2026
2026
07
07
Onboarding
Onboarding
Context
Context
What onboarding looks like when the product learns.
What onboarding looks like when the product learns.
What onboarding looks like when the product learns.
There is a moment early in every AI product when the user decides whether to stay.
There is a moment early in every AI product when the user decides whether to stay.
There is a moment early in every AI product when the user decides whether to stay.
It's not the moment they sign up. It's not the moment they complete the tutorial, if there is one. It's the moment when the product demonstrates, for the first time, that it understood something about them. That it received not just their input, but something of who they are and what they're trying to do.
It's not the moment they sign up. It's not the moment they complete the tutorial, if there is one. It's the moment when the product demonstrates, for the first time, that it understood something about them. That it received not just their input, but something of who they are and what they're trying to do.
It's not the moment they sign up. It's not the moment they complete the tutorial, if there is one. It's the moment when the product demonstrates, for the first time, that it understood something about them. That it received not just their input, but something of who they are and what they're trying to do.
Most AI products never reach that moment in onboarding. They're too busy explaining features.
Most AI products never reach that moment in onboarding. They're too busy explaining features.
Most AI products never reach that moment in onboarding. They're too busy explaining features.
What classical onboarding assumes
What classical onboarding assumes
What classical onboarding assumes
Traditional onboarding is built on a straightforward premise: the user doesn't know how to use the product, and the product's job is to teach them. Show them where things are. Walk them through the key actions. Demonstrate the value proposition quickly enough that they don't leave before they've experienced it.
Traditional onboarding is built on a straightforward premise: the user doesn't know how to use the product, and the product's job is to teach them. Show them where things are. Walk them through the key actions. Demonstrate the value proposition quickly enough that they don't leave before they've experienced it.
Traditional onboarding is built on a straightforward premise: the user doesn't know how to use the product, and the product's job is to teach them. Show them where things are. Walk them through the key actions. Demonstrate the value proposition quickly enough that they don't leave before they've experienced it.
This model works for products that are static. The product is the same on day one as it is on day one hundred. The user changes, they learn, they develop habits, but the product itself doesn't adapt to them. Onboarding is a one-time transfer of knowledge from the product to the user.
This model works for products that are static. The product is the same on day one as it is on day one hundred. The user changes, they learn, they develop habits, but the product itself doesn't adapt to them. Onboarding is a one-time transfer of knowledge from the product to the user.
This model works for products that are static. The product is the same on day one as it is on day one hundred. The user changes, they learn, they develop habits, but the product itself doesn't adapt to them. Onboarding is a one-time transfer of knowledge from the product to the user.
AI products are different in a way that makes this model inadequate. The product is not static. It adapts. It learns. The experience on day one hundred is genuinely different from the experience on day one, not because the user has learned more, but because the system knows more about the user. The relationship develops in both directions.
AI products are different in a way that makes this model inadequate. The product is not static. It adapts. It learns. The experience on day one hundred is genuinely different from the experience on day one, not because the user has learned more, but because the system knows more about the user. The relationship develops in both directions.
AI products are different in a way that makes this model inadequate. The product is not static. It adapts. It learns. The experience on day one hundred is genuinely different from the experience on day one, not because the user has learned more, but because the system knows more about the user. The relationship develops in both directions.
Classical onboarding ignores this entirely. It treats the AI product as if it were a sophisticated but ultimately fixed tool, and it focuses exclusively on teaching the user how to operate it. It says nothing about the other half of what's happening: the system beginning to build its understanding of the person using it.
Classical onboarding ignores this entirely. It treats the AI product as if it were a sophisticated but ultimately fixed tool, and it focuses exclusively on teaching the user how to operate it. It says nothing about the other half of what's happening: the system beginning to build its understanding of the person using it.
Classical onboarding ignores this entirely. It treats the AI product as if it were a sophisticated but ultimately fixed tool, and it focuses exclusively on teaching the user how to operate it. It says nothing about the other half of what's happening: the system beginning to build its understanding of the person using it.
Onboarding a product that learns is not a one-directional knowledge transfer. It's the beginning of a calibration process that will continue for as long as the user stays. Designing it as anything less is designing for the wrong product.
Onboarding a product that learns is not a one-directional knowledge transfer. It's the beginning of a calibration process that will continue for as long as the user stays. Designing it as anything less is designing for the wrong product.
