Skip to content
Nerdy beta

AI in Discovery

Add to Space
 Product Discovery AI

LLMs help discovery where the team produces more material than humans can hold in their heads. They hurt when the team uses them to skip the parts that actually produce insight.

Overview

Product discovery has always been throttled by the workday schedule. How many interviews can you run? How many transcripts can you read? How many ideas can you hold in your head before the picture gets fuzzy?

AI elevates this ceiling. You can cluster a hundred transcripts in an afternoon, find the contradictions across sessions, pull verbatim quotes by theme, and code transcripts against your framework. There is significant leverage here for the product professional.

There are downsides, though. Shortcuts produce something that looks like discovery output. None of them produce the thing discovery is for: a team with sharper judgment about which bets are worth making. Synthetic personas flatter the prompt. Ranked opportunity lists pattern-match to consensus. Polished summaries get quoted in place of the customer.

The value-add shift isn’t from “doing discovery” to “AI does discovery.” The bottleneck has moved from human processing to human judgment. Teams that don’t sharpen the judgment layer just get faster at producing plausible-looking artifacts that don’t connect to impact.

Stage by Stage

Framing. Use the model as an adversary: “argue the strongest case against this outcome.” Generate alternative framings before committing. Don’t let it decide the outcome, and don’t anchor on the first framing it produces; you’ll get convergent, plausible-sounding defaults.

Generative research. Stress-test interview guides for leading questions. Generate screener variations to reach harder segments. Don’t substitute synthetic personas for real conversations. You’ll get plausible-sounding nonsense that flatters the prompt. Don’t skip the interview because “the model can probably predict what they’d say.”

Shallow analysis and synthesis. The highest-leverage stage, and the most dangerous. Feed it transcripts and ask for theme clusters, contradictions across sessions, verbatim quotes organized by theme, and what you didn’t hear that you expected to. Don’t quote the summary instead of the customer. Don’t trust the first synthesis pass. The model smooths over outliers, and outliers are where the insight lives.

Opportunity mapping. Force divergence: “give me five approaches, including weird ones.” Generate analogies from other industries to break category lock-in. Don’t use it to score or rank opportunities; it pattern-matches to consensus, not strategy. The value is in keeping the option space open longer, not collapsing it faster. This pairs naturally with Opportunity Solution Trees: use AI to widen the branches, pruning them with your team after.

Assumption testing. Generate copy variants, low-fi prototypes, survey instruments, and falsifiable experiment designs. Don’t treat plausible artifacts as validated ones. Don’t use AI-generated content in a test of whether users want AI-generated content. You’ll get a confounded result. Feed what you learn back into Assumption Mapping so the riskiest bets get tested first.

Decisions, handoff, communications. Draft the decision memo, the pre-read, the exec narrative. Don’t let the model make the call. Don’t ship its prose unedited; it has a tell, and your audience knows it.

Resources