Surveys look easy because everybody has answered one. Write a few questions, add some radio buttons, send a link, and wait for Truth to arrive in a spreadsheet. Then twelve coworkers interpret “often” twelve different ways and the team calls a 4% response rate “the voice of the customer.”
AI survey design prompts can help turn a fuzzy business question into a research objective, draft neutral wording, critique answer options, organize a pilot, and summarize de-identified responses. They cannot make a convenience sample representative, remove bias by rewriting one adjective, infer what respondents secretly meant, or decide what the organization should do.
AI can help edit the questionnaire. Humans still choose the research method, recruit responsibly, protect respondents, validate the analysis, and own the decision.
These ten templates are for researchers, product managers, marketers, customer-experience teams, HR teams, nonprofit staff, and founders who need useful evidence rather than decorative percentages. For broader desk research, use AI research prompts. For quantitative cleanup after collection, start with AI data analysis prompts—inside an approved environment.
What a survey can and cannot tell you
A survey collects self-reported answers from people who were invited, noticed the invitation, chose to respond, understood the questions, and felt able to answer. Every step creates possible distortion.
A useful survey needs:
- One decision: what someone will do differently after seeing the results.
- A defined population: exactly whose experience or opinion matters.
- A defensible sample: how people are recruited and who may be missing.
- A clear construct: what each question is actually trying to measure.
- Neutral wording: no pressure toward the answer the sponsor wants.
- Sensible response options: complete enough to fit real people.
- A short pilot: observation of how respondents interpret the instrument.
- An analysis plan: written before interesting answers tempt the team.
- Transparent limits: what the results do not support.
Surveys are useful for reported attitudes, preferences, recollections, intentions, and experiences. They are weaker at explaining observed behavior, proving causality, or representing a population that was poorly sampled. If you need to see where people actually struggle, pair the survey with usability testing. If you need causal evidence about a product change, consider a properly reviewed A/B test.
The reusable survey design prompt formula
Use this instruction with any template below:
“Act as a survey-design assistant. I am researching [decision, target population, research objective, recruitment method, constraints, and deadline]. Use only the verified information I provide. Produce [artifact] with assumptions, risks, exclusions, open questions, privacy considerations, accessibility checks, and required human review. Separate supplied facts, respondent reports, analysis, interpretations, and decisions. Do not invent respondents, response rates, sample sizes, quotes, benchmarks, statistical significance, or representative claims.”
That separation matters. “Forty of 100 respondents selected option B” is an observation. “Customers prefer B” is an inference requiring a defensible sample, clean question, and appropriate analysis. “Build B” is a decision involving costs, strategy, ethics, and judgment.
Never paste respondent names, email addresses, employee IDs, health or financial details, credentials, raw survey exports, confidential comments, or demographic combinations that could identify someone into an unapproved AI tool. Remove free-text details carefully; deleting a name does not make a story anonymous. Use synthetic examples, approved systems, access controls, documented retention, and human privacy review.
What to collect before prompting
“Make me a customer survey” gives the model room to manufacture your research plan. Provide a compact, verified brief instead.
| Input | Why it matters | Human check |
|---|---|---|
| Decision to inform | Prevents curiosity theater | Accountable owner confirms |
| Research objective | Defines what must be learned | Research lead reviews |
| Target population | Sets the boundary of claims | Domain owner validates |
| Recruitment method | Reveals selection risk | Researcher documents |
| Known evidence | Avoids asking settled questions | Evidence owner verifies |
| Sensitive topics | Triggers consent and safety review | Privacy/legal reviews |
| Delivery channel | Affects length and accessibility | Operations owner tests |
| Analysis plan | Reduces post-hoc storytelling | Analyst approves |
| Timeline and budget | Keeps the design realistic | Project owner confirms |
| Decision threshold | Explains how evidence will be used | Decision owner signs off |
Do not ask for data merely because it might be interesting later. Every additional question costs attention and creates more personal information to protect.
This came from a book.
Don't Replace Me
200+ pages. 24 chapters. The honest version of what AI means for your career, written by someone who actually builds this stuff.
Get the Book →10 AI survey design prompts
Replace brackets with sanitized, verified details. These templates draft artifacts; they do not certify the research.
1. Turn a business question into a research objective
“Using this proposed decision, current evidence, stakeholder questions, target population, constraints, and unknowns: [paste], draft three possible research objectives. For each, state the construct to measure, why a survey is or is not suitable, required population, likely limitations, and what decision the result could inform. Reject objectives that ask a survey to prove causality or predict behavior with certainty.”
“Do customers like our product?” is not an objective. Which customers, which experience, what does “like” mean, and what changes if the answer is no?
A useful objective might be: “Estimate how recently activated administrators rate setup clarity, identify commonly reported obstacles, and decide which three issues deserve follow-up interviews.” It has a population, topic, output, and decision. A researcher still decides whether a survey is the right method.
