Route · Public Design Atlas
Design a product with AI
To design a product or service that uses artificial intelligence, from understanding people to testing it carefully.
Understand what they need
Start with the person, not the technology.
In-depth interview
A long, open, guided conversation with a single person to understand in detail their experiences, motivations and way of seeing a topic.
How it’s used
How it’s done
- Prepare a guide of topics, not a closed questionnaire.
- Have a conversation following what the person tells you.
- Record and analyze looking for themes and quotes.
- Output
- Detailed accounts of experiences and motivations
- Avoid
- Asking for general opinions instead of concrete stories (“the last time you…”), asking leading questions or selling your idea.
Jobs to be done
Understanding customers by the progress they are trying to make in a specific situation, the “job” they hire a product for, not by their demographics.
How it’s used
How it’s done
- Look at the situation. When, where and with whom the product is chosen.
- Find the job. What progress is sought and what alternatives are used.
- Design for that job. Improve what matters in that situation.
- Avoid
- Confusing the job with a product feature.
Personas
A method for summarizing research in a few fictional characters, with a name, goals and context, who represent groups of users.
How it’s used
How it’s done
- Research with real users.
- Group similar goals and behaviors.
- Each group becomes a character.
- Output
- A few fictional characters, with a name, goals and context
- Avoid
- Making them up in an afternoon or filling them with stereotypes and details that don't help decide.
Mental model
The often incomplete idea a person forms of how something works, and uses to predict what will happen when they act.
How it’s used
Why it happens
- We can't see how a system works on the inside.
- We form an idea from what we see and what we already know.
- If the idea doesn't match the system, we make mistakes.
- Avoid
- Assuming people understand the system the way the team does.
Decide whether AI adds anything
Not everything needs artificial intelligence.
Three criteria an idea must balance: it meets what people need and want, it can be built with the available means, and the organization can sustain it.
How it’s used
How it works
- Assess each idea against the three criteria.
- Spot which of the three raises the most doubts.
- Test that doubt first, with a prototype or with data.
- Avoid
- Using it as a quick vote, without data.
How might we…?
A technique for turning a problem or an insight into an open, optimistic question that invites ideas: “How might we…?”
How it’s used
How it’s done
- You start from an insight or a specific problem.
- You reframe it as a question neither too broad nor too narrow.
- You use it as the starting point for ideation.
- Output
- Open questions that invite ideas
- Avoid
- Putting the solution inside the question (“How might we build an app that…?”).
Value Proposition Canvas
A template to see whether an offer fits a customer type: on one side, their profile (jobs, pains and gains); on the other, the offer’s value map.
How it’s used
How it’s done
- Describe the profile of a customer type: jobs, pains and gains.
- Fill in the value map with products and services, pain relievers and gain creators.
- Test the fit between both parts with real customers.
- Output
- A canvas with the customer profile and the value map
- Avoid
- Filling the profile with assumptions and taking it as valid.
Define how it behaves
What it does, how it explains itself, and what happens when it gets something wrong.
Eighteen guidelines for designing interaction with AI systems, across four moments: initially, during interaction, when the system is wrong and over time.
How it’s used
How it works
- The guidelines are grouped into four moments of the interaction.
- At each moment, the team checks whether its product follows them.
- Whatever falls short is fixed before launch.
- Avoid
- Applying them only at the end, when nothing can be changed.
Conversational design
Designs how a person converses with a system, by voice or in writing, as in voice assistants and chatbots, drawing on how conversations between people are organized.
How it’s used
- When to use it
- Write sample dialogues and read them aloud before designing the flow.
- Avoid
- Designing a menu disguised as a conversation.
Storyboard
A sequence of panels that shows a person using a product or a service in context, step by step.
How it’s used
How it’s done
- Choose the situation and the person.
- Draw the steps in a few panels.
- Share it and gather feedback.
- Output
- A sequence of panels showing use, step by step
- Avoid
- Drawing only screens instead of the situation.
Prototype it before building it
Test the experience without programming the AI.
Wizard of Oz
A prototype in which a hidden person, the “wizard,” produces the system's responses by hand while the participant uses it as if it worked on its own.
How it’s used
How it’s done
- Prepare the tasks, the wizard's script, and a place where the wizard can't be seen.
- The person tries it and the wizard responds live.
- Adjust the script and test again.
- Output
- People's reactions to a simulated system and what the real system needs to do
- Avoid
- Letting the wizard improvise: if the answers change every time, you won't know what failed.
Prototyping
Building quick, incomplete versions of an idea, from a paper sketch or a cardboard mock-up to something that works, to answer a design question before developing the idea.
How it’s used
How it’s done
- Choose the question: whether it works for people, how it looks or whether it can be built.
- Build the minimum that answers it: paper, cardboard, mock-up, wireframe or code.
- Test it and decide what to change.
- Output
- Prototypes, from a sketch to something that works
- Avoid
- Polishing the look when you don't yet know if the idea works.
Role playing
The team, and sometimes users, act out a situation with simple or cardboard objects to imagine how a product or service would work.
How it’s used
How it’s done
- Prepare the scene and the roles.
- Act out the situation with the imagined object.
- Discuss what worked and what didn't.
- Output
- Acted-out situations that show how an idea would work
- Avoid
- Turning it into theater without a design question.
Test it and take care of people
That it works and doesn’t manipulate.
Usability testing
A method in which real people try to do specific tasks with a product while being observed, to find out where they get stuck and why.
How it’s used
How it’s done
- Prepare realistic tasks, without saying how to do them.
- Observe each person as they try and think aloud.
- Note where they get stuck and decide what to fix first.
- Output
- A list of usability problems, with where they appear and why
- Avoid
- Asking “do you like it?”: measure what people do, not what they think.
Ten general rules that evaluators use to review an interface and spot usability problems before testing it with users.
How it’s used
How it works
- Several people review the interface separately.
- Each problem is noted next to the heuristic it breaks.
- The lists are combined and the problems are ranked by severity.
- Avoid
- Letting a single person review the interface and call it done.
Interface designs that lead people to do something they didn't want to, like buying, subscribing or giving up data, by exploiting how we decide.
How it’s used
Why it happens
- They exploit mental shortcuts such as haste, inertia or social pressure.
- They hide or make costlier the option that best serves the person.
- They work, which is why they have spread so widely.
- Avoid
- Hiding cancellation, preselecting extras or using guilt-tripping wording.
HEART framework
A framework for choosing what to measure in a digital product (happiness, engagement, adoption, retention and task success), starting from goals rather than available data.
How it’s used
How it works
- Goals what the product should achieve.
- Signals what behavior or response would show it.
- Metrics how to count it to track it over time.
- Avoid
- Always using all five categories; choose the ones that make sense for your product.