--- title: "How to Use AI for Customer Interviews" description: "Use AI for customer interview prep, live note capture and follow-up drafting, and keep the moderating human, with the checks that catch a smoothed quote." canonical_url: "https://builderscamp.com/guides/tools/ai-for-customer-interviews" date_published: "2026-09-18" date_modified: "2026-09-18" author: "Andre Albuquerque, Mihaela Draghici" publisher: "Builders Camp" guide_class: "tools" --- # How to use AI for customer interviews without handing it the room **TL;DR:** AI belongs in three slots around a customer interview and in none of them during it: drafting and de-biasing the guide beforehand, transcribing afterwards, and turning what you heard into the next session's probes. Nielsen Norman Group names four specific traps that turn a question leading, and an AI-drafted guide falls into all four by default unless you ask it not to. The interview itself stays yours because the value is in the follow-up nobody scripted. ## What should AI do before the interview starts? Three jobs, in order: draft the screener, draft the guide, then attack the guide. The first two are speed. The third is the one that changes the quality of what you hear, and most teams skip it. An AI-drafted interview guide reads well and leads badly. Ask for six questions about a new checkout flow and you will get questions like "how much time would this save you during your workday", which Nielsen Norman Group lists as a textbook leading question because it hands the participant the answer before they have formed one. The same article names three more traps that appear in almost every first draft: rephrasing what the participant said into your own words, naming an interface element before they name it, and assuming they felt a certain way about a task. A guide written in one pass will contain several of each. So write the second prompt before you write the first. Paste the draft back, name the four traps explicitly, and ask for each question to be rewritten as a probe about a specific past instance. "Tell me about the last time you abandoned a basket" survives that rewrite. "How useful would one-click checkout be for you" does not, and should not. ## Why does the guide need past instances rather than opinions? Because what people say and what they do are different, and the gap is not dishonesty. Memory is unreliable and people construct rationalisations after the fact, so a question about a typical week returns an idealised week. The critical incident method is the standard workaround: ask for the specific extreme case, the time it went badly or unusually well, because those stay sharp in memory when the average day has already blurred. That has a direct consequence for how you prompt. A model asked for "customer interview questions about onboarding" produces attitude questions, since that is what the phrase pulls in its training data. A model asked for "questions that make the participant narrate one specific recent occasion, with no hypotheticals and no future-tense verbs" produces something you can actually code later. The second prompt is longer and the difference is the entire interview. ## Can AI take the notes while you run the session? Transcription, yes. Live interpretation, no. The split matters because they feel like the same feature. Recording and transcribing removes the real cost of note-taking, which was never the typing, it was the attention. A moderator writing notes misses the pause before an answer, the hedge at the end of a sentence, and the moment the participant's story contradicts what they said earlier. Otter's plans put 300 transcription minutes a month on the free tier and 1,200 on Pro, which covers two 45-minute interviews a week without anyone reaching for a budget line. Whatever tool you use, get consent on the recording and strip names, emails and account identifiers before the file goes anywhere else. A live AI panel suggesting follow-up questions is the opposite trade. It puts a second demand on the attention the participant needs from you, and it optimises for the question that fits the script rather than the one that chases the contradiction you just heard. The follow-up is where interviews earn their cost, and it is the one move the model has the least context to make. ## What does AI do well in the ten minutes after the call? This is the slot people underuse. While the transcript is processing, ask the model for three things: - Every claim the participant asserted without giving a specific example, quoted verbatim, so those become planned probes next time. - Every moment where two answers pull in different directions, with both quotes side by side. - The three questions from your guide that produced the shortest answers, which usually means the question was badly framed rather than the topic being empty. Run that immediately, and the next interview is measurably better than the last one. Run it a week later and you have a summary instead of a correction. ## How do you stop the transcript becoming a story? Require a verbatim quote with a speaker code next to every claim in any output, then spot check them against the source file. Anthropic's own guidance for reducing hallucinations is to extract word-for-word quotes before the analysis and to make the model retract any claim it cannot support with one afterwards. The failure mode you are checking for is not fabrication, which is rare and obvious. It is the smoothed quote: the participant said "I think I'd probably use it, maybe, if it was free", and the summary reads "would use it if free". The hedge was the finding and it is gone. Two minutes of searching the transcript for three quotes at random catches that. If one of the three comes back paraphrased rather than exact, treat the whole document as unverified and re-check every quote in it, because the same smoothing ran across all of them. The second thing worth counting is how many distinct participants sit behind each claim. Summaries default to the passive voice of consensus, so a sentence like "users found the upload step confusing" reads identically whether eight people said