The design sprint was built to replace weeks of debate with five days of structured work and a real test. AI tools don't change that goal, but they do change how long some of the steps take. This guide walks through the publicly documented Google Ventures sprint, from the book Sprint by Jake Knapp, John Zeratsky and Braden Kowitz, one day at a time, and shows where AI genuinely helps and where it quietly makes the sprint worse.
How do you run a design sprint with AI?
You run a design sprint with AI by keeping the five-day Google Ventures structure intact and using AI for the slow, mechanical parts: summarizing research, gathering references, building the prototype and processing interview notes. The decisions, the sketches and the user test stay with people.
That split matters because the sprint's value comes from two things AI can't supply: a team committing to one direction, and real customers reacting to it. Everything else is preparation, and preparation is where AI saves time. If you're new to the method itself, start with our guide to running a design sprint for a startup, and see how sprints fit with other approaches in product prototyping.
Day | GV sprint activity | Where AI helps | Where people lead |
|---|---|---|---|
Monday | Long-term goal, map, ask the experts, pick a target | Summarizing tickets, reviews and past research | Choosing the goal and target |
Tuesday | Lightning demos, four-step sketch | Gathering references for demos | Sketching alone, on paper |
Wednesday | Art museum, heat map, supervote, storyboard | Transcribing notes, tidying the storyboard | The decider's call |
Thursday | Build a realistic facade prototype | Generating screens, copy and sample content | Keeping it to the storyboard |
Friday | Five one-on-one user interviews | Transcripts and first-pass summaries | Running and watching the interviews |
Where does AI help on Monday's map?
On Monday, AI helps most before the sprint starts: summarizing support tickets, sales call notes, reviews and past research into a short briefing the team reads on the first morning. That gives the map and the expert interviews a better starting point.
The day itself stays human. The team agrees the long-term goal, writes the sprint questions, draws a simple map of the customer journey and interviews the people inside the company who know customers best, capturing the problems as How Might We notes. AI can cluster those notes afterward, but the choice of target, the one moment on the map the week will focus on, has to be made by the team and the decider. A tidy AI summary is easy to agree with, which is exactly why it shouldn't pick the target for you.
How can AI speed up Tuesday's sketches without causing groupthink?
Use AI to gather material for the lightning demos, then put it away while everyone sketches alone. The four-step sketch, notes, ideas, Crazy 8s and a detailed solution sketch, works because each person thinks independently before seeing anyone else's ideas.
AI is good at finding examples of how other products solve a similar problem, which shortens the research behind lightning demos. It's a poor sketching partner. Ask a model for ideas and the whole team gets variations of the same, most common answer, which removes the range the sketch session exists to produce. If someone wants AI help, let them use it after their own Crazy 8s, to push one idea further. Our shortlist of AI design tools covers which tools are useful for research and which for drawing.
Why should the Wednesday decision stay human?
The Wednesday decision should stay human because it's a commitment, not an analysis: the decider chooses the direction the company will test and stand behind. AI can support the process around the decision, but it can't own the outcome.
The GV structure is deliberate here. Solution sketches go up in an art museum, the team marks interesting parts with dots for a heat map, runs a speed critique and a straw poll, and then the decider makes the supervote. After that, the team turns the winning sketch into a storyboard of around a dozen frames. AI can transcribe the critique, list the dotted ideas, and tidy the storyboard into a shot list for Thursday. Asking a model which sketch is best is a way of avoiding the decision, and the sprint stalls the same way it does when the decider skips the session.
How do AI tools change Thursday's prototype?
Thursday changes the most: AI design and code tools can turn a storyboard into a realistic, clickable facade in hours, with believable copy and sample data. That frees time to polish the parts the test depends on.
The GV rule still applies: build a facade, real enough that participants react honestly, and only along the storyboard's path. AI makes it tempting to build more, because extra screens feel free. They aren't, because every extra screen is something the team has to check before Friday. Choose the medium by the question. A clickable Figma prototype suits most flow and messaging questions; a thin code prototype from an AI builder is more honest when the test depends on real data or real model output. Our comparison of Figma prototype vs code prototype helps you decide, and if you're testing an AI feature itself, prototyping AI features covers Wizard of Oz setups and real models.
For example, suppose a startup sprints on an AI onboarding flow
Hypothetical example: a seed-stage HR software startup wants to know whether new managers will trust an AI assistant to draft their first team goals. The sprint question: will a new manager accept a drafted set of goals after answering five short questions?
Monday, a pre-sprint AI summary of support tickets shows that managers often ask what good goals look like. The team targets the moment after the manager's first login. Tuesday, everyone sketches alone; one sketch shows the draft side by side with the manager's answers. Wednesday, the founder picks it.
Thursday, the designer builds the screens in Figma and uses a model to write realistic sample goals for three made-up teams, then checks every line by hand. Friday, five managers from the target market try it. The team watches each session live and uses AI only to produce first-pass transcripts. This walkthrough is illustrative; it shows a process, not a real company or measured result.
Can AI replace Friday's user interviews?
No. AI can't replace Friday's interviews, because the point of the test is to see how real people from your target market react, and a simulated user can only echo patterns from its training data. Use AI for the paperwork around the interviews, not for the interviews themselves.
The GV format runs five one-on-one interviews, with the rest of the team watching and taking notes, and looks for patterns across them. AI helps with recruiting messages, transcripts, and a first-pass summary of what each participant said. It doesn't help with the moment that matters: a participant hesitating, misreading a label or saying they'd never trust it. Watch those live. Then write the patterns yourselves and use the AI summary only to check you didn't miss something.
Should you run an AI-assisted sprint yourself or bring in help?
Run it yourself if you have a decider who can commit the time and someone who can build a convincing prototype; bring in help when the prototype is the gap. AI tools lower the bar for building, but they don't supply the design judgment that makes a facade believable.
- Run it in-house when your team includes a designer, the founder can attend the map and decide days, and test participants are already booked.
- Bring in help for the design work when nobody can turn a storyboard into a credible prototype, or when you want several directions explored before choosing. Our product design and MVP design work covers that part.
- Skip the sprint when the question is small enough to ship and measure in a day.
Whichever route you take, the prototype should look like the product you mean to build. If the sprint is preparing a demo for investors rather than a test with users, read our guide to investor demo prototypes first, because the two have different jobs.
