The AI Customer Research Loop

--- ## Who this is for You are a founder 0-6 months from launch (or stuck pre-PMF) who has done 5-10 customer interviews and feels you aren't learning enough. You want a system, not theory. You are willing to use AI to augment the work but not to skip the actual conversations. If you are running a mature product with 1,000+ customers, this book is too early-stage. If you sell B2C at <$50 ACV, the interview cost does not justify the methodology. ## What you'll have after reading - [ ] A recruitment system that books 30 interviews in 48 hours - [ ] A 25-question interview script that doesn't lead the witness - [ ] The transcription + cleaning pipeline (Granola + Whisper) - [ ] The single-interview synthesis prompt (Claude) - [ ] The cross-interview pattern recognition prompt - [ ] The opportunity-solution tree template - [ ] A 7-day research sprint schedule - [ ] The decision rule: build / iterate / kill --- # Chapter 1 — Why most customer interviews fail You've probably done customer interviews that felt productive but didn't move the needle on your product. The reason is not the conversation itself. It is what happens before, during, and after. ## The 5 failure modes 1. **Recruiting the wrong people.** You interview people who match your demographic profile but not your pain profile. They are polite. They are not your customer. 2. **Leading the witness.** "Would you use a tool that..." is a leading question. The respondent says yes to be helpful. You build the wrong thing. 3. **Surface-level synthesis.** You take notes. You remember the highlights. You forget the 3 sentences that actually mattered. 4. **No cross-interview pattern recognition.** You hear each interview in isolation. You never see the pattern across 20 conversations. 5. **No action artifact.** You finish the interview sprint and don't have a roadmap. Just a Google Doc full of quotes. AI fixes #3 and #4 well. It improves #1 and #5 modestly. It does not fix #2 — that is on you. ## The Mom Test, applied Rob Fitzpatrick's *The Mom Test* gives the 3 rules that apply in 2026 just as they did in 2013: 1. **Talk about their life, not your idea.** Ask about their current behavior. Not about your product. 2. **Ask about specifics in the past, not generics or hypotheticals.** "Last time you had [problem], what did you do?" not "Would you use a tool that...?" 3. **Talk less.** The customer should be talking 80% of the time. Your job is to listen. AI helps you listen better (it summarizes the transcript), but it cannot make you ask better questions. ## What AI does well in research 1. **Transcription.** Granola, Otter, Fireflies — all hit 95%+ accuracy. Cost: $0-$20/mo. 2. **Single-interview synthesis.** Claude or ChatGPT can extract the key insights, pain points, and quotes from a 30-minute transcript in 30 seconds. Cost: $0.05-$0.20 per interview. 3. **Cross-interview pattern recognition.** Once you have 10-20 transcripts, AI can find the patterns across all of them — the recurring pain points, the unmet needs, the language customers use. 4. **Opportunity-solution tree construction.** Teresa Torres's framework (opportunities → solutions → experiments) can be drafted by AI in minutes after the synthesis. ## What AI does poorly in research 1. **Asking the follow-up question.** AI cannot read body language, hear the hesitation, ask the unexpected question. 2. **Knowing when to push.** "Tell me more about that" requires judgment. 3. **Synthesizing qualitative into actionable.** AI can extract themes. It cannot tell you which theme matters most for your business. The rule: AI accelerates the mechanical parts of research. The judgment parts are still you. ## The 2026 research tooling landscape | Tool | Use case | Cost | |---|---|---| | Granola | AI meeting notes, transcript → summary | Free trial / $18/mo | | Listen Labs | AI-moderated interviews at scale (B2C, large samples) | Enterprise pricing | | UserResearch.ai | B2B participant recruitment | $99/mo | | Atypica.AI | Synthetic persona interviews (consumer) | Contact sales | | Speak | AI transcription + analysis | $50-$500/mo | | Claude / ChatGPT | Single + cross-interview synthesis | $20/mo each | | Notion | Synthesis storage + opportunity-solution tree | Free | For most founders, the right stack is: UserResearch.ai (recruit) + Granola (transcribe) + Claude (synthesize) + Notion (store). $120/mo total. ## Do this in the next 48 hours 1. Open `templates/05-the-ai-customer-research-loop-templates/01-icp-profile.md`. Define your ICP for research. 