2. Define the audience and screening criteria
“Given this target population, decision, product eligibility rules, geography, time window, recruitment channels, exclusions, and known coverage gaps: [paste], draft a recruitment and screening brief. Distinguish the target population, sampling frame, invited sample, respondents, and analysis sample. Flag selection bias, nonresponse risk, duplicate responses, professional respondents, and groups likely to be underrepresented. Do not claim representativeness.”
A customer email list excludes people who churned before joining it. An in-product intercept excludes people who cannot access the product. An employee Slack poll excludes silence, fear, leave, time zones, and anyone who muted the channel for emotional survival.
Document who had a realistic chance to respond. If the sample is directional, say so. “Respondents to this survey” is honest. “Our customers” may be fiction.
3. Draft neutral questions
“Using this research objective, constructs, audience language, reading-level goal, sensitive-topic rules, and prohibited assumptions: [paste], draft up to 12 concise survey questions. Use one concept per question. Avoid leading wording, loaded terms, absolutes, hypotheticals presented as facts, unnecessary jargon, and assumed experience. For each question, explain its purpose and list one possible misinterpretation for human review.”
Questions should not contain the sponsor’s preferred answer. “How helpful was our improved dashboard?” smuggles in two claims: that it improved and that it helped. Ask about a specific task or experience instead.
AI can spot textbook problems. It may miss political pressure, cultural meaning, product history, or why “simple” irritates people who struggled. Have humans from the target audience review the wording.
4. Choose consistent response scales
“Review these draft questions and intended analyses: [paste]. Recommend a response format for each using only justified categories, such as frequency, agreement, satisfaction, ease, confidence, number, date, or open text. Provide balanced labels for every scale point, a suitable recall period, and options for not applicable, do not know, or prefer not to answer where needed. Flag mismatches between the question and scale.”
Do not mix “strongly disagree” with a question asking frequency. Avoid unlabeled mystery numbers unless respondents understand the anchors. Keep scale direction consistent so people do not accidentally answer the opposite.
More scale points do not automatically create more truth. The researcher should choose a scale that respondents can distinguish and that supports the planned analysis.
5. Find leading and double-barreled questions
“Audit this questionnaire: [paste]. Create a table with question number, exact wording, issue, severity, likely effect, and a neutral rewrite. Check for leading language, double-barreled questions, hidden assumptions, absolutes, social-desirability pressure, emotionally loaded terms, unclear reference periods, jargon, negation, and questions that require knowledge respondents may not have. Do not silently rewrite; preserve the audit trail.”
“How satisfied are you with the speed and accuracy of support?” asks two questions. A respondent may love the speed and hate the answer. One radio button cannot express both.
An audit is a useful second pass, not validation. Pilot the revised instrument with real members of the audience and ask them to explain what each question means in their own words.
6. Improve answer options
“For each closed-ended question and current answer list: [paste], check whether options are mutually understandable, sufficiently complete, consistently ordered, non-overlapping where appropriate, and aligned with the question. Identify missing ordinary answers, false binaries, unequal ranges, ambiguous ‘other,’ and cases needing ‘none,’ ‘not applicable,’ ‘do not know,’ or ‘prefer not to answer.’ Propose revisions and explain tradeoffs.”
Age ranges that overlap at 35 are broken. Job-role lists that omit contractors distort the answer. “Other” can be useful, but a required free-text explanation can turn it into a privacy trap.
Randomizing options can reduce order effects in some cases; it is wrong for ordered scales and may confuse respondents. Humans should approve the logic in the actual survey platform.
7. Create an accessible, mobile-friendly version
“Review this survey for a respondent using a phone, keyboard, screen reader, magnification, limited bandwidth, or plain-language translation: [paste]. Produce an accessibility checklist covering question length, heading structure, labels, focus order, error messages, contrast, matrix questions, required fields, progress indicators, time estimate, save-and-return behavior, and alternatives to visual information. Mark every item unverified until tested.”
A ten-column matrix may look efficient on a laptop and become thumb-operated punishment on a phone. Break complex grids into shorter questions when possible. Do not use color alone to carry meaning.
Automated suggestions cannot replace testing with assistive technology and people with relevant access needs. Accessibility belongs in the design, not in an apology after launch.
8. Plan a small pilot
“Using this draft survey, audience, delivery channel, risks, and deadline: [paste], create a pilot plan. Include recruitment criteria, a small purposeful range of participants, think-aloud or debrief questions, completion-time observation, comprehension checks, technical tests, privacy checks, stopping conditions, issue logging, owners, and revision rules. Do not invent a universally valid pilot size.”
A pilot is not a miniature launch whose only metric is completion. Watch people interpret the questions. Ask what they thought each item meant and whether any answer option failed to fit.
Fix severe confusion, broken logic, inaccessible controls, or unexpected disclosure risk before broader distribution. Record revisions so the team knows which version produced which data.
