AI can simulate user interviews, analyze feedback at scale, and find patterns no human analyst would catch in time. Here's how to use it in your discovery process — and where human judgment is still irreplaceable.
Product discovery has always had a scaling problem. You need to talk to users to understand their problems. But "talking to users" doesn't scale easily — every interview takes scheduling, conducting, transcribing, and synthesizing. A team of two PMs can realistically do 10-15 interviews per quarter if everything goes well. Meanwhile, the product generates thousands of user signals every day through support tickets, app reviews, NPS surveys, and behavioral analytics. Most of that signal goes unread.
AI is changing this. Not by replacing user interviews — qualitative conversations with real humans remain irreplaceable for the depth of understanding they provide. But by dramatically expanding the signal surface that PMs can process and by helping synthesize insights from large volumes of unstructured data.
Where AI Genuinely Helps Discovery
Large-scale text analysis
Feed an LLM a year's worth of support tickets, app store reviews, customer success call transcripts, or NPS survey responses, and it can surface thematic clusters, identify the most commonly expressed frustrations, and highlight patterns that your team has been too busy to read systematically. This isn't a substitute for reading selected samples yourself — it's a tool for deciding which samples to read, by understanding the distribution before you dive in.
Interview synthesis
After conducting user interviews, AI can synthesize transcripts across multiple sessions to identify common threads, contradictions between what different users said, and hypotheses that emerge from the patterns. The PM still needs to do the interpretation, but the initial synthesis is faster and more comprehensive than manual review.
Assumption mapping
When you describe a product idea to an LLM and ask it to identify all the assumptions the idea depends on, it will produce a more complete list than most PMs generate alone. This is useful not because the AI knows which assumptions are correct, but because it's good at identifying what needs to be validated. The PM still has to go validate those assumptions — but starting from a complete map is better than starting from an incomplete one.
Where AI Discovery Falls Short
Simulated user interviews: Some tools offer to simulate user interviews using AI personas trained on your user data. These are interesting for exploring hypothesis space but dangerous if mistaken for real user insight. AI personas reflect the aggregate patterns in existing data. They cannot represent users who don't yet exist in your data — which is often exactly who you should be talking to when doing discovery for new product areas.
Emotional context: User interviews are valuable partly because of what users don't say — the pause before they answer, the way they light up when describing a workaround they're proud of, the frustration that comes through in their tone. AI can process text. It cannot process emotional context from non-verbal signals. For deeply human products, this limitation matters a lot.
Novel problem discovery: AI synthesizes what's already been expressed. The problems users don't know how to articulate, the needs that haven't yet surfaced in feedback, the opportunities that exist in adjacent spaces — these require the kind of creative, exploratory thinking that emerges from genuine human-to-human conversation and observation. AI-assisted discovery is excellent at going deep on known problems. It's not a tool for finding unknown ones.
Improving your discovery process?
The best discovery combines AI-assisted synthesis with genuine human conversation. I've worked on how to integrate both effectively. If you're thinking about your discovery workflow, let's talk through what might work for your context.
Work with Yogesh →