Discovery Sprint with AI: A Five-Day Plan and a Three-Hour Version
A discovery sprint time-boxes one big product question into a week. Use AI to mine the evidence you already have and to run interviews faster, and know what three hours can and cannot tell you.
Share
A discovery sprint is a time-boxed effort, usually one week, in which a small team answers one substantial product question or risk before committing to build. With AI, the first day of evidence review can shrink to about three hours, and interviews can run without waiting on calendars. The customer contact still takes days, and the decision still belongs to the team.
Key takeaways
Marty Cagan describes this format as a one-week time box aimed at a substantial problem or risk; he prefers the term to "design sprint" because good discovery goes beyond design (INSPIRED, 2nd ed.).
GV's design sprint, the best-known format, runs five days with seven people or fewer and ends with five customer interviews on Friday (Character).
AI is fastest at the evidence you already have: searching calls, tickets and notes, clustering themes and drafting a brief. It cannot create new customer evidence in three hours.
AI interviewers suit short, structured questions. In a 2026 Nielsen Norman Group study, only 3 of 10 participants found an AI-moderated conversation natural (NN/g, 2026).
A sprint ends in a written decision, not a deck.
What is a discovery sprint?
It takes one big question, such as "Should we build team workspaces?" or "Why do trials stall in week two?", and gives it a week of focused attention from the people who will build the answer. Cagan places it in product discovery: the work of checking value, usability, feasibility and business viability before engineering time is committed (SVPG, 2017).
It is not the same as a delivery sprint. A Scrum sprint ships increments; this kind ships a decision and the evidence behind it.
How is a discovery sprint different from a design sprint?
The names get used interchangeably, but the formats answer different questions.
The design sprint was developed at GV: Jake Knapp began running sprints at Google in 2010 and brought them to GV in 2012. Monday maps the problem, Tuesday sketches, Wednesday decides, Thursday prototypes and Friday tests with customers (GV). Knapp and John Zeratsky later added the two-day Foundation Sprint for the step before: no more than five people, including the Decider, agree a hypothesis that design sprints can then test (Lenny's Newsletter, 2025).
Choose Cagan's format when the problem itself is uncertain. Use a design sprint when you trust the problem and need to test a solution.
How do you run a discovery sprint with AI in five days?
Block the week for a small team: a product person, a designer and an engineer, plus whoever makes the final call. The design sprint calls that person the Decider, and its guide is blunt that decisions without one won't stick (Character).
Day 1: Frame the question and mine existing evidence
Write the sprint question as a decision: "By Friday, we will decide whether to [act] for [user] because of [evidence]."
Use AI search across calls, support tickets, sales notes and feedback to pull every mention of the topic from the last 90 days. Ask for quotes with sources, not summaries.
Cluster what comes back into three to five themes. For each, record how many sources support it ("7 tickets, 3 calls") and what is missing.
Ask the AI to argue against your favourite theme. If the evidence doesn't hold up, say so now.
This is the step AI shortens most: the machine does the reading, and the team does the judging.
Day 2: Map opportunities and pick the riskiest assumption
Put the themes on an opportunity solution tree: the desired outcome at the root, then customer opportunities, then candidate solutions, then the assumption tests underneath (Product Talk).
List assumptions for the top one or two opportunities (the free assumption map helps) and score each 1–5 for impact and uncertainty. Test the highest product.
Write the pass signal for each test before you see any data.
Start recruiting for days 3 and 4. Invite people who hit the problem in the last month.
Day 3: Talk to customers
Run five to eight short interviews with people who match your screener, using story-based questions about the last time the problem happened. Our customer interview template has a 20-minute script.
AI helps in two ways. It transcribes and tags live calls, and an AI interviewer can run short, structured conversations from a link, so people answer when it suits them. Keep at least three live interviews for anything exploratory: Nielsen Norman Group found AI interviewers struggled to probe unexpected answers, ran anywhere from 13 to 56 minutes because they couldn't hold a target length, and recommends them as a complement to human moderation (NN/g, 2026).
Day 4: Test the riskiest assumption
Build the smallest thing that puts the assumption in front of users: a clickable mock, a concierge version, a fake door or a thin prototype built with a coding agent. For usability questions, five participants is a sound default; Jakob Nielsen's model shows five users finding about 85% of usability problems (Nielsen Norman Group, 2000).
If the risk is value rather than usability, test behaviour: a sign-up, a booked call, a pre-order. Our guide to choosing a product experiment maps each risk to a method.
Day 5: Decide and write it down
Compare each result with the pass signal you wrote on day 2. Then write a one-page decision record:
The question and the decision (continue, change, pause or stop)
The evidence for and against, with sources
What you are deliberately not doing yet, and why
The next test, if any
Share it with the team that wasn't in the room. A decision nobody can trace back to evidence gets reopened at the next planning meeting.