2. Set up Granola or Otter for transcription. 3. Set up a Notion database for storing interview summaries. 4. Schedule 3 interviews for next week. --- # Chapter 2 — Recruit 30 interviews in 48 hours The hard part of customer research is not the interview. It is getting 30 people on the calendar in 48 hours. Most founders spend 3 weeks recruiting, lose momentum, and end up with 5 interviews. ## The 5 channels | Channel | Time to fill | Quality | |---|---|---| | Your personal network | 1-2 days | Highest (warm) | | LinkedIn outreach | 2-3 days | High (targeted) | | Twitter / X outreach | 1-2 days | Medium (cold) | | UserResearch.ai | 2-4 days | High (paid panel) | | Reddit / community posts | 3-7 days | Medium | Run all 5 in parallel. Goal: 30 interviews in 48 hours. ## The recruitment message The message that gets replies: ``` Subject: 15-min chat about [problem area]? Hi [name] — I'm researching how [ICP] handle [specific problem] today. No pitch, no product — just trying to learn. Worth 15 minutes? I'll send a [day] or [day] option that works. [Your name] ``` The key: "No pitch, no product." Prospects agree because they are curious, not because they want to buy. ## The incentive For B2B research, the standard incentive is: - **No incentive** — works for warm network, sometimes for cold. - **$25-$50 Amazon gift card** — works for cold. - **Free access to a resource** — works if the resource is relevant. - **Donation to a charity** — works for some audiences. Avoid: equity, large cash incentives, anything that biases the respondent. ## The screener A screener ensures you talk to the right people. Use a 5-question form: ``` 1. What is your role? [Dropdown: CEO, Founder, VP Sales, ...] 2. How long have you been in this role? [<1y, 1-3y, 3-5y, 5y+] 3. How big is your company? [1, 2-10, 11-50, 51-200, 200+] 4. How often do you deal with [problem]? [Daily, Weekly, Monthly, Rarely] 5. Are you willing to do a 15-min video call? [Yes / No] ``` Auto-reject anyone who answers "rarely" to #4 (not acute enough) or "No" to #5. ## Do this in the next 48 hours 1. Send 50 recruitment messages across 3 channels. 2. Set up the screener form. 3. Book interviews on Calendly with 2-3 slots/day for the next 14 days. 4. Goal: 30 interviews booked by Friday. --- # Chapter 3 — The interview script that doesn't lead The script is a guide, not a teleprompter. The best interviewers deviate constantly to follow unexpected threads. But starting from a structure prevents the "so... tell me about yourself" trap. ## The 25-question script Use this as the starting point. Customize for your ICP. **Section A: Context (5 minutes)** 1. Walk me through your role. What do you own? 2. What does your typical day look like? 3. What's the biggest challenge in your role right now? **Section B: The pain (10 minutes)** 4. How do you currently handle [problem area] today? 5. Walk me through the last time you had this problem. What happened? 6. What did you try first? Why? 7. What worked? What didn't? 8. How much time/money does this cost you? 9. How do you feel when this happens? (Emotion question) 10. If a magic wand fixed this overnight, what would change? **Section C: Current solutions (5 minutes)** 11. What tools do you currently use for [problem area]? 12. What do you like about them? 13. What do you hate about them? 14. Have you looked for alternatives? What stopped you? 15. What would the perfect tool look like? **Section D: Willingness to pay (3 minutes)** 16. If a tool solved this, what would you pay monthly? 17. What would you expect in return? 18. Who else would need to approve the purchase? **Section E: Closing (2 minutes)** 19. Is there anything I didn't ask that I should have? 20. Can I follow up if I have more questions? 21. Who else do you know who has this problem? 22. May I contact them? ## The "no leading" rules 1. **Never ask "would you use..."** Always ask "what do you currently do..." or "what did you try..." 2. **Never describe your solution.** The interview is about their world, not yours. 3. **Never pitch.** If they ask what you're building, say "I'll share after the research phase." 4. **Use silence.** After<p style="opacity:0.6;font-style:italic;margin-top:24px;">[Sample chapter — buy the book for the full 30-50 page edition]</p>