Onboarding a product that learns is not a one-directional knowledge transfer. It's the beginning of a calibration process that will continue for as long as the user stays. Designing it as anything less is designing for the wrong product.
Mutual calibration: the user learns the product, the product learns the user
Mutual calibration: the user learns the product, the product learns the user
Mutual calibration: the user learns the product, the product learns the user
In the first interactions with an AI product, two things are happening simultaneously. The user is forming their mental model of what the system can do, how it behaves, and how to work with it effectively. And the system is accumulating the signals it needs to begin personalizing the experience: preferences expressed explicitly, behaviors that reveal working styles, inputs that indicate expertise level and domain knowledge.
In the first interactions with an AI product, two things are happening simultaneously. The user is forming their mental model of what the system can do, how it behaves, and how to work with it effectively. And the system is accumulating the signals it needs to begin personalizing the experience: preferences expressed explicitly, behaviors that reveal working styles, inputs that indicate expertise level and domain knowledge.
In the first interactions with an AI product, two things are happening simultaneously. The user is forming their mental model of what the system can do, how it behaves, and how to work with it effectively. And the system is accumulating the signals it needs to begin personalizing the experience: preferences expressed explicitly, behaviors that reveal working styles, inputs that indicate expertise level and domain knowledge.
Most onboarding designs capture only one side of this. They focus entirely on the user's learning curve and treat the system's learning as a backend process that happens invisibly.
Most onboarding designs capture only one side of this. They focus entirely on the user's learning curve and treat the system's learning as a backend process that happens invisibly.
Most onboarding designs capture only one side of this. They focus entirely on the user's learning curve and treat the system's learning as a backend process that happens invisibly.
That invisibility is a missed opportunity. When users understand that the product is learning about them, and when they can see that learning happening in ways that benefit them, their relationship with the product changes. They're not just users of a tool. They're participants in a process that improves over time because of their involvement.
That invisibility is a missed opportunity. When users understand that the product is learning about them, and when they can see that learning happening in ways that benefit them, their relationship with the product changes. They're not just users of a tool. They're participants in a process that improves over time because of their involvement.
That invisibility is a missed opportunity. When users understand that the product is learning about them, and when they can see that learning happening in ways that benefit them, their relationship with the product changes. They're not just users of a tool. They're participants in a process that improves over time because of their involvement.
This reframe has real design implications. It means onboarding should include moments that make the system's learning visible, not as a technical feature to be explained, but as a value to be demonstrated. It means the early interactions should be designed to elicit the signals the system needs, not just to showcase the features the team is proud of. It means the definition of onboarding success shifts from "user completed the tutorial" to "system has enough context to begin personalizing meaningfully."
This reframe has real design implications. It means onboarding should include moments that make the system's learning visible, not as a technical feature to be explained, but as a value to be demonstrated. It means the early interactions should be designed to elicit the signals the system needs, not just to showcase the features the team is proud of. It means the definition of onboarding success shifts from "user completed the tutorial" to "system has enough context to begin personalizing meaningfully."
This reframe has real design implications. It means onboarding should include moments that make the system's learning visible, not as a technical feature to be explained, but as a value to be demonstrated. It means the early interactions should be designed to elicit the signals the system needs, not just to showcase the features the team is proud of. It means the definition of onboarding success shifts from "user completed the tutorial" to "system has enough context to begin personalizing meaningfully."
The best onboarding for an AI product doesn't end when the user has learned the product. It ends when the product has learned enough about the user to start being genuinely useful to them specifically.
The best onboarding for an AI product doesn't end when the user has learned the product. It ends when the product has learned enough about the user to start being genuinely useful to them specifically.
The best onboarding for an AI product doesn't end when the user has learned the product. It ends when the product has learned enough about the user to start being genuinely useful to them specifically.
Designing a first conversation, not a tunnel
Designing a first conversation, not a tunnel
Designing a first conversation, not a tunnel
The metaphor of the onboarding tunnel has shaped how design teams think about new user experiences for years. A tunnel has an entrance and an exit. It has a defined path. Progress is linear. The job is to move the user through it as efficiently as possible.
The metaphor of the onboarding tunnel has shaped how design teams think about new user experiences for years. A tunnel has an entrance and an exit. It has a defined path. Progress is linear. The job is to move the user through it as efficiently as possible.
