The task
You ran a customer survey two weeks ago. 200 responses across 12 questions, sitting in a CSV. Your PM wants a one-pager by EOD Friday to feed three decisions: whether to raise prices, whether to kill the free tier, and which integration to build next. The native response of “I’ll just make 30 charts” is what produces decks nobody reads — and worse, every reader walks away with a different conclusion. You want a one-page narrative organized around the 3 decisions, with the actual percentages, sample sizes, and verbatim quotes that make the findings stick — plus an honest section on what this survey is not ready to answer.
Where AI helps — and where it does not
AI is strong at clustering open-end responses into themes, pulling verbatim quotes that exemplify each theme, computing percentages with sample-size context, and reorganizing the survey by decision instead of by question order. It is also good at flagging where the data contradicts your stated prior — those are the findings that change behavior.
What AI cannot do reliably: tell you which finding matters most to the business. The same 84% number can support “lower prices” or “redo the upgrade page” depending on the verbatim under it. Feed the model your 2-3 decisions explicitly so it can organize around them. AI also cannot tell when your sample is biased — if you surveyed only active users, the survey says nothing about churn, no matter how clean the percentages look.
A specific failure mode: AI happily reports percentages without sample sizes when the underlying N is tiny (“75% of respondents said X” on N=8). Tell it: “every percentage needs an N in parentheses, and any finding on N<30 must be flagged as directional only.”
What to feed the AI
- Raw survey responses (CSV upload, or paste responses if smaller)
- The 2-3 business decisions this survey is supposed to inform — specifically
- Your prior expectation for each decision (so the model flags contradictions)
- Who you surveyed and how they were sampled — active users? churned users? mailing list? (changes interpretation entirely)
- The total N and the response rate (low response rates change confidence)
- Demographic or segment splits that matter (B2B vs B2C, free vs paid, US vs EU)
- Questions you regret asking (so the model deprioritizes them)
- Any verbatim quote that already caught your eye — to anchor the model on the qualitative signal
Copy-ready prompt
Interpret these survey results.
Responses (CSV or pasted): {paste}
Who I surveyed + sampling method: {paste}
Total N + response rate: {paste}
The 2-3 business decisions this survey must inform: {paste}
My priors for each decision: {paste}
Segments that matter (B2B/B2C, free/paid, etc.): {paste}
Questions I regret asking: {paste}
Return:
1) One-line headline per decision (max 3 decisions). Lead with what the data says, then the action implication.
2) Quantitative finding per decision: the percentage, the N in parentheses, and the sample-size context. Anything on N<30 is flagged as "directional only."
3) Open-end themes per decision: 3 themes max, each with one verbatim quote in quotation marks. No theme without a quote.
4) Prior-contradiction flags — anywhere the data contradicts my stated prior. These get extra attention because they are where the survey changes behavior.
5) Too-thin-to-conclude — explicitly list the questions and segments where the sample is insufficient. Do not paper over.
6) Final two sections: "Decisions this survey is ready to inform" and "Questions for round 2."
Rules:
- Every percentage must include the N in parentheses.
- Every open-end theme must include a verbatim quote — paraphrases do not count.
- Do not collapse "I don't know" and blank answers into the same bucket — they are different signals.
- If a finding contradicts my prior, mark it with [PRIOR-CONTRADICTION] so I cannot miss it.
- Total length under 500 words. This is a one-pager, not a report.Shorter variant — one-question deep read
Below is one survey question with all open-end responses. Cluster into 3-5 themes. For each theme, give me the count, percent of total (with N), and one verbatim quote. Drop any theme with fewer than 5 mentions or no usable quote.
Responses: {paste}Sample output
A useful headline: “Pricing decision: 84% of churned users cited ‘too expensive’ (N=92 of 110 churned) — but the open-end shows the real issue is unclear value, not price level. Action: redo the upgrade page before considering a price cut. [PRIOR-CONTRADICTION: you expected price level to be the issue.]”
