Customer interviews look simple: ask people what they want, write down the answer, then build it. That is also how teams spend six months shipping a feature five unusually polite customers said sounded “interesting.”
AI customer interview prompts can help turn a fuzzy assumption into a research question, draft a neutral guide, critique leading language, prepare follow-ups, organize de-identified notes, and surface candidate themes. They cannot recruit a representative group, build trust, read discomfort reliably, verify what someone meant, or decide what your team should build.
AI can help prepare and organize the interview. Humans still earn consent, conduct the conversation, protect participants, interpret the evidence, and own the decision.
These ten templates are for product managers, UX researchers, founders, marketers, customer-success teams, and service designers who want evidence rather than customer-flavored fan fiction. For broader source work, start with AI research prompts. If a structured questionnaire is the better method, use these AI survey design prompts.
What customer interviews can actually tell you
An interview is good at revealing context: what someone recently tried, what happened, where they struggled, what workaround they invented, and why the situation mattered. It is weaker at estimating prevalence, predicting future purchases, or proving that one product change caused an outcome.
A useful interview needs:
- One learning goal: the uncertainty the team needs to reduce.
- A real decision: what might change after the research.
- Relevant participants: people with recent experience of the behavior.
- Neutral questions: no answer hidden inside the wording.
- Concrete recall: recent events instead of fantasy futures.
- Responsive follow-ups: curiosity, not a robotic script.
- Consent and privacy: clear rules for recording, storage, and use.
- An evidence trail: observations separated from interpretation.
- Visible limitations: who was missing and what remains unknown.
Interviews do not produce “the customer’s opinion.” They produce accounts from specific people, in a specific context, speaking to a specific interviewer. Courtesy bias, memory limits, recruitment bias, incentives, power differences, and interviewer behavior all affect the result.
If you need to observe someone using a product, pair interviews with AI usability testing prompts. Watching a participant fail to find a button is different evidence from asking whether navigation feels easy.
The reusable customer interview prompt formula
Add this instruction to any template below:
“Act as a customer-research preparation assistant. I am investigating [learning goal, decision, audience, recent behavior, existing evidence, recruitment method, constraints, and deadline]. Use only verified information I provide. Produce [artifact] with assumptions, exclusions, open questions, privacy risks, accessibility needs, and required human review. Separate participant statements, observed behavior, interviewer notes, interpretations, and decisions. Do not invent participants, quotes, prevalence, motivations, consensus, or confidence.”
That separation prevents a common mess. “Participant 4 said setup took two hours” is a statement. “Setup is too slow” is an interpretation. “Rebuild onboarding” is a decision. Each step needs more evidence and judgment than the one before it.
Never paste names, email addresses, recordings, identifiable transcripts, health or financial details, employer-confidential information, credentials, unreleased strategy, or legally sensitive comments into an unapproved AI tool. Removing a name does not automatically anonymize a distinctive story. Use approved research systems, explicit consent, access controls, retention limits, synthetic examples, and human privacy review.
What to collect before prompting
“Write customer discovery questions” is an invitation for generic sludge. Give the model a compact research brief instead.
| Input | Why it matters | Human check |
|---|---|---|
| Decision to inform | Stops research theater | Decision owner confirms |
| Learning goal | Defines the uncertainty | Research lead reviews |
| Recent behavior | Anchors questions in reality | Domain owner verifies |
| Participant criteria | Connects people to the question | Researcher approves |
| Existing evidence | Avoids asking settled questions | Evidence owner checks |
| Recruitment channel | Exposes selection bias | Researcher documents |
| Sensitive topics | Triggers safety planning | Privacy/legal reviews |
| Recording plan | Sets consent and storage rules | Participant explicitly agrees |
| Accessibility needs | Makes participation possible | Research ops confirms |
| Analysis approach | Reduces cherry-picking later | Research lead approves |
Keep the brief focused. An interview guide is not a database schema for every fact the company may someday want.
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 customer interview prompts
Replace brackets with sanitized, verified details. These templates create drafts, not validated research.
1. Turn an assumption into a research question
“Using this product assumption, proposed decision, existing evidence, target audience, constraints, and unknowns: [paste], draft three neutral research questions. For each, explain what an interview could reveal, what it cannot establish, which recent behavior matters, suitable participants, and what decision the evidence may inform. Reject questions that ask interviews to estimate market prevalence, prove causality, or predict purchases with certainty.”
“Would customers pay for automation?” is mostly an invitation to be agreeable. A stronger question investigates how people currently complete the task, what it costs them, what they have already tried, and which tradeoffs drove their choices.
The human researcher should select one learning goal. Ten fuzzy goals create a long interview and shallow evidence.
2. Define participant and screening criteria
“Given this research question, relevant behavior, recency window, product access, geography, accessibility needs, exclusions, recruitment channels, and known gaps: [paste], draft participant criteria and a short screener. Separate essential criteria from convenient preferences. Flag professional-participant risk, power dynamics, channel bias, missing groups, and sensitive questions. Do not claim the resulting sample is representative.”