The metaphor of the onboarding tunnel has shaped how design teams think about new user experiences for years. A tunnel has an entrance and an exit. It has a defined path. Progress is linear. The job is to move the user through it as efficiently as possible.
This metaphor produces bad AI onboarding. A tunnel assumes the destination is fixed and the same for every user. In an AI product, the destination is a personalized experience that looks different for different users. The path to that destination is not linear. It's iterative, responsive, shaped by what the system learns along the way.
This metaphor produces bad AI onboarding. A tunnel assumes the destination is fixed and the same for every user. In an AI product, the destination is a personalized experience that looks different for different users. The path to that destination is not linear. It's iterative, responsive, shaped by what the system learns along the way.
This metaphor produces bad AI onboarding. A tunnel assumes the destination is fixed and the same for every user. In an AI product, the destination is a personalized experience that looks different for different users. The path to that destination is not linear. It's iterative, responsive, shaped by what the system learns along the way.
The better metaphor is a first conversation. A conversation has a direction but not a script. It responds to what's said. It produces understanding that neither party had at the start. It feels like progress even when it's exploratory. And it ends not at a predetermined exit point, but at a natural moment when both parties have enough to move forward.
The better metaphor is a first conversation. A conversation has a direction but not a script. It responds to what's said. It produces understanding that neither party had at the start. It feels like progress even when it's exploratory. And it ends not at a predetermined exit point, but at a natural moment when both parties have enough to move forward.
The better metaphor is a first conversation. A conversation has a direction but not a script. It responds to what's said. It produces understanding that neither party had at the start. It feels like progress even when it's exploratory. And it ends not at a predetermined exit point, but at a natural moment when both parties have enough to move forward.
Designing a first conversation means creating onboarding flows that genuinely respond to the user's inputs instead of following a fixed path regardless of what the user does. It means building in moments of reflection, where the system demonstrates that it has heard something. It means allowing the user to move at their own pace and in their own direction, while still ensuring the system collects what it needs to be useful.
Designing a first conversation means creating onboarding flows that genuinely respond to the user's inputs instead of following a fixed path regardless of what the user does. It means building in moments of reflection, where the system demonstrates that it has heard something. It means allowing the user to move at their own pace and in their own direction, while still ensuring the system collects what it needs to be useful.
Designing a first conversation means creating onboarding flows that genuinely respond to the user's inputs instead of following a fixed path regardless of what the user does. It means building in moments of reflection, where the system demonstrates that it has heard something. It means allowing the user to move at their own pace and in their own direction, while still ensuring the system collects what it needs to be useful.
A tunnel delivers users to a destination. A conversation begins a relationship. AI products need the second, and most of them have built the first.
A tunnel delivers users to a destination. A conversation begins a relationship. AI products need the second, and most of them have built the first.
A tunnel delivers users to a destination. A conversation begins a relationship. AI products need the second, and most of them have built the first.
Showing the system learning without being unsettling
Showing the system learning without being unsettling
Showing the system learning without being unsettling
There is a real risk in making the system's learning visible: it can feel like surveillance. The user who notices that the product has been paying close attention to their behavior, and who hasn't explicitly consented to that attention, doesn't feel served. They feel watched.
There is a real risk in making the system's learning visible: it can feel like surveillance. The user who notices that the product has been paying close attention to their behavior, and who hasn't explicitly consented to that attention, doesn't feel served. They feel watched.
There is a real risk in making the system's learning visible: it can feel like surveillance. The user who notices that the product has been paying close attention to their behavior, and who hasn't explicitly consented to that attention, doesn't feel served. They feel watched.
The design challenge is making the learning feel collaborative rather than observational. The difference is agency. When the user feels they are an active participant in the system's learning, contributing deliberately to the model it's building of them, the experience feels empowering. When the system's learning feels like something happening to them without their participation, it feels intrusive.
The design challenge is making the learning feel collaborative rather than observational. The difference is agency. When the user feels they are an active participant in the system's learning, contributing deliberately to the model it's building of them, the experience feels empowering. When the system's learning feels like something happening to them without their participation, it feels intrusive.
The design challenge is making the learning feel collaborative rather than observational. The difference is agency. When the user feels they are an active participant in the system's learning, contributing deliberately to the model it's building of them, the experience feels empowering. When the system's learning feels like something happening to them without their participation, it feels intrusive.
