Before any AI touches the data, I design and run the study myself. I write the discussion guide, set up the unmoderated tasks, recruit participants, and take observation notes throughout. That hands-on involvement is not optional. It is what makes my review of the AI output meaningful later. I can only catch what the AI gets wrong if I was the one in the room.

After sessions wrap, I clean and label the transcripts, organise my observation notes, and structure everything before it goes anywhere near an AI. I treat this step with the same care as the study design itself. Poorly structured input produces poorly structured output, and the AI has no way of knowing the difference. I do.

With the data prepared, I use AI to surface patterns, cluster themes, and flag connections across sessions. It can process a full set of transcripts faster than I can and sometimes surfaces links I might not have noticed on my own. But it cannot tell me which patterns matter and which are noise. That judgment comes from having been in those sessions, and it stays with me.

I go through every theme the AI surfaces and test it against the raw data. Some hold up exactly as presented. Some need to be merged or reframed. Some get cut entirely, because the context I carry from those sessions makes them misleading in ways the AI would have no way of knowing. This is the step I care about most. It is where the actual insight gets made.

I turn the refined findings into something other teams can actually use. That might be a synthesis deck for a stakeholder presentation, or something more practical. At Canada Post, I built a structured feedback template for a development team with no prior experience collecting user feedback. I used AI to draft and iterate the tool quickly, then calibrated it with my domain knowledge to make sure it captured the right signals. AI made it faster to build. I made sure it was worth using.