I Used to Spend Weeks on Discovery Calls. Now AI Does the First Pass in Hours.

I used to think product discovery meant hopping on Zoom calls, taking notes, and looking for themes across 15 user conversations. That was the job. Synthesize feedback, spot patterns, build a story.

Then at Sonic Linker, we started feeding customer support tickets, user session transcripts, and sales call recordings into GPT-4. Not to replace discovery, but just to see what would happen.

What happened was: it found 3 pain points I had completely missed after weeks of manual analysis. And it did it in about 90 minutes.

That's when I realized AI isn't changing discovery by making it faster. It's changing what counts as discovery in the first place.

The old way was expensive, so we pretended it was thorough

Before AI, discovery had a built-in limit: your time. You could realistically talk to maybe 10-15 users a month if you were hustling. So we built entire frameworks around small sample sizes. We called it "qualitative research" and acted like 12 interviews were enough to understand thousands of users.

I did this at Finvestfx. We had 20+ enterprise clients, and I would pick 4-5 treasury managers to interview every quarter. I'd write up themes, add some quotes to make it feel rigorous, and call it validated learning. But honestly? I have no idea if those 5 people represented the other 15. I just didn't have time to talk to everyone.

AI removes that constraint. Now I can analyze feedback from 200 users in the same time it used to take me to manually review 10 call transcripts. That's not a small shift. That's a different game.

Discovery used to be about talking. Now it's about listening at scale.

At Sonic Linker, we integrated a simple feedback widget into the product. Users could drop a sentence about what wasn't working. Nothing fancy. We got about 60-80 pieces of feedback a week.

Old me would have skimmed them, maybe tagged a few, and moved on. There's only so much unstructured text you can process before your brain starts pattern-matching based on what you already believe.

But we started running all of it through an LLM with a prompt like: "What are the top 3 problems users are trying to solve that we're not addressing?" And then: "What language are they using to describe this problem?"

The second question turned out to be more important. Because users don't say "I need better data pipeline orchestration." They say "I just want this thing to stop breaking when I add a new source." The AI caught that. I wouldn't have, because I was too busy translating everything into product-speak.

We shipped a feature based on that insight in two weeks. Retention for that cohort jumped 18%. Not because we built something groundbreaking, but because we finally understood what the problem actually felt like to users.

AI makes you honest about what you don't know

Here's the uncomfortable part: AI is really good at showing you when you're wrong.

I had a hypothesis at Finvestfx that our enterprise clients wanted more customization options. It felt true. A few loud customers had asked for it. I was ready to build.

But when we analyzed support tickets and call logs with AI, the data said something different. The top issue wasn't lack of features. It was that people didn't understand how to use the features we already had. The real problem was onboarding and education, not customization.

That sucked to hear. Because I had already sold the customization story internally. But the AI didn't care about my narrative. It just showed me what users were actually saying, at scale, without my confirmation bias filtering it.

That's the shift. Discovery used to be about building a story from limited data. Now it's about pressure-testing your story against way more data than you can ignore.

What this means for how I do discovery now

I still talk to users. That hasn't changed. But now the conversations are different.

I use AI to do the first pass: analyze feedback, highlight patterns, flag edge cases I wouldn't have noticed. Then I take those insights into user calls and ask: "Is this actually what's going on, or is the data misleading me?"

The calls are shorter. More focused. I'm not trying to discover everything from scratch. I'm validating or challenging what the AI already surfaced. It's faster, and honestly, it's better.

At Sonic Linker, this meant we could talk to 10x more users (through async feedback and AI analysis) and still do deep-dive calls with the most interesting cases. We weren't choosing between scale and depth anymore. We could do both.

The real change is that 'I don't know' is harder to hide behind

The biggest shift AI brings to discovery is accountability. You can't just say "we talked to users" anymore. You have to say "we analyzed 200 data points, here's what we found, and here's why we're prioritizing this."

That's uncomfortable. But it's also way better.

Because discovery was never supposed to be about having good taste or trusting your gut. It was supposed to be about understanding users better than anyone else. AI just makes it harder to fake that.