This means giving users explicit opportunities to teach the system. Not just passive data collection, but active moments where the user can express preferences, correct misunderstandings, or tell the system something about themselves that it couldn't have inferred. These moments serve two purposes. They give the system better signal. And they give the user a sense of authorship over their own experience.
This means giving users explicit opportunities to teach the system. Not just passive data collection, but active moments where the user can express preferences, correct misunderstandings, or tell the system something about themselves that it couldn't have inferred. These moments serve two purposes. They give the system better signal. And they give the user a sense of authorship over their own experience.
This means giving users explicit opportunities to teach the system. Not just passive data collection, but active moments where the user can express preferences, correct misunderstandings, or tell the system something about themselves that it couldn't have inferred. These moments serve two purposes. They give the system better signal. And they give the user a sense of authorship over their own experience.
It also means surfacing the learning at the right moments and in the right tone. Not as a technical announcement: "your preferences have been updated." As a natural demonstration of understanding: the system responding in a way that shows it has internalized something the user cared about, without drawing explicit attention to the mechanism behind it.
It also means surfacing the learning at the right moments and in the right tone. Not as a technical announcement: "your preferences have been updated." As a natural demonstration of understanding: the system responding in a way that shows it has internalized something the user cared about, without drawing explicit attention to the mechanism behind it.
It also means surfacing the learning at the right moments and in the right tone. Not as a technical announcement: "your preferences have been updated." As a natural demonstration of understanding: the system responding in a way that shows it has internalized something the user cared about, without drawing explicit attention to the mechanism behind it.
The system that learns visibly but quietly, that demonstrates understanding without announcing it, earns something that no onboarding tutorial can produce: the user's genuine belief that the product gets them.
The system that learns visibly but quietly, that demonstrates understanding without announcing it, earns something that no onboarding tutorial can produce: the user's genuine belief that the product gets them.
The system that learns visibly but quietly, that demonstrates understanding without announcing it, earns something that no onboarding tutorial can produce: the user's genuine belief that the product gets them.
The first moment of recognition
The first moment of recognition
The first moment of recognition
There is a specific moment in AI product onboarding that determines, more than any other, whether the user will stay.
There is a specific moment in AI product onboarding that determines, more than any other, whether the user will stay.
There is a specific moment in AI product onboarding that determines, more than any other, whether the user will stay.
It's the first time the product does something that makes the user think: it understood me. Not just my words. Me.
It's the first time the product does something that makes the user think: it understood me. Not just my words. Me.
It's the first time the product does something that makes the user think: it understood me. Not just my words. Me.
This moment is not accidental. It has to be designed. It requires knowing, from the signals available in the first few interactions, when the system has accumulated enough context to produce a response that will feel genuinely personal. And it requires designing the interface around that moment so that when it happens, it lands.
This moment is not accidental. It has to be designed. It requires knowing, from the signals available in the first few interactions, when the system has accumulated enough context to produce a response that will feel genuinely personal. And it requires designing the interface around that moment so that when it happens, it lands.
This moment is not accidental. It has to be designed. It requires knowing, from the signals available in the first few interactions, when the system has accumulated enough context to produce a response that will feel genuinely personal. And it requires designing the interface around that moment so that when it happens, it lands.
What it looks like varies by product. It might be a recommendation that maps precisely to something the user mentioned in passing. A response that uses the same framing the user used, without mimicking them. An anticipation of a need the user hadn't yet expressed but was clearly building toward. A correction that shows the system noticed a pattern the user themselves hadn't consciously identified.
What it looks like varies by product. It might be a recommendation that maps precisely to something the user mentioned in passing. A response that uses the same framing the user used, without mimicking them. An anticipation of a need the user hadn't yet expressed but was clearly building toward. A correction that shows the system noticed a pattern the user themselves hadn't consciously identified.
What it looks like varies by product. It might be a recommendation that maps precisely to something the user mentioned in passing. A response that uses the same framing the user used, without mimicking them. An anticipation of a need the user hadn't yet expressed but was clearly building toward. A correction that shows the system noticed a pattern the user themselves hadn't consciously identified.
The specifics matter less than the effect. The user feels seen. Not surveilled. Seen. Understood in a way that makes the product feel less like a tool and more like a collaborator.
The specifics matter less than the effect. The user feels seen. Not surveilled. Seen. Understood in a way that makes the product feel less like a tool and more like a collaborator.