Can you compress the sprint into three hours?
Partly. A three-hour version can answer one question well: what does the evidence we already have say, and what is missing? It cannot create new customer evidence, because the interviews and tests on days 3 and 4 depend on other people's time. Treat it as day 1 compressed, not as the whole sprint.
Hour
Work
Output
1
Search calls, tickets, Slack threads and notes on the topic with AI; pull quotes with sources
3–5 evidence clusters, each with a source count
2
Turn clusters into opportunities; list assumptions; score impact × uncertainty; send a short interview link to 10–20 relevant users
A ranked list and one riskiest assumption
3
Draft a one-page brief: what we know, what we're recommending, what we're deferring, what the interviews must show
A brief ready for the team, plus a pass signal
Then read the interview answers as they arrive over the next few days, before acting on the brief. The three-hour version suits single-topic questions with plenty of existing evidence. It does not suit a new market, where you have no evidence to mine.
What does a five-day sprint look like in practice?
The following is illustrative, not a real company.
Say a six-person B2B analytics startup sees trials stall in week two and wants to know why before redesigning onboarding.
Day 1. AI search across 140 support tickets and 22 sales-call notes finds three clusters: data import errors (31 mentions), unclear first dashboard (18) and pricing questions (9). The team had assumed pricing was the main issue.
Day 2. The riskiest assumption: "trials stall because the first import fails, not because of the dashboard." Pass signal: at least 4 of 6 stalled trialists describe an import problem unprompted.
Day 3. Six interviews (three live, three by AI interviewer link). Five of six describe an import failure; two never reached a dashboard at all.
Day 4. A concierge test: the team imports data by hand for five new trialists. Four of five reach a first dashboard within a day.
Day 5. Decision: change. Fix import first, defer the onboarding redesign, and recheck week-two activation after four weeks.
The sprint cost about 15 person-days. The redesign it postponed would have taken far more, and would have fixed the wrong step.
What mistakes derail a sprint?
Picking a question too big for a week. "What should our strategy be?" isn't a sprint question. "Should we build X for Y?" is.
Starting with a solution. If Monday opens with mock-ups, it's a design sprint with a different name.
Skipping the Decider. Without one person who can say yes, the decision drifts back into meetings.
Treating AI summaries as evidence. A summary without sources is an opinion. Insist on quotes and counts.
Letting synthetic users fill the interview slots. Simulated answers are not customer evidence.
Recruiting on Wednesday. Without participants booked on day 2, day 3 is empty.
When not to run one: if the question needs prevalence data (run a survey), if you can't reach any users this week, or if the team will build regardless of the outcome. In that last case, save the week.
Where Specky fits
Specky shortens day 1. It reads customer calls, feedback, Slack, Jira and analytics into a Product Graph, so a search for the sprint topic returns quotes with their sources and counts. Alex, Specky's research agent, sends short interviews from a shareable link and files transcripts, themes and sentiment back into the graph. The Opportunity Tree holds day 2's map, and the Decision Log keeps day 5's record next to the evidence that drove it. For a lighter, asynchronous format, see how to run discovery asynchronously.
Frequently asked questions
What is a discovery sprint?
A discovery sprint is a short, time-boxed effort, usually one week, in which a small product team answers one substantial question or risk before committing engineering time. It combines a review of existing evidence, customer interviews and a small test, and it ends with a written decision. Marty Cagan uses the term in INSPIRED in preference to "design sprint".
How long should a discovery sprint be?
Most teams use five working days, which leaves room to recruit, interview and test. Some run two weeks so the sprint matches their delivery rhythm. A one-day or three-hour version can review existing evidence and frame the question, but it cannot replace the days needed to talk to customers and observe a test.
What is the difference between a discovery sprint and a design sprint?
A discovery sprint asks whether a problem or risk is worth acting on; a design sprint asks whether a specific solution works. GV's design sprint follows a fixed five-day agenda ending in five customer tests of a prototype. Cagan's version is looser and often ends with a decision rather than a prototype.
Who should be on the sprint team?
Keep it small: a product manager, a designer and an engineer who will build the result, plus a Decider with authority to act on the outcome. GV's design sprint guide caps the team at seven people. Bring in experts, such as support or sales leads, for short interviews on day 1 rather than for the whole week.
Can AI run the customer interviews in a sprint?
For short, structured questions, yes: an AI interviewer can collect answers from more people without scheduling. For exploratory questions, keep humans in at least some interviews. Nielsen Norman Group's 2026 study found AI interviewers struggled to follow unexpected threads, and only 3 of 10 participants said the conversation felt natural.
GDPR-aligned. EU workspace data is stored in Frankfurt, Germany. GDPR and EU data Built in Vienna, Austria. A product of influence solutions fm GmbH (FN 675374y). Imprint 14-day trial. Card required, and nothing is charged until day 14. 30-day refund after subscribing. See pricing