I Used to Spend 3 Weeks Synthesizing User Interviews. Now AI Does It in 20 Minutes. Here's What Changed.

I remember sitting in a coffee shop in Koramangala last year, drowning in 47 pages of user interview notes. I'd talked to 12 Forex traders for Finvestfx over two weeks, transcribed everything manually, and now I had to find patterns. Three days later, I had a synthesis doc. By the time I shared it, two of those users had already churned.

That was discovery before AI. It worked, but it was brutally slow.

The Real Shift Isn't Speed, It's Scale

Everyone talks about AI making things faster. Sure, fine. But the actual change is this: I can now talk to 10x more users because the synthesis bottleneck is gone.

At Sonic Linker, we were building an AI platform for link optimization. In the first month, I did maybe 8 user calls. Standard PM stuff. Took notes, tagged themes, built a mental model of what mattered. But I kept hitting this ceiling where I'd hear something interesting and think, "Is this just one guy's opinion or a real pattern?" I didn't know because I couldn't afford to talk to 30 more people just to validate a hunch.

Then I started recording calls (with permission, obviously) and feeding transcripts into Claude and GPT-4. Not for answers, but for pattern recognition. I'd ask: "What are the top 3 friction points mentioned across these 15 calls?" or "Which users talked about speed vs. accuracy, and what were their actual words?"

Suddenly I could talk to 40 users in a month. The synthesis happened in near real-time. I'd finish a call at 3 PM, have the transcript processed by 4 PM, and update my discovery doc by 5 PM. That pace changes everything. You stop guessing. You start knowing.

AI Catches the Stuff You Miss (Because You're Human)

Here's the uncomfortable truth: I'm biased. You're biased. We all go into discovery calls with a hypothesis, and we hear what confirms it.

I was convinced Sonic Linker users cared most about link performance metrics. I kept steering conversations there. But when I fed the raw transcripts into an LLM and asked it to surface the most-mentioned pain points without any framing, it kept flagging "setup complexity" and "unclear pricing."

I'd heard those words in calls. I just didn't weight them properly because I was anchored on my hypothesis. AI doesn't have that problem. It's ruthlessly literal. It counts. It clusters. It doesn't care what you think matters.

This isn't about replacing judgment. It's about getting a second pair of eyes that doesn't share your blind spots. I still made the final call on what to prioritize, but now I had data that wasn't filtered through my confirmation bias.

The Questions You Can Afford to Ask Change

Before AI, I'd batch discovery. Talk to 10 users, synthesize for a week, build something, repeat. The cycle time was long, so I had to be selective. I couldn't afford to explore tangents.

Now? I can chase hunches in real time.

At Finvestfx, a enterprise client offhand mentioned they were manually exporting data to reconcile with their ERP system. I flagged it but didn't think much of it because we were focused on core treasury workflows. Two weeks later, I'd talked to 8 more clients. AI synthesis showed me that 6 of them had mentioned "reconciliation" or "export" in some form. I went back, asked targeted follow-ups, and realized we were sitting on a retention play. We shipped a bulk export feature in two weeks. Retention among those clients jumped 18% in the next quarter.

I would have missed that pattern entirely in the old model. The synthesis lag was too long. The signal would have decayed.

What This Means for How I Work Now

I talk to more users. I ask weirder questions because I know I can process the answers. I don't worry as much about "wasting" a call on an edge case because synthesis is cheap now.

But here's the catch: AI makes bad discovery faster, too. If you're asking the wrong questions or talking to the wrong users, you'll just get confident about the wrong thing more quickly. The tools don't fix strategy. They amplify execution.

I still spend time upfront defining who I need to talk to and why. I still write a discussion guide. I still listen for tone and emotion, not just keywords. AI handles the grunt work of synthesis, tagging, and pattern recognition. I handle the judgment calls on what it means and what to do next.

The Takeaway

AI didn't replace discovery. It removed the excuse that "talking to 50 users is too expensive." It's not anymore. The synthesis bottleneck is gone. If you're still doing discovery at the same pace you were two years ago, you're leaving insights on the table.

The PMs who win in the next few years won't be the ones with the best AI tools. They'll be the ones who realized the constraint shifted from "how do I synthesize this" to "who should I talk to next" and acted accordingly.