Teresa Torres' continuous discovery framework prevents AI feature bets from becoming expensive guesses. Learn how to validate opportunity-solution trees before shipping AI, not after.
By OperatorRadar Editorial
Teresa Torres introduced continuous discovery habits in 2017 as a structured approach to ongoing customer research. The core practice: PMs spend 5–8 hours weekly interviewing customers, mapping their problems and opportunities, and testing solution hypotheses before committing to roadmaps. The opportunity solution tree (OST) became her signature tool—a visual framework that branches from a desired outcome into customer opportunities, then into solution ideas, each validated through weekly discovery conversations.
Before continuous discovery became mainstream, product teams relied on quarterly roadmap planning cycles and post-launch analytics to validate features. Customer research was often siloed to dedicated research teams or skipped entirely in favor of founder intuition or competitive copying. Torres' framework emerged during the shift toward lean product development and agile methodologies, offering PMs a repeatable, lightweight way to stay connected to customer problems without waiting for formal research sprints.
Torres advocates for discovery as a habit, not a project. The goal is to reduce the time between identifying a customer problem and testing a solution hypothesis. By embedding weekly interviews into the PM role, teams avoid building features in isolation. The OST specifically helps PMs avoid the trap of jumping to solutions: it forces explicit mapping of the customer opportunity before evaluating which solution to build. Torres emphasizes that discovery is never 'done'—it's continuous because customer needs, market conditions, and competitive landscapes shift.
AI feature bets amplify both the upside and downside of skipping discovery. LLMs enable rapid prototyping—a PM can spec an AI feature in hours. But without grounded customer discovery, teams build AI solutions to problems customers don't have or in ways that don't match how customers actually work. AI also introduces new discovery questions: Does the customer trust the AI output? How do they verify accuracy? What happens when the AI fails? These aren't answered by traditional feature discovery alone. Continuous discovery becomes more critical because AI features often require behavioral change from users, and that change only sticks if the underlying opportunity is real.
The core discipline remains unchanged: talk to customers weekly, map opportunities before solutions, test hypotheses before committing resources. The OST structure is equally valid for AI features—it simply adds new branches. For example, an opportunity might be 'reduce time spent on data entry,' and solutions might include 'AI-powered form auto-fill' or 'manual template system.' The weekly cadence prevents teams from over-investing in the AI solution before validating that customers actually want to reduce data entry time or that they trust AI to do it accurately. Discovery habits also catch the 'AI for AI's sake' trap: if no customer opportunity maps to the feature, it shouldn't be built, regardless of technical feasibility.
The assumption that discovery conversations happen primarily in-person or via video calls has shifted. Remote-first discovery is now standard, expanding the pool of customers PMs can reach. Additionally, Torres' original framework predated the need to validate AI-specific concerns—hallucination tolerance, output variability, data privacy with third-party LLM providers. Teams now need to extend discovery to include technical validation questions that weren't part of the original OST. Also, the 'one PM, 5–8 hours weekly' model may not scale for teams building multiple AI features simultaneously; some teams now rotate discovery responsibilities or use structured customer panels to maintain the cadence.
Before shipping any AI feature, map it to a customer opportunity using an OST. Spend 2–3 weeks in discovery mode: interview 8–12 customers about the underlying opportunity (not the AI solution). Ask: What's the current workaround? Why does it matter? How would they verify the AI output is correct? Only after validating the opportunity should you prototype the AI solution and test it with the same customers. This prevents building polished AI features that solve non-problems. For teams under pressure to ship fast, treat discovery as a gate: if you can't articulate the customer opportunity in one sentence, you're not ready to build the feature, no matter how technically feasible the AI is.
Take your current AI feature roadmap. For each feature, write down: (1) the customer opportunity it solves, (2) how many customers you've interviewed about that opportunity (not the solution), and (3) what would make a customer trust the AI output. If you can't answer all three, that feature needs discovery before design.
Teresa Torres
Continuous Discovery Habits: Discover Products That Create Customer Value and Market Fit
“Torres advocates for PMs to spend 5–8 hours weekly in customer discovery conversations, using an opportunity solution tree to map customer opportunities before evaluating solutions. The framework emphasizes discovery as an ongoing habit, not a one-time project.”
Teresa Torres
“The OST is a visual framework that branches from a desired outcome into customer opportunities, then into solution ideas. Each branch is validated through customer discovery before committing resources.”
OperatorRadar Editorial
AI Feature Validation and Discovery
“AI features amplify the cost of skipping discovery because they enable rapid prototyping but also introduce new validation questions around user trust, accuracy verification, and behavioral change.”
Struggling to validate your AI feature roadmap? OperatorRadar's discovery coaching service helps PMs run structured customer interviews and map opportunities before building. Schedule a 30-minute consultation to audit your current discovery process.
Request implementationNeed a custom implementation?
Have Ekofi Lyrae design and implement AI agents and automations.