The specifics matter less than the effect. The user feels seen. Not surveilled. Seen. Understood in a way that makes the product feel less like a tool and more like a collaborator.
That moment is the real end of onboarding. Everything before it is preamble. Everything after it is a relationship.
That moment is the real end of onboarding. Everything before it is preamble. Everything after it is a relationship.
That moment is the real end of onboarding. Everything before it is preamble. Everything after it is a relationship.
Designing the progression, not just the entry
Designing the progression, not just the entry
Designing the progression, not just the entry
Most onboarding design stops at activation. The user has completed the setup. They've experienced the core value. They're in the product. Job done.
Most onboarding design stops at activation. The user has completed the setup. They've experienced the core value. They're in the product. Job done.
Most onboarding design stops at activation. The user has completed the setup. They've experienced the core value. They're in the product. Job done.
For AI products, this is where the design work actually begins.
For AI products, this is where the design work actually begins.
For AI products, this is where the design work actually begins.
The relationship between a user and an AI product develops over time in ways that need to be designed as deliberately as the first session. The system's model of the user becomes more accurate. The user's model of the system becomes more sophisticated. The interactions that were effortful in the first week become natural in the second month. New capabilities become accessible as the user's fluency grows.
The relationship between a user and an AI product develops over time in ways that need to be designed as deliberately as the first session. The system's model of the user becomes more accurate. The user's model of the system becomes more sophisticated. The interactions that were effortful in the first week become natural in the second month. New capabilities become accessible as the user's fluency grows.
The relationship between a user and an AI product develops over time in ways that need to be designed as deliberately as the first session. The system's model of the user becomes more accurate. The user's model of the system becomes more sophisticated. The interactions that were effortful in the first week become natural in the second month. New capabilities become accessible as the user's fluency grows.
This progression doesn't happen automatically just because the AI is learning. It has to be shaped. The designer has to think about what the experience looks like at week one, month one, month six. What changes. What the system surfaces at each stage. How the interface evolves to reflect the depth of the relationship that has developed.
This progression doesn't happen automatically just because the AI is learning. It has to be shaped. The designer has to think about what the experience looks like at week one, month one, month six. What changes. What the system surfaces at each stage. How the interface evolves to reflect the depth of the relationship that has developed.
This progression doesn't happen automatically just because the AI is learning. It has to be shaped. The designer has to think about what the experience looks like at week one, month one, month six. What changes. What the system surfaces at each stage. How the interface evolves to reflect the depth of the relationship that has developed.
This means designing moments of progression that make the development of the relationship visible. Small signals that show the user the product has grown with them. New capabilities unlocked not by completing a tutorial, but by the system recognizing that the user is ready for them. A product that feels, over time, like it was made for this specific person.
This means designing moments of progression that make the development of the relationship visible. Small signals that show the user the product has grown with them. New capabilities unlocked not by completing a tutorial, but by the system recognizing that the user is ready for them. A product that feels, over time, like it was made for this specific person.
This means designing moments of progression that make the development of the relationship visible. Small signals that show the user the product has grown with them. New capabilities unlocked not by completing a tutorial, but by the system recognizing that the user is ready for them. A product that feels, over time, like it was made for this specific person.
Onboarding is not the beginning of the user's experience with the product. It's the beginning of the product's experience with the user. Design it as if the relationship it starts will last for years. Because for the users who matter most, it will.
Onboarding is not the beginning of the user's experience with the product. It's the beginning of the product's experience with the user. Design it as if the relationship it starts will last for years. Because for the users who matter most, it will.
Onboarding is not the beginning of the user's experience with the product. It's the beginning of the product's experience with the user. Design it as if the relationship it starts will last for years. Because for the users who matter most, it will.
Raphaël D. - Head of Product Design, designing at the intersection of AI infrastructure and human experience.
Raphaël D. - Head of Product Design, designing at the intersection of AI infrastructure and human experience.
Raphaël D. - Head of Product Design, designing at the intersection of AI infrastructure and human experience.
Design for AI
Thinking through the design problems that AI products create.
Not how to use AI as a designer. How to design for it.
© 2026 Design for AI. All rights reserved.
Design for AI
Thinking through the design problems that AI products create. Not how to use AI as a designer. How to design for it.
© 2026 Design for AI. All rights reserved.
Design for AI
Thinking through the design problems that AI products create. Not how to use AI as a designer. How to design for it.
© 2026 Design for AI. All rights reserved.