orders. The analyst checks: the tables show both rose, but nothing links them. The revised line: repeat orders rose nine percent and satisfaction scores improved; we haven't established whether one caused the other. The recommendation stays clear, and the director isn't misled into over-investing in a single lever.

4:00 Example 2: UAE hotel pricing

Second example, a business case. An analyst at a UAE hotel group drafts a note recommending a weekend rate increase. The checker prompt flags two numbers that don't appear in the tables, because the AI had rounded occupancy into a different figure, and one causal phrase: the event drove occupancy, based on observational data. The corrected note uses the exact figures, says occupancy was higher during the event weekends, and recommends a two-weekend price test at two properties, owned by revenue management, reviewed in six weeks.

4:37 Watch me do it, part 1

Watch me run the checker. I paste my draft and the source tables, and ask four things. One, list every number in the draft and the table cell it comes from, marking any number not in the tables. Two, list every causal phrase, like drove, caused, because of or led to, and say whether the data is from a controlled test; if not, suggest non-causal wording. Three, is the recommendation clear, owned and dated? Four, is the main caveat stated near the top? Only then, return a corrected draft.

5:16 Watch me do it, part 2

The response maps eleven numbers to cells and flags one it can't find. It was a percentage I'd calculated in my head. I recompute it properly and it changes slightly. It lists two causal phrases; one is fine because it refers to our controlled email test, the other I rewrite. And it notes my recommendation has no owner. I add one: the head of e-commerce, by the fifteenth. The final note is shorter, more precise, and much harder to misread.

5:51 Length and format

On length and format: most decision narratives should fit on one screen. Lead with the recommendation for senior readers, put supporting numbers in a short table, and link to the full analysis. Ask the AI for a version that works as a chat message as well as a longer document, because many decisions are actually made from the short version in a chat thread.

6:19 Confidence language

Confidence language deserves one more minute, because it's where narratives quietly go wrong. Agree on a small set of phrases and what they mean: almost certain, likely, uncertain, unlikely. Use ranges where the data supports them. And say whether the uncertainty changes the decision: even at the low end of the range, the campaign pays back, is far more useful than a paragraph of hedges. We'll go deeper on this in module six, but build the habit in every summary you write.

6:55 Measuring narratives

How do you know your narratives work? Decisions get made from the summary without a meeting to explain it. Stakeholders quote your numbers correctly later. Recommendations have owners who actually act by the date. And nobody feels misled when they read the full analysis afterward. That last one is the real test of an honest narrative.

7:19 Common mistakes

Common mistakes. Numbers not traceable to a table. Causal language without causal evidence. Caveats hidden in footnotes, or so many caveats that the important one is lost. Recommendations without owners or dates. Changed numbers between audience versions. And words like proves or guarantees.

7:38 Recap

Recap. Structure narratives as Situation, Insight, Implication, Action, plus one confidence line. Prompt with guardrails: only provided numbers, no unsupported causation, plain English. Apply the so what test, tailor by audience while keeping facts identical, and include counter-evidence. And run a checker pass before sending.

7:58 Try this now

Try this now. Write a two-hundred-word Situation, Insight, Implication, Action summary of a recent analysis with AI help, using the guardrail prompt. Then run the checker prompt on it. Trace every number to its source, fix any causal phrase you can't support, and add one line of counter-evidence and an owner with a date.