Interviewing active power users because they answer Slack quickly tells you about active power users. It may tell you nothing about people who abandoned setup, churned, could not access the product, or never trusted the invitation.
Screen only for what the study needs. Collecting extra demographics “just in case” creates privacy risk without automatically creating insight.
3. Draft a neutral interview guide
“Using this approved research question, participant criteria, interview length, known evidence, prohibited topics, and decision context: [paste], draft a conversational interview guide. Include an introduction, consent check, easy context questions, recent-behavior questions, neutral probes, closing reflection, and time boxes. Use one idea per question. Avoid pitching, praise-seeking, jargon, hypotheticals presented as evidence, and language that reveals the preferred answer.”
A guide is a map, not courtroom testimony. The interviewer should follow relevant threads and skip questions already answered naturally.
Ask “Tell me about the last time you…” before “Would you ever…?” Recent behavior is imperfect, but usually more useful than imagined behavior in an imagined future with an imaginary budget.
4. Replace leading and loaded questions
“Audit this draft interview guide: [paste]. Create a table with exact question, issue, likely effect, neutral rewrite, and a suggested behavior-based follow-up. Check for leading wording, false assumptions, double-barreled questions, praise-seeking, solution pitching, social-desirability pressure, jargon, absolutes, blame, and questions that make disagreement awkward. Preserve the original wording for an audit trail.”
“How helpful would our smart dashboard be?” has already declared the dashboard smart and helpful. “Walk me through how you currently monitor this work” leaves room for reality.
AI can catch textbook leading language. It may miss internal politics, cultural meaning, power differences, or the fact that the interviewer designed the feature being discussed. Human review remains mandatory.
5. Generate behavior-based follow-ups
“For each core question in this guide: [paste], suggest three optional neutral probes that ask for a recent example, sequence of events, workaround, trigger, consequence, artifact, or tradeoff. Include reminders to tolerate silence and ask for clarification using the participant’s own words. Do not suggest ‘why didn’t you’ blame questions, debate the participant, diagnose motivation, or push toward the product concept.”
Useful probes include “What happened next?”, “Can you show me how you handled that?”, and “When you said ‘confusing,’ what did that mean in that moment?”
Do not machine-gun every follow-up. Listening beats completing the template. A participant’s unexpected detour may contain the actual answer.
6. Prepare consent, recording, and safety checks
“Using this research setting, participant relationship, location, topics, incentive, recording method, storage system, retention period, observers, and intended uses: [paste], draft a pre-interview checklist and plain-language consent script for human legal/privacy review. Cover voluntary participation, recording choice, withdrawal, confidentiality limits, observers, quote use, storage, access, retention, incentive, accessibility, and escalation if distress or sensitive information appears. Mark legal requirements unverified.”
Consent is not a decorative sentence buried before the first question. Participants should understand what happens to their words and be able to decline recording without being tricked.
If someone shares credentials, personal health details, legal allegations, or another person’s private information, do not casually feed it into the synthesis pipeline. Stop, follow the approved protocol, and minimize exposure.
7. Create an evidence-first note-taking template
“Create a note-taking template for this interview guide and research question: [paste]. Include timestamp or question reference, direct quote, observed behavior or artifact, factual context, interviewer observation, interpretation, confidence, contradiction, follow-up needed, and privacy flag. Keep participant statements separate from analyst conclusions. Add a post-interview section for immediate reflections and possible interviewer bias.”
Notes become dangerous when “participant said export took 20 minutes” silently turns into “exports are broken.” Preserve what was said, what was observed, and what the researcher inferred.
If recording is allowed, notes still matter. They capture context, nonverbal events cautiously, and what the interviewer thought was important at the time. Never treat automated transcription as perfect.
8. De-identify notes before AI-assisted synthesis
“Using this approved de-identification policy and a synthetic example of interview notes: [paste], create a redaction checklist and transformation procedure. Cover direct identifiers, contact details, company and product names, precise job titles, locations, dates, rare attributes, linked events, credentials, health or financial details, third-party information, and distinctive quotations. Recommend participant IDs and a separate access-controlled key. Do not claim text is anonymous merely because names were removed.”
A quote can identify someone through context. “The only pediatric surgeon using Product X in a small Montana hospital” does not need a name attached.
The safest dataset is the smallest one that answers the research question. Keep raw recordings and identity keys outside general-purpose model workflows unless the approved system and consent explicitly support that use.
9. Compare themes without flattening disagreement
“Using these de-identified, human-checked interview notes and the approved research question: [paste], create a candidate-theme matrix. For each theme, cite participant IDs and source excerpts, list supporting and contradicting evidence, relevant context, possible alternative explanations, missing groups, and follow-up questions. Distinguish repeated patterns from isolated but high-impact cases. Do not invent quotes, count vague mentions as prevalence, or resolve disagreement by majority vote.”
Three people mentioning a workaround is a pattern in your interviews. It is not automatically “most customers.” One participant describing a severe accessibility blocker may deserve action even if nobody else encountered it.
