I was recently re-reading Continuous Discovery Habits by Teresa Torres — the go-to resource for PMs learning to apply agile product development principles in their practice. It holds up remarkably well, and there's a wealth of knowledge every PM can benefit from. But it suffers from one thing that makes it feel dated: it doesn't mention AI. Given that the book was published in 2021 — firmly in the pre-ChatGPT era — that's understandable.
Torres has since acknowledged the opportunities AI provides, particularly for summarizing client conversations and identifying common themes in discovery. But I think we need to go further and take stock of just how much AI affords us today. It has become far more mature and versatile than most PMs give it credit for. That's why I want to share the best practices I've developed using AI in my own PM work. Hopefully you can take something from these techniques — and if you have approaches I haven't covered, I'd love to hear them.
Let's dive in.
Start With the Obvious Question — Then Go Deeper
The first step in product discovery is always: "What problem does this customer face?" If you've tried asking an LLM this directly, you'll notice it blends a response from a broad range of sources — which can be both a blessing and a curse.
Let's use car mechanics as an example.
Prompt: What are the main problems a car mechanic faces?
Output: - Increasing vehicle complexity — advanced electronics, software, EV and hybrid systems require constant upskilling. - High cost of tools and equipment — diagnostic scanners, specialized tools, lifts, and software subscriptions are expensive. - Physical strain and injury risk — heavy lifting, awkward positions, exposure to chemicals, and long-term wear on the body.
This is a solid starting point. It saves hours of desk research and helps you imagine customer scenarios you might not have considered. But it has a few core problems.
First, these problems are very broad — they leave a ton of work on the table and aren't remotely actionable yet. Second, there's no clear segment here. There are nearly 29,000 auto mechanics in Canada alone; they don't all face the same problems in the same way. Third, there's no sense of frequency or severity. "High cost of tools" is a real problem early in a mechanic's career, but once they've built out their toolkit, it largely goes away. Not all problems are equal.
> Don't outsource your discovery to AI — use it to expedite your discovery instead.
Niche Down With Follow-Up Questions
Now let's go deeper on one of those problems: "Increasing vehicle complexity — advanced electronics, software, EV and hybrid systems require constant upskilling."
Ask the AI: How often does this problem actually show up, and what types of mechanics face it most?
* General auto technician: Every year (often ongoing)
* EV / hybrid specialist: Constant / continuous
* Diesel / heavy-duty tech: Every 2–3 years (big shifts)
* Diagnostics / electronics: Monthly / ongoing
Now we're getting somewhere. We've moved from a vague problem affecting a massive segment, to one specific type of mechanic facing one specific problem — continuously. That's a much more interesting foundation for a product or business.
That said, none of this is validated yet. There's no guarantee that if you walk into your local general auto shop and pitch a "car training program," anyone will bite. Those techs might exclusively work on older vehicles. The AI's output is a hypothesis, not a fact.
> What AI gives you is a well-reasoned starting point. What you do with it determines whether it's valuable.
Find Your Customer
Here's where AI becomes genuinely powerful for early-stage discovery. Once you've narrowed down a customer segment and a problem worth exploring, you need to talk to real people. In the past, that meant scouring Reddit threads and forums just to figure out where your potential customers hung out. Now you can just ask.
Prompt: How do I find general automotive technicians I could interview about this problem in my local area?
The AI will return a list of local shops to reach out to, tips for approaching conversations, suggestions for what to say, and alternatives if you'd rather not drop by in person. It's a remarkably useful output — and it would have taken hours to compile manually.
But again: none of it matters until you actually go and do it. Until AI can knock on doors autonomously, that part is still your job.
What We've Covered
To summarize the workflow:
1. Use AI to identify potential problems a target customer might face — it's a fast, broad starting point. 2. Ask follow-up questions to niche down to a specific customer type, a specific problem, and how often they face it. 3. Use AI to find your customer — where they work, where they spend time, and how to reach them for interviews.
One Final Point
AI can be wrong. You've heard this before. But here's the thing — so can humans, who also make assumptions and guesses. That's exactly why we test and validate hypotheses rather than acting on them blindly. AI's guess is about as good as yours at this stage, and it gets you there much faster.
The real value is in the division of labour: let AI handle the guessing, and you handle the scientific methodology. That's a powerful combination — especially given how much time and effort proper validation takes.
You might go through all of this and have a mechanic stare back at you blankly. That's okay. That's where learning to pivot during product discovery becomes essential — which I'll cover in the next article. Stay tuned.