This week, I used Claude to help me prepare for and process customer interviews for my project. I learned more about where AI is genuinely useful in this process, and where it cannot do the work for me.

In class this week we covered customer validation, including how to plan interviews and surveys, and we practiced with mock customer interviews. The assignment was to conduct real interviews, including at least one person outside of my own network.

Firstly, I used Claude to brainstorm interview questions. I gave it context on who I was interviewing, the AI-generated persona responses from Week 3, and the assumptions that came out of that work which still needed to be tested with real people. It gave me a list of questions to work from. However, I removed several of them because they did not make sense given the context, and because they were too specific for the people I was actually speaking to. They were questions my interviewees most likely would not have had an answer to.

This experience highlighted that Claude could generate questions that sounded reasonable on their own, but it did not know who I was actually sitting across from. I had the context about these individuals that the model did not, so deciding which questions to keep was something only I could do.

I also used Claude after the interviews to help structure and organize my notes. What it gave me back was overly structured and far more elaborate than what I needed. More importantly, it did not reflect the analysis I had personally done from the interviews. I rejected a lot of it and rewrote the synthesis myself so that the documentation actually represented my own thinking rather than a framework the model applied on top of my notes.

This was a different limitation from the first one. With the interview questions, the issue was missing context. Here, the issue was that AI can make something look more analyzed than it really is. The conclusions I drew from talking to these people were mine, and if I had kept what Claude produced, the write-up would have represented its organization of my notes rather than what I actually took away from the conversations.