AI is useful for sorting. Humans must return to the source notes, check whether excerpts were stripped of context, and decide what differences matter.
10. Draft a findings brief with limitations
“Using this verified theme matrix, participant criteria, recruitment method, research question, decision context, and unresolved contradictions: [paste], draft a findings brief. Include executive summary, method, participant boundaries, evidence-backed findings, counterevidence, limitations, open questions, implications, and options for next research or product action. Label observations, interpretations, and recommendations. Every claim must trace to supplied evidence. Do not fabricate consensus, prevalence, certainty, quotes, or ROI.”
A useful brief helps a decision owner see what changed, why, and where uncertainty remains. It does not pretend six interviews are a referendum.
Recommendations should be framed as options with tradeoffs and evidence needs. Product, research, legal, accessibility, security, and business owners still make the call.
A practical interview workflow
Use the prompts in a controlled sequence:
- Define one decision and one learning goal.
- Review existing evidence before recruiting anyone.
- Choose participants connected to recent relevant behavior.
- Draft and human-review the guide.
- Pilot the guide with one or two appropriate people.
- Confirm consent, accessibility, recording, and retention rules.
- Conduct interviews with a human interviewer who listens.
- Secure, minimize, and de-identify notes under policy.
- Use AI only for approved organization and candidate synthesis.
- Verify every theme against source evidence and document limits.
For turning findings into clearer product boundaries, see AI requirements gathering prompts. For the eventual choice among options, use AI decision-making prompts without outsourcing accountability.
Common ways teams fool themselves
They interview whoever is easiest. Friendly customers, internal champions, and loud community members are convenient. Document who is absent.
They pitch instead of research. If half the session explains the concept, the participant is reviewing your presentation, not describing their world.
They ask about hypothetical intent. “Would you use this?” costs nothing to say yes to. Ask about recent behavior, existing spending, workarounds, and tradeoffs.
They count mentions as market size. Interviews reveal mechanisms and context. Surveys, behavioral data, experiments, or other methods may be needed for magnitude.
They summarize too early. A neat theme created after the second interview can become a lens that hides later contradictions.
They paste raw transcripts into a consumer chatbot. Convenience does not cancel consent, contracts, policy, or privacy obligations.
They let generated prose outrank evidence. A polished findings brief can still be wrong. Trace every claim back to a participant statement, observation, or verified artifact.
If your organization needs clearer rules for approved tools and data classes, these AI policy prompts can help draft the policy. Humans with actual authority must review it.
Frequently asked questions
Can ChatGPT write customer interview questions?
Yes, it can draft and critique questions when given a specific learning goal, audience, recent behavior, constraints, and existing evidence. A trained human should review the guide for bias, sensitivity, accessibility, context, and research fit before using it.
Can AI conduct customer interviews by itself?
A bot can collect structured responses, but that is not equivalent to a skilled human interview. It may miss discomfort, rapport, ambiguity, unexpected context, accessibility needs, and ethical boundaries. For meaningful or sensitive research, use an accountable human interviewer and approved protocol.
How many customer interviews are enough?
There is no universal magic number. The answer depends on the decision, audience diversity, recruitment quality, risk, method, and whether new interviews still change understanding. Report the actual participant boundaries and avoid pretending a small qualitative sample estimates prevalence.
Should I record customer interviews?
Only with clear consent and an approved plan for storage, access, use, retention, deletion, and withdrawal. Recording laws and organizational requirements vary. Participants should know who can access the recording and whether quotes may be used.
Can I paste interview transcripts into an AI tool?
Only if participant consent, contracts, organizational policy, data classification, vendor terms, retention controls, and the approved research system allow it. Prefer de-identified notes and data minimization. Do not assume removing names makes a transcript anonymous.
How do I stop AI from inventing customer quotes?
Explicitly forbid invented quotations, require participant IDs and source excerpts for every claim, and verify the output against source notes. If the tool cannot provide a traceable source, treat the statement as unsupported and remove it.
What is the difference between interviews and surveys?
Interviews explore context, behavior, language, and mechanisms in depth. Surveys collect structured self-reports at broader scale when sampling and questionnaire design are defensible. Use the method that matches the question; sometimes the right sequence is interviews first, then a survey designed from what you learned.
Who owns the final product decision?
A human decision owner. AI may organize evidence and outline options, but strategy, ethics, feasibility, customer impact, and accountability remain human work. If that boundary feels fuzzy, read the five-minute guide to what AI can and cannot do.
The point is better listening, not automated certainty
AI can make interview preparation faster and synthesis less chaotic. That is useful. The value disappears when teams use generated questions to lead participants, upload sensitive transcripts without permission, or turn six conversations into a universal truth.
Use the machine for structure. Use humans for trust, curiosity, judgment, privacy, and responsibility. That division of labor is the larger argument in Don’t Replace Me by Dmitry Kargaev: the goal is not to cosplay as a machine. It is to use one without surrendering the parts of work that require an accountable person.