A useful theme with verbatim: “Theme: ‘I don’t know what I’m paying for’ (N=38 of 92 cited-pricing). Representative quote: ‘I don’t know what I get for the $20 that I don’t get for free. I keep meaning to look at the upgrade page but it doesn’t tell me.’ This theme outweighs straight ‘price too high’ (N=22 of 92) by almost 2:1.”
A useful too-thin-to-conclude section: “Integration decision: too thin to conclude. Only 32 respondents answered the integration question, and 18 of those said ‘no preference.’ The remaining 14 are split across 6 integrations with N≤3 each. Round 2 should re-ask integration as a forced-choice ranked question with a list of 5 candidates.”
A useful “ready to inform” closer: “This survey is ready to inform: (1) the pricing decision — re-do upgrade page first, defer price cut. (2) the free-tier decision — keep, with the qualifier that free-to-paid conversion needs investigation. This survey is NOT ready to inform: (1) which integration to build next — sample too thin. Round 2 needs an integration-ranking question with forced choice.”
How to refine
- Force N in every percentage: “Re-read the doc. Every percentage must have the N in parentheses. If any percentage is missing an N, add it. If a percentage is based on N<30, append ’— directional only.’”
- Demand verbatim per theme: “Each open-end theme must include a real verbatim quote in quotation marks. No quote, no theme — replace with a different theme that does have one, or merge it.”
- Flag prior-contradictions prominently: “Re-check the findings against my stated priors. Any finding that contradicts a prior gets a [PRIOR-CONTRADICTION] tag prepended. These are the most important findings; do not bury them.”
- Surface the ‘I don’t know’ answers: “Treat ‘I don’t know,’ ‘not applicable,’ and blank as three separate buckets, not one. Each carries different signal — ‘I don’t know’ on pricing usually means unclear value; blank usually means survey fatigue.”
- End with the readiness split: “Close every survey readout with two sections: ‘Decisions this survey is ready to inform’ and ‘Questions for round 2.’ If everything is in ‘ready,’ you are overclaiming; usually 1 of 3 decisions belongs in round 2.”
Common mistakes
- Citing percentages without sample size — 75% of 8 respondents is not a finding, it is a coincidence; every percent needs N in parentheses
- Themes without verbatim quotes — when AI paraphrases, audiences treat the paraphrase as authoritative; verbatim grounds the theme in real customer language
- Ignoring “I don’t know” and blank answers — they are signal, often the most important one; collapsing them into “other” hides the real story
- Re-organizing the report by survey question order — readers care about decisions, not the order you wrote the survey; reorganize by the 2-3 decisions the survey informs
- Over-extending sub-segment claims — 12 respondents in a sub-segment can support directional language only; treating it as conclusive misleads
- Same survey informing too many decisions — a survey designed for 5 decisions usually answers none well; pick 2-3 going in
- Not feeding the model your priors — without priors, the model cannot tell you what is surprising; surprises are what move decisions
- Hiding the “too thin to conclude” findings — the temptation is to paper over weak samples; doing so kills credibility on the strong findings too
FAQ
- What’s the minimum sample size for a finding?: For overall percentages: 30+ for directional language, 200+ for confident claims. For sub-segments: 30+ per segment, ideally pre-stratified so segment N is known going in. Anything under 30 is “this hints at” not “this shows.”
- Should I include charts in the one-pager?: One chart per decision, if any. Beyond that, charts become page-padding; the narrative with verbatims does more work. If a chart doesn’t change the action implication, drop it.
- What if my survey sample is biased — only active users responded?: Be explicit in the readout: “This survey reflects active users only; nothing in this report should be read as a statement about churned users.” Then plan a churned-user follow-up.
- The model keeps overclaiming from thin samples — what changes?: Add: “Any finding based on N<30 must include the phrase ‘directional only — sample too small for confidence’ in the headline itself. Do not let small-N findings appear without that caveat.” Then re-run.
- How should I share the readout with the team?: Send the one-pager first, in writing. Then do a 20-min live walkthrough. Avoid presenting as 30 slides — that’s the failure mode you started with.