User research has traditionally been handled by dedicated research teams, and over the years I have been deeply involved in multiple end-to-end projects — from defining research goals and recruiting participants to conducting interviews and synthesizing insights. At my most recent job, however, we didn't have an in-house research team. So I took on the qualitative research independently, using AI to work efficiently throughout the process.
Many people assume that once you know how to use AI tools, you can automate most research work. My experience has been the opposite: AI can significantly improve research efficiency, but only when it is guided by strong domain expertise and sound professional judgment.
That experience was essential because research requires judgment at every step: defining the right target users, designing effective screening criteria, identifying appropriate questions, recognizing meaningful signals during interviews, and turning raw feedback into actionable insights for product, design, engineering, and marketing teams. AI could support these tasks, but it could not replace the underlying research judgment.
In this article, I will walk through the full process I used — from planning and participant recruitment to interviews, synthesis, and reporting — and explain where AI created leverage and where professional experience remained essential.
The Process
Stage 1: Planning and Preparation
Before getting started, it's important to define the purpose and strategy of the research. Our team builds an AI video product, so my goal was to understand the video creation process of AI creators. At the same time, a new feature had recently been launched in the product, so the team wanted me to collect early feedback on it.
This was an important stage where professional judgment mattered more than tooling. AI could help me organize materials, but it could not decide what we actually needed to learn, which users were most relevant, or how to balance exploratory research with product feedback goals.
Stage 2: Define participant criteria
Because most of our users were based in North America, I defined the initial target profile as follows:
- North American users
- Paid users of our product
- Heavy users of AI tools
I then contacted users who matched these criteria and asked whether they were interested in participating. At the same time, I shared the relevant details with them, including the interview format, session length, incentive amount, NDA requirements, and available time slots.
After identifying more than 40 potential participants, I sent out a survey to gather more detail and narrow the pool further.
AI helped me draft the survey questions, but I did not rely on its output blindly. I provided detailed context, specified the kind of information I needed, and manually revised the final version. This was necessary because AI does not reliably understand what makes an interview question useful, what introduces bias, or what information is actually decision-critical.
After receiving responses, I used an additional set of filters:
- Frequent AI video creators who created content more than five times per week
- Creators with several years of relevant experience
- Users with a meaningful monthly budget for AI video tools
This stage typically took about a week. Once I had enough valid responses and at least five qualified participants, I moved to the interview stage.
Stage 3: Scheduling Interviews
I first identified the time slots that worked for my schedule, then sent confirmation emails to a small group of candidates first to test the process. This helped me catch potential issues before contacting the rest of the group.
Because the participants and I were in different time zones, scheduling required careful coordination. I made sure the proposed times worked for both sides and highlighted time zone details clearly in the confirmation emails.
This may sound operational, but it also reflects research discipline. Small coordination mistakes can reduce participation rates or create friction before the interview even starts.
Stage 4: Preparing the Interview Script
To prepare the interview guide, I gave Claude detailed context: the background of the project, the research goals, my hypotheses, and the screening survey results. Based on that input, the AI helped me generate a draft structure that included:
- An introduction to the session
- Participant background
- The participant’s video creation workflow
- Usability testing within our product
The AI generated the first draft quickly, but the initial version was not good enough to use directly. I iterated several times, added more specific requirements, and revised the structure manually. I also rehearsed the guide before the actual interviews.
This stage clearly showed the difference between tool usage and professional capability. Claude could draft questions, but it could not reliably determine which questions were too broad, which ones would lead to superficial answers, or when a prompt would fail to uncover motivations.
For example, instead of asking, “Can you describe your content creation process?”, it was much more effective to ask, “Can you walk me through the specific steps, the AI tools you used at each step, and how you used them?” The latter question is more concrete and much more likely to reveal real behavior.
I also learned not to overload the interview guide. A one-hour interview should usually include no more than about 15 main questions. The goal is not to ask as many questions as possible, but to create enough space for follow-up questions and deeper investigation.
At its core, qualitative research is about going beyond stated behavior to understand decision-making, motivations, tradeoffs, and pain points.
Stage 5: Conducting the Interviews
Before each interview, I made sure everything was ready: the meeting software, recording setup, note-taking workflow, and checklist.
During the interviews, I generally followed the script, but not strictly. Different participants communicated in different ways. Some answered in great detail and naturally covered multiple topics at once. In those cases, I avoided interrupting them unless they drifted too far off track. Instead, I adjusted the order of questions dynamically.
If a participant was less talkative, I used follow-up questions to help them go deeper and provide more concrete details. If something was unclear, I clarified it immediately rather than moving on with uncertainty.
This is another point where research experience matters. Running a strong interview is not about reading questions in order. It is noticing what matters, knowing when to dig deeper, and adapting in real time.
I also used AI meeting tools to record and transcribe each session automatically. That saved time and improved documentation quality, but the transcript itself was not the insight. It was only raw material for later synthesis.
Stage 6: Organizing Results and Writing the Report
After the interviews, I used AI to help organize the findings and draft the final report. I compared outputs from two different AI tools (Claude and OpenClaw) and cross-checked them against my own notes and understanding of the interviews.
I gave both systems the research context, interview transcripts, target audience for the report, and the desired report structure. This helped them produce more relevant outputs.
Even so, the results were not equally strong. I used both outputs as supporting materials, then created the final report myself through manual synthesis and editing.
This stage reinforced an important lesson: AI becomes much more effective when the user provides a clear analytical frame, a defined audience, and a strong output structure. Without that, it tends to produce generic summaries.
After the report was complete, I asked both systems to polish the language and summarize key insights into a concise version that I could quickly share with cross-functional teams.