9. Build a de-identified analysis framework
“Using this approved research objective, questionnaire, sampling notes, response coding, missing-data rules, and decision needs: [paste], draft an analysis plan before viewing outcomes. Map every question to an objective; define valid denominators, exclusions, weighting decisions requiring expert review, planned subgroup comparisons, open-text coding steps, uncertainty reporting, and limits. Ban post-hoc segment fishing and invented significance.”
Write the plan early because dashboards encourage storytelling. If twenty subgroups are searched after collection, one surprising pattern will eventually appear by luck.
Free-text responses need special care. Remove direct identifiers, review rare details, restrict access, and report themes without exposing respondents. AI-assisted coding requires human calibration, disagreement review, and documentation of the model and prompt used.
10. Summarize findings without overclaiming
“Using only these approved aggregate tables, documented sample, field dates, recruitment method, analysis notes, limitations, and de-identified themes: [paste], draft a findings brief. Separate observed results, plausible interpretations, unanswered questions, and recommendations. State who responded, who may be missing, uncertainty, contradictory evidence, and what follow-up research is needed. Never fabricate quotes, causal claims, population claims, or confidence.”
A readable report does not need fake certainty. Say “62% of 214 respondents recruited through the customer newsletter selected…” rather than “62% of customers believe…” when the sample does not justify the larger claim.
Recommendations need accountable owners. Use a documented decision-making process and preserve the result in a decision log. The model can draft the memo. It does not absorb the consequences.
A human review gate before launch
Run a short review with research, privacy, accessibility, analytics, domain, and decision owners as appropriate.
- Does every question support the stated objective?
- Is a survey the right method for each claim?
- Can the target population realistically be reached?
- Are consent and sensitive-topic protections adequate?
- Could wording pressure, shame, confuse, or identify respondents?
- Do answer options fit ordinary real-world situations?
- Does skip logic work on mobile and assistive technology?
- Is the analysis plan written before results arrive?
- Are claims limited to what the sample supports?
- Is someone accountable for acting—or explicitly not acting—on the result?
Use AI policy prompts if the organization lacks rules for approved tools, retention, or sensitive research data. For general safe operating habits, read the no-BS guide to using AI at work.
Frequently asked questions
Can ChatGPT write a whole survey for me?
It can draft one, which is different from designing valid research. Give it a verified objective, audience, constraints, and review criteria. Then have a qualified human review the method, wording, privacy, accessibility, sampling, and analysis. Never launch raw model output because it looks polished.
How many questions should a survey have?
As few as needed to inform the decision. There is no universal number. Length depends on audience, topic, channel, burden, sensitivity, and value to the respondent. Pilot completion time and remove questions that are merely “nice to know.”
Can AI calculate the required sample size?
It can explain inputs or reproduce a calculation under supervision, but it should not choose assumptions by vibes. Sample planning depends on population, design, expected effect or precision, variability, clustering, subgroup needs, nonresponse, and analysis method. Use a qualified researcher or statistician.
How do I stop AI from writing leading questions?
Explicitly prohibit leading language and require an audit explaining each risk. Ask for alternative neutral wording. Then pilot with real people. Models can reproduce bias from the brief, especially when stakeholders describe their preferred conclusion as a fact.
Is survey data safe to paste into an AI tool?
Not by default. Raw exports can contain direct identifiers, sensitive free text, timestamps, rare demographics, and combinations that re-identify people. Use approved tools and data-handling rules. Prefer aggregated or synthetic data, minimize fields, restrict access, and involve privacy or security owners.
Can AI analyze open-ended survey responses?
It can assist with a documented coding process in an approved environment. Humans should build and test the codebook, review disagreements, inspect minority themes, check whether summaries erase criticism, and retain traceability to de-identified evidence. Do not present generated themes as objective truth.
Does a large response count make a survey representative?
No. A huge biased sample is still biased. Representation depends on who could be selected, who was invited, who responded, coverage gaps, nonresponse, weighting, and the claim being made. Report the recruitment method and limits alongside the count.
When should I use interviews instead of a survey?
Use interviews or observation when you need depth, context, language, motivations, or understanding of an unfamiliar problem. Use a survey when the constructs and answer space are sufficiently understood and structured responses can inform a defined decision. Often the best sequence is interviews first, survey second.
The point is better evidence, not faster forms
AI is genuinely useful at the boring parts: generating alternatives, checking consistency, building review tables, and turning notes into an organized draft. That speed matters. It also makes it easier to launch a bad survey before anyone asks whether the sample, question, or decision makes sense.
The durable advantage is human judgment: choosing a worthwhile question, respecting respondents, noticing context, and resisting the convenient conclusion. If you need a broader field guide for keeping that advantage while using the tools, Don’t Replace Me by Dmitry Kargaev is the low-drama version.
Use the machine to sharpen the instrument. Do not let it manufacture the evidence.
