AI Made My Discovery Faster. I'm Not Sure It Made It Better.

I used to rewatch every customer interview at 2x speed.

Not because it was a good use of time, but because I didn't trust the notes I'd taken during the call. Nine discovery interviews meant two or three hours of rewatching, then another stretch pulling scattered notes into something that resembled an insight.

AI does that in minutes now. Cleaner transcripts than I'd ever type, themes clustered, quotes pulled, contradictions flagged. The output is genuinely good.

Tasks and thinking are not the same work

The frame I keep coming back to is the difference between a task and a thought.

Writing a PRD is a task. Transcribing an interview is a task. Turning fifty pages of notes into five themes is a task. These are important, repetitive, and largely mechanical. Hand the same raw material to five competent PMs and you'll get five broadly similar outputs.

Deciding which of those five themes deserves a quarter of engineering time is not a task. Deciding which customer to interview, and what to ask them, is not a task. Killing a well-researched idea because something shifted in the market is not a task. Hand the same inputs to five PMs and you get five different answers. The spread between those answers is where the value of the role actually sits.

The bit I got wrong about my own example

I need to be honest about the interview story, because I framed it too conveniently.

Rewatching those calls wasn't pure tedium. Somewhere in the third hour, sitting with the raw material (the pause before an answer, the thing a drilling supervisor mentioned twice that I brushed past both times, the question I asked badly), I'd notice something. Not because I was analysing. Because I was immersed.

The synthesis was a task. The immersion was not. They just happened to occupy the same three hours.

That's the sharper risk in automating discovery work. Not that AI produces bad summaries. It doesn't. It's that the summary arrives so clean, so complete, so finished, that you never sit with the mess. And you never notice the question you should have asked.

DORA's 2025 research points at the same mechanism from the engineering side: AI lowers the entry barrier to unfamiliar work, but in doing so it can bypass the productive struggle that builds real expertise. You get the answer without acquiring the understanding that would let you judge whether the answer is any good.

What AI still cannot do in a discovery call

It cannot make sure the right question gets asked.

Maybe future models will run the interview themselves, catch the hesitation before an answer, resist leading the witness. Maybe. But asking the right question depends on context the model doesn't have: that this customer's ops director is politically exposed on the decision you're probing; that the last three accounts said the same thing and all three were wrong; that the thing they're complaining about is a symptom of something they haven't got words for yet.

That context sits in the overlap between the PM, the designer and the engineer, each arriving with different priors and different scar tissue. It gets surfaced through argument, not synthesis. No agent automates that, and I don't think it's close.

Faster is not the same as better

AI will produce more software. More features shipped, more prototypes, more variants tested. That much is already visible.

Google's DORA research, covering roughly 5,000 respondents in 2025, found that AI adoption now correlates with higher delivery throughput, a reversal from the previous year. It also correlates with higher delivery instability. Teams adapted for speed; the systems around them didn't adapt to absorb it.

DORA's framing is that AI is an amplifier. It magnifies whatever is already there, strength or dysfunction. And of the seven capabilities they identify as determining which way it goes, one is a user-centric focus. Without it, their data suggests AI adoption can make team performance worse, not better.

That's the whole argument in a single finding. Speed applied to the wrong thing is just faster waste.

So what actually changes

Nothing about good product work changes. Craft, genuine collaboration between engineer, designer and PM, trial and error through discovery: those were always the ingredients, and no agent replaces them.

What changes is the ratio. When synthesis takes minutes instead of hours, the constraint moves upstream, into the decisions. Which problem, which customer, which question, which idea to kill.

The practical version: take one week of your own work and split it into two columns, tasks and thinking. Automate the first column aggressively. Then check whether you actually spent the reclaimed hours in the second column, or whether you just did more tasks.

Most of us, I suspect, will find we just did more tasks.