Talking to Users at Scale with AI

Paul Graham's advice is to talk to your users. YC partners say it constantly. Every founder nods. Most don't do it enough.

The bottleneck isn't usually motivation — it's time and synthesis. You have 50 users to talk to. Each call takes an hour. Synthesizing 50 hours of notes takes another 20 hours. If you're also building, something has to give.

AI changes the synthesis bottleneck significantly. It doesn't eliminate the calls — but it makes each call more valuable and turns 50 conversations into actionable insight much faster.

What AI Actually Helps With in User Research#

Let me be specific about where AI adds value and where it doesn't.

AI helps: Synthesizing patterns across many conversations. Generating questions to ask. Transcribing and organizing notes. Finding contradictions in what users say. Surfacing the outlier who mentioned something unusual.

AI doesn't help: The empathy that makes users open up. The intuition that tells you to probe deeper when someone hesitates. The implicit communication — tone, pauses, energy — that often carries more signal than words.

The human part of user research is irreplaceable. The synthesis part is where AI earns its keep.

Setting Up a User Research System in DenchClaw#

This is how we structure user research for DenchClaw:

Object: User Interviews
Fields:
  - Person (relation to contacts)
  - Date
  - Duration
  - Use Case (text)
  - Current Solution (text)
  - Pain Level (1-5)
  - Enthusiasm Level (1-5)
  - Key Quotes (rich text)
  - Would Pay? (yes/no/maybe)
  - Feature Requests (tags)
  - Segment (enum: founder/developer/sales/other)
  - Status (scheduled/completed/follow-up)

Every user interview gets an entry. The Key Quotes field captures the literal words users used — not paraphrases. After 10 interviews, the agent can start surfacing patterns.

The Pre-Interview Preparation Flow#

Before each user interview, query what you already know:

"Show me everything we know about Sarah Chen — her company, 
previous conversations, how she found us, what she said she 
needed when she signed up."

The agent pulls together all your CRM data on this person and generates a prep brief. You walk into the call knowing their context without manual thread-diving.

Also use AI to generate interview questions tailored to their situation:

"Generate 7 discovery questions for a conversation with a 
founder who signed up after our Show HN post, works in 
SaaS, and mentioned they're frustrated with HubSpot pricing."

The questions are better when they're specific to the person's situation, not generic.

During the Interview: What to Track#

The most important thing is capturing exact quotes, not paraphrases. When a user says "I hate that I have to export a CSV every time I want to run a custom report," write that down word for word. Paraphrasing loses signal.

After the call, update DenchClaw immediately:

• Pain level (1-5) and enthusiasm level (1-5) based on your impression
• Key quotes (exact words they used)
• Feature requests they mentioned explicitly
• One sentence describing their primary use case

Five minutes of logging while the conversation is fresh is worth more than 30 minutes trying to reconstruct it the next day.

The Synthesis Loop#

After 10-15 interviews, ask DenchClaw:

• "What are the most common pain points across user interviews?"
• "Which features have been requested by multiple users?"
• "Which user segments show the highest pain level?"
• "Are there any contradictions between what different users said about X?"

The agent queries your interview entries and surfaces patterns. This takes 2 minutes instead of 2 hours.

After 30 interviews, you can ask more sophisticated questions:

• "Compare pain levels between founder-segment users and developer-segment users"
• "Show me users with pain level >= 4 who mentioned competitor X"
• "What do the high-enthusiasm users have in common that low-enthusiasm users don't?"

This kind of cross-tabulation of qualitative data is where DuckDB's SQL capabilities become genuinely useful for user research synthesis.

What AI Can't Do for User Research#

I keep coming back to this because I see founders over-rotate on the AI angle.

The single most valuable thing in user research is the moment when a user says something that surprises you and you lean in and say "interesting, tell me more about that." That's a human judgment — recognizing the unexpected and being curious about it. An AI listening to a transcript would probably skip over it.

The insights that change your product are often in the silences, the hesitations, the "well, kind of — it's more like..." moments. Those require human presence.

Use AI to scale the operational parts of user research. Keep the actual conversations human.

Frequently Asked Questions#

How many user interviews do you need before you can draw conclusions?#

It depends on the signal strength. For very consistent feedback (everyone says the same thing), 10-15 interviews may be enough. For nuanced questions about pricing or positioning, 30-50 is more reliable. The synthesis loop helps you know when you're hearing new things vs. confirming patterns you've already seen.

Should I record my user interviews?#

With permission, yes. Recordings let you review moments you missed and get accurate quotes. Transcripts with tools like Whisper or Otter.ai are fast and good enough. Upload transcripts to DenchClaw entry documents for searchability.

How do I get users to agree to interviews?#

Ask directly and make it easy. "Can I get 20 minutes of your time on a video call this week?" works better than a scheduling link in an email blast. User interview conversion is much higher when the ask is personal and the subject is helping you improve a product they already use.

What if users are too positive in interviews?#

Ask "what almost made you not try this?" and "what would need to be true for you to stop using this?" These questions surface dissatisfaction that users are too polite to volunteer.

Ready to try DenchClaw? Install in one command: npx denchclaw. Full setup guide →