Primary Research
Overview
Primary research is the UX phase with the highest data sensitivity and the highest AI leverage. AI compresses the 70% of research time spent on logistics, transcription, and initial analysis, freeing practitioners to focus on the 30% that requires human interpretation: why users behave the way they do, what they cannot articulate, and what the data does not say. Critically, never run client user data through AI tools without explicit approval. Most platforms use your data for model training by default.
AI workflow
8 steps
- 01
Research Objectives Definition
Claude
Before writing a single question, clarify objectives. Prompt: "Here are my project goals. What are the 5 most important questions I need answered to make informed design decisions? What would make this research fail to be actionable?" This prevents unfocused guides.
- 02
Interview Guide Generation
Claude
Feed objectives to Claude and generate a full interview guide. Then run the stress-test: "What are the 5 most common blind spots in interview guides like this? What leading questions did I accidentally include? What assumptions is this guide making about users?" Revise before the first session.
- 03
Screener Survey Creation
Tally
Build participant screeners in Tally - unlimited free forms, no watermark. Use Claude to write screening logic that identifies your target user profile without telegraphing who you want to recruit.
- 04
Session Transcription
Otter.ai
Real-time transcription: 300 minutes/month free. Enable auto-delete for sensitive sessions. Accuracy drops with strong accents and cross-talk - always review the first 5 minutes manually to catch systematic errors.
- 05
Async Research Sessions
Loom
25 free videos per month. Replace synchronous interviews for fixed-bid projects. Send participants a structured video prompt, receive recorded responses. Saves 2-3 hours of scheduling per participant.
- 06
Pattern Identification
Dovetail
AI-powered tagging across transcripts. Upload multiple transcripts and use AI to surface recurring themes, sentiment patterns, and high-frequency quotes. Critical: Dovetail is NOT client-safe by default. Get explicit approval before loading real user data.
- 07
Insight Synthesis
Claude + Notion AI
Paste anonymised themes into Claude: "Here are 12 research themes from user interviews. Synthesize these into 5 actionable insight statements. For each, identify what it means for the design, what we are uncertain about, and what follow-up questions it raises." Use Notion AI to reformat for stakeholder delivery.
- 08
Research Readout Creation
Notion AI
Generate the stakeholder readout structure from synthesized insights. Include: key findings, design implications, confidence levels, and recommended next steps. Mark all AI-synthesized content for human review before sharing.
Tools
| Tool | Use | Free tier | Client-safe |
|---|---|---|---|
| Claude | Research objectives, interview guides, insight synthesis | Free tier available | Yes |
| Otter.ai | Real-time transcription | 300 min/month free | Yes |
| Dovetail | AI tagging, pattern identification across transcripts | Free tier available | Review |
| Notably | Research repository, collaborative synthesis | Free tier available | Review |
| Tally | Screener surveys, unlimited free | Unlimited free | Yes |
| Loom | Async video research sessions | 25 videos free | Yes |
| Synthetic Users | Simulated user feedback for early concept validation | Free tier available | Yes |
Constraints
4 constraints
- Client Data Warning: Dovetail, Notably, and most AI research tools use uploaded data for model training by default. Check privacy settings and get explicit client approval before uploading real user transcripts.
- Honest Assessment: Synthetic Users is useful for early-stage directional validation but cannot replace real user research. Use it to pressure-test research plans, not as a substitute for human participants.
- Fixed Bid: Cap at 5 interviews maximum. Use Loom for async sessions to reduce scheduling cost. AI synthesis compresses analysis from 2 days to 4 hours.
- AI transcription accuracy is 85-92% on average. Always review before using quotes in deliverables - misattributed quotes in client reports are a serious credibility risk.