I Used to Spend 2 Weeks Analyzing User Interviews. Now AI Does It in 20 Minutes (And Forces Me to Ask Better Questions)

I used to think product discovery was about collecting enough signal to make a confident bet. Talk to 15 users, find the pattern, build the thing. At Sonic Linker, I ran over 40 discovery calls in our first three months. I'd spend hours after each call tagging themes in Notion, looking for overlap, trying to spot the hidden insight.

Then I started feeding those transcripts into Claude and GPT-4. And honestly, it broke something in how I think about discovery.

The Real Shift Isn't Speed (Though That's Nice)

Yes, AI can summarize 10 call transcripts in minutes. Yes, it can pull out sentiment patterns I'd miss on my third pass. That's useful. But the actual shift is weirder and more uncomfortable.

AI doesn't just summarize what users said. It shows you what you didn't ask.

I was running discovery for a link management feature at Sonic Linker. Standard stuff: how do you currently handle this, what's frustrating, would you pay for X. I thought I was being thorough. Then I dumped five transcripts into Claude and asked it to flag assumptions I was embedding in my questions.

It came back with a list. Things like: you're assuming they want centralized control (they don't, they want delegation with guardrails). You're asking about pain with the current tool, but not whether solving this is a priority compared to three other things they mentioned. You keep anchoring pricing discussions around per-user when they're thinking about it as cost-per-campaign.

I went back and re-listened to two of those calls. The AI was right. I'd been steering the conversation toward the solution I already half-believed in.

Discovery Used to Be About Finding Patterns. Now It's About Challenging Your Framing

Before AI, synthesis was the expensive part. So I'd batch my learning. Talk to eight people, then spend two days finding themes. By the time I spotted a pattern, I'd already shaped how I was interpreting it.

Now, I can run a call in the morning and have a structured debrief by lunch. That sounds like a productivity hack. But the real difference is I can ask the AI to argue with my interpretation. I'll feed it a transcript and say: here's what I think the core need is, tell me what I'm missing or over-indexing on.

Half the time, it points out something I glossed over because it didn't fit my hypothesis. A user at one of our early Sonic Linker pilots kept talking about "making sure the links stay live." I tagged it as a reliability concern. The AI flagged it as a trust/accountability issue (they didn't trust their team to maintain links, not the tool). That reframe changed the feature spec completely.

At Finvestfx, I was working on retention features for enterprise treasury clients. I had a theory that we needed better onboarding. I ran six calls, fed them into GPT-4, and asked it to summarize the primary retention blockers. It came back with: integration friction with their ERP system (mentioned by five of six), and lack of role-based permissions (mentioned by four of six). Onboarding quality was mentioned once, in passing.

I'd been solving the wrong problem because I'd already decided what the problem was.

The Uncomfortable Part: AI Makes You Confront How Lazy Your Questions Are

Here's what I didn't expect. Using AI for discovery doesn't just make synthesis faster. It makes you realize how often you're asking soft questions that let users give you the answer you want to hear.

"Would this feature be useful?" is a garbage question. Everyone knows that. But I still caught myself asking versions of it because it felt like forward progress. When you feed that into an AI and ask it to evaluate question quality, it's blunt. It'll tell you: this question assumes the user has already thought about this problem in your framing. It's leading. You got a Yes, but you didn't learn anything.

Now I use AI as a pre-call reviewer. I'll draft my discussion guide and ask Claude: which of these questions are actually open, and which are just fishing for validation? It's humbling. And it's made my calls way more useful.

What I Actually Do Now

I still do the same number of calls. But the loop is tighter and the quality is different. I'll run a call, immediately feed the transcript into AI with a prompt like: "What did this user care about that I didn't follow up on?" or "What's the simplest explanation for their behavior that doesn't require them to be irrational?"

Then I adjust my next call. I'm not batching learning anymore. I'm iterating on my understanding in real time.

At Sonic Linker, this meant we killed two features in our backlog within a month of starting discovery, because the AI kept surfacing that users were solving those problems in ways we hadn't considered (and didn't need us for). That's not a win for our feature count. But it probably saved us from building something mediocre that we'd have defended for six months.

The Takeaway: AI Didn't Replace Discovery. It Made Me Stop Pretending I Was Good at It

I used to think discovery was about gathering evidence. Now I think it's about stress-testing my assumptions faster than I can get attached to them. AI is good at that because it doesn't care about being right. It just reflects back what's actually in the data, not what I want to be in the data.

The bottleneck in discovery was never synthesis time. It was my own confirmation bias and the lag between learning and adjusting. AI just made that lag so short I can't ignore it anymore.