it or one person said it twice. Force the participant codes into the output and that distinction survives to the readout, where it changes what anyone does about it. ## Where does the AI-moderated interview genuinely fit? Not as a replacement for the ten conversations that shape your roadmap, but as a widener at the edges. If you have run eight interviews yourself and want to check whether a specific behaviour shows up in a different segment or a different market, a scripted, automated session asking three closed behavioural questions is defensible, as long as you report it as what it is and never blend those transcripts into the same theme counts as your moderated ones. The line to hold is about who owns the deviation. In a moderated session, you decide in real time to abandon the guide because something more interesting appeared. That decision is the product of the interview, and no automation currently makes it. ## Where this sits in the rest of the discovery job Interviews feed a wider system. Builders Camp's AI Prompting for Customer Discovery runs 1 week and covers the loop end to end, from turning an assumption into a hypothesis with a stated falsification bar through to the decision it supports. Voice of the Customer extends the same discipline past interviews into support tickets, sales calls and reviews, where volume is higher and context is thinner. The Discovery Expert Track bundles the sequence if you want the path rather than a single week. For the moderating craft itself, [how to run customer interviews](https://builderscamp.com/guides/other/how-to-run-customer-interviews) covers the part no tool touches, and the [discovery interview script](https://builderscamp.com/guides/templates/discovery-interview-script) gives you a starting structure to make AI rewrite rather than invent. For what happens to the transcripts afterwards, see [AI for product discovery](https://builderscamp.com/guides/tools/ai-for-product-discovery) and [synthesising user research with Claude Code](https://builderscamp.com/guides/tools/claude-code-user-research-synthesis). ## The question to ask yourself before the next session Look at your guide and count how many questions could be answered without the participant remembering anything specific. If it is more than one, the session will produce opinions you cannot act on, and no amount of AI synthesis afterwards will turn them into evidence. Fix that before the call, because it is the only part of this workflow that cannot be repaired later. [See the AI Prompting for Customer Discovery bootcamp](https://builderscamp.com/bootcamps/ai-prompting-for-customer-discovery?utm_source=guide&utm_medium=organic&utm_campaign=ai-for-customer-interviews) ## Frequently asked questions ### Can AI moderate a customer interview on its own? It can run a scripted session, and several research platforms now sell exactly that. What it cannot do is notice that the participant just contradicted something they said four minutes ago and chase it. Nielsen Norman Group's guidance is that interviews are worth running for the attitudes and impressions that stuck in people's minds, and those surface through unplanned follow-ups, not through the script. ### What is the safest AI task in an interview workflow? Transcription, followed by a first pass at the guide. Transcription is a mechanical conversion with an obvious correctness check: you were in the room. Otter's free tier allows 300 transcription minutes per user per month and the Pro tier 1,200, which covers a normal discovery cadence of two interviews a week without a paid seat for everyone on the team. ### How do I stop an AI-drafted guide from being full of leading questions? Run every question through the four traps Nielsen Norman Group names: rephrasing the participant's words in your own, suggesting the answer, naming interface elements before the participant does, and assuming how they felt. Ask the model to rewrite each question to a behavioural probe about a specific past instance instead of an opinion about the future. ### Should I record and transcribe every interview? Record with consent, transcribe everything, and strip identifiers before any transcript leaves your systems. Replace each participant with a stable code such as P4 in the file name and in every speaker turn, and keep the mapping from code to person in a file the AI tool never sees. ### Can AI write my follow-up questions during the call? A live suggestion feed competes for the attention the participant needs from you. Use AI between sessions instead: after each interview, have it list the moments where an answer was asserted without a specific example, and take those into the next conversation as planned probes. ### How many interviews should I run before synthesising with AI? Torres recommends three to four story-based customer interviews before building a first opportunity solution tree, which is also a sensible floor before asking a model for themes. Below that, a model will confidently cluster three anecdotes into a pattern, and there is no pattern yet. ### Does Builders Camp teach this? AI Prompting for Customer Discovery is the direct match: a 1 week bootcamp covering interview planning prompts, recruiting and outreach, thematic synthesis and insight-to-decision. Its practical challenge requires you to generate a six-question guide with AI, then edit it to cut leading questions and add behavioural probes before you run it. ## Sources - [Nielsen Norman Group: Interviewing Users](https://www.nngroup.com/articles/interviewing-users/) - [Nielsen Norman Group: 4 Traps to Avoid When Asking Users Questions](https://www.nngroup.com/articles/leading-questions/) - [Product Talk: Why I Prefer Opportunity Solution Trees](https://www.producttalk.org/2016/08/opportunity-solution-tree/) - [Anthropic: Reduce hallucinations](https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) - [Otter.ai: Pricing](https://otter.ai/pricing) ## How this guide was made Researched from Builders Camp's bootcamp, track and masterclass material and the sources listed on this page, drafted with AI, and fact-checked against every source cited.