--- title: "AI in product discovery: do the synthesis yourself first" description: "Where a language model speeds up discovery, what it drops from an interview, and the second-reader rule that keeps a team’s customer understanding intact." url: https://torn.studio/en/insights/ai-in-product-discovery locale: en published: 2026-09-07 --- # AI in product discovery: do the synthesis yourself first > **Short answer:** Use AI in product discovery as a second reader. A person synthesizes every interview first, a model synthesizes it second, and the two readings are compared. Teresa Torres’s December 2025 rule is that dumping transcripts into a model is no shortcut. Models are good at desk research, clustering and drafting, and they miss what a customer left unsaid. A product team that runs interviews now has a model that will summarize them in seconds, cluster the feedback and draft the opportunity tree before lunch. The question is what that does to the team’s understanding of its customers, and the honest answer from the people who have tested it is that everything depends on who reads first. ## What is a model good at in discovery? Three things, and they are the laborious parts. Desk research: reading a market, a competitor set or a regulation and returning a structured brief with the gaps marked. Clustering: taking two hundred support tickets or a hundred survey answers and grouping them by the problem underneath, which a person then checks. Drafting: turning a synthesis into an interview guide, a hypothesis list or a first cut of an opportunity solution tree. The studio used a model this way on its own product, this website: the article questions come from an autocomplete harvest across ten markets, run on 30 August 2026 with a script any reader can re-run, and the ordering it produced set the priority for three content clusters. The harvest is a measurement, so it is checkable, and it has a known hole: it gives an ordering and no volumes, which is why the priority stays a hypothesis until a paid tool fills that column. ## What does a model miss in an interview? The thing the customer did not say. Teresa Torres, whose continuous discovery method most product teams use in some form, published her position on 2 December 2025 after six months of experiments: “You still need to synthesize every interview individually. Dumping transcripts into an LLM isn’t a shortcut.” Her practice, restated in an April 2026 interview, is the one this article recommends: do your own synthesis first, then ask the model for its own, then compare. Each time, the person catches things the model missed and the model catches things the person missed. Nielsen Norman Group reached the same conclusion from the research side in June 2024, on AI-generated “synthetic users”: “UX without real-user research isn’t UX.” The legitimate uses they allow — desk research, hypothesis generation, preparing a guide — are the three things above. Concept validation and decisions are the two things a synthetic respondent may never carry. ## What is the second-reader rule? Every interview is read twice, in a fixed order. - The person who ran the interview writes the synthesis first: the problem in the customer’s words, the workaround they use today, and what surprised the interviewer. - The model reads the same transcript and writes its own synthesis, from a prompt that asks for the same three things and for the quotes it relied on. - The two are laid side by side. The disagreements are the output: each one is either a thing the person missed or a thing the model invented, and both are worth a minute. The order is the whole method. Read the model first and the person’s synthesis becomes an edit of it. That is the lazy synthesis Torres warns about, and the fastest way for a team to stop understanding its customers while producing more documents about them. ## How does this change what a product manager does? It moves the hours. Torres’s own framing, from the same December 2025 conversation, is that the real gains come from expert plus AI, and that beginner plus AI is usually better than nothing. The expert’s job is to be the first reader, to write the prompt that asks the right three questions, and to decide what the disagreements mean. That is a product decision, and it has the same shape as [deciding what to build next](https://torn.studio/en/insights/deciding-what-to-build-next): written down, with what was set aside. The same move — measure first, rank second, decide in writing — is how the studio orders processes in [which processes to automate first](https://torn.studio/en/insights/which-processes-to-automate-first), and it is why the two articles read alike. ## How the studio runs discovery Torn Studio takes the product role for founders and product teams, priced per engagement, and discovery is the first two to four weeks of it: interviews synthesized by a person and then by a model, a written problem list with the disagreements kept, and one page per decision. What the team keeps afterwards is the list and the reasoning, in a form it goes on using once the engagement ends. **Read next** - [Product Management at Torn Studio](https://torn.studio/en/services/product-management) - [For founders on a runway](https://torn.studio/en/for/founders) ## Common questions ### Can we let a model summarize our customer interviews? Yes, as the second reader. Write your own synthesis first, then ask the model for its own from the same transcript, then compare. Torres’s December 2025 position is that dumping transcripts into a model is no shortcut, and her practice is to compare the two readings every time. ### What does an AI summary of an interview typically leave out? What the customer did not say: the workaround they use today, the hesitation before an answer, the question they asked back. A model summarizes what is in the transcript; the person in the room carries the rest, which is why the person reads first. ### What can we use a model for in desk research? Reading a market, a competitor set or a regulation and returning a structured brief with the gaps marked, and clustering a large pile of tickets or survey answers by the problem underneath. Both outputs are hypotheses a person checks; neither is a finding. ### How do we know our discovery has become lazy? When the model’s summary is the first thing anyone reads. The tell is a team producing more documents about customers while fewer people can quote one. The fix is the order: person first, model second, compare. ### What did the studio measure with a model in its own discovery? The search demand behind this site’s articles: an autocomplete harvest across ten markets and six service areas on 30 August 2026, run with a script a reader can re-run. It gives an ordering and no volumes, and the priority it set stays a hypothesis until a paid tool fills that column. ### How long does discovery take with Torn Studio? Two to four weeks at the start of an engagement, with the price fixed before work starts. The output is a written problem list with the disagreements between the human and model readings kept, and one page per decision that the team keeps using afterwards. ## Sources - [Customer Interview Analysis — All Things Product Podcast with Teresa Torres & Petra Wille — Product Talk](https://www.producttalk.org/customer-interview-analysis-all-things-product-podcast-with-teresa-torres-petra-wille/) — Backs Torres’s 2 December 2025 position: every interview is synthesized individually, dumping transcripts into a model is no shortcut, and the real gains come from expert plus AI. - [“Building is cheap now, but don’t build everything”: Teresa Torres on AI product management — airfocus](https://airfocus.com/blog/teresa-torres-ai-product-management/) — Backs the April 2026 practice: own synthesis first, then the model, then compare, with each reading catching what the other missed. - [Synthetic Users: If, When, and How to Use AI-Generated “Research” — Nielsen Norman Group](https://www.nngroup.com/articles/synthetic-users/) — Backs the June 2024 conclusion that UX requires research with real users, and the three permitted uses of synthetic respondents.