2025
2025
05
05
Agency
Agency
Trust
Trust
Designing the handoff between human and AI.
Designing the handoff between human and AI.
Designing the handoff between human and AI.
There is a moment in every AI product that almost nobody designs deliberately. It happens in seconds. It's rarely documented in a flow. It doesn't have a name in most design systems.
There is a moment in every AI product that almost nobody designs deliberately. It happens in seconds. It's rarely documented in a flow. It doesn't have a name in most design systems.
There is a moment in every AI product that almost nobody designs deliberately. It happens in seconds. It's rarely documented in a flow. It doesn't have a name in most design systems.
It's the moment when control changes hands. When the user stops driving and the system takes over. Or when the system has done what it can do and hands the result back to a human being who now has to decide what to do with it.
It's the moment when control changes hands. When the user stops driving and the system takes over. Or when the system has done what it can do and hands the result back to a human being who now has to decide what to do with it.
It's the moment when control changes hands. When the user stops driving and the system takes over. Or when the system has done what it can do and hands the result back to a human being who now has to decide what to do with it.
That moment is where most AI products lose their users. Not dramatically. Quietly. In a accumulation of small discomforts that compound until the product feels unreliable, opaque, or simply not worth the effort.
That moment is where most AI products lose their users. Not dramatically. Quietly. In a accumulation of small discomforts that compound until the product feels unreliable, opaque, or simply not worth the effort.
That moment is where most AI products lose their users. Not dramatically. Quietly. In a accumulation of small discomforts that compound until the product feels unreliable, opaque, or simply not worth the effort.
What a handoff actually is
What a handoff actually is
What a handoff actually is
In traditional software, transitions are mechanical. You submit a form. The system processes it. You see the result. The user is always in control of what happens next, because the system only acts when explicitly told to.
In traditional software, transitions are mechanical. You submit a form. The system processes it. You see the result. The user is always in control of what happens next, because the system only acts when explicitly told to.
In traditional software, transitions are mechanical. You submit a form. The system processes it. You see the result. The user is always in control of what happens next, because the system only acts when explicitly told to.
AI products don't work that way. The system has agency. It interprets, decides, generates, and sometimes acts, based on inputs the user provided without fully understanding how they would be used. The line between "the user is doing something" and "the system is doing something" is blurry, shifting, and often invisible.
AI products don't work that way. The system has agency. It interprets, decides, generates, and sometimes acts, based on inputs the user provided without fully understanding how they would be used. The line between "the user is doing something" and "the system is doing something" is blurry, shifting, and often invisible.
AI products don't work that way. The system has agency. It interprets, decides, generates, and sometimes acts, based on inputs the user provided without fully understanding how they would be used. The line between "the user is doing something" and "the system is doing something" is blurry, shifting, and often invisible.
A handoff is any moment where that line moves. Where the locus of action transfers from the user to the system, or from the system back to the user. Where one party stops and the other begins.
A handoff is any moment where that line moves. Where the locus of action transfers from the user to the system, or from the system back to the user. Where one party stops and the other begins.
A handoff is any moment where that line moves. Where the locus of action transfers from the user to the system, or from the system back to the user. Where one party stops and the other begins.
These moments happen constantly in AI products. They're just not being designed.
These moments happen constantly in AI products. They're just not being designed.
These moments happen constantly in AI products. They're just not being designed.
A handoff is not a loading state. It's not a transition animation. It's a transfer of agency between a human being and a system. Treating it as anything less is what makes it go wrong.
A handoff is not a loading state. It's not a transition animation. It's a transfer of agency between a human being and a system. Treating it as anything less is what makes it go wrong.
A handoff is not a loading state. It's not a transition animation. It's a transfer of agency between a human being and a system. Treating it as anything less is what makes it go wrong.
The two directions, and why they're different problems
The two directions, and why they're different problems
The two directions, and why they're different problems
Handoffs run in two directions, and they create fundamentally different design challenges.
Handoffs run in two directions, and they create fundamentally different design challenges.
Handoffs run in two directions, and they create fundamentally different design challenges.
When the user hands off to the AI, the challenge is letting go. The user has been in control. They've been making decisions, directing the flow, holding the context in their head. Now they're asking the system to take over, and they have to trust that it understood what they meant, not just what they said. The design challenge is making that transfer feel safe. Not by hiding what the system is doing, but by giving the user enough signal to feel that the system has received the right instructions and is working on the right problem.
When the user hands off to the AI, the challenge is letting go. The user has been in control. They've been making decisions, directing the flow, holding the context in their head. Now they're asking the system to take over, and they have to trust that it understood what they meant, not just what they said. The design challenge is making that transfer feel safe. Not by hiding what the system is doing, but by giving the user enough signal to feel that the system has received the right instructions and is working on the right problem.
When the user hands off to the AI, the challenge is letting go. The user has been in control. They've been making decisions, directing the flow, holding the context in their head. Now they're asking the system to take over, and they have to trust that it understood what they meant, not just what they said. The design challenge is making that transfer feel safe. Not by hiding what the system is doing, but by giving the user enough signal to feel that the system has received the right instructions and is working on the right problem.
When the AI hands back to the user, the challenge is re-entry. The system has done something. Now the user has to pick up where the AI left off, evaluate what was produced, and decide what happens next. The design challenge is making that re-entry feel natural. The user shouldn't have to reconstruct context from scratch. They shouldn't have to figure out where the system stopped and where they need to start. The interface should orient them immediately.
When the AI hands back to the user, the challenge is re-entry. The system has done something. Now the user has to pick up where the AI left off, evaluate what was produced, and decide what happens next. The design challenge is making that re-entry feel natural. The user shouldn't have to reconstruct context from scratch. They shouldn't have to figure out where the system stopped and where they need to start. The interface should orient them immediately.
When the AI hands back to the user, the challenge is re-entry. The system has done something. Now the user has to pick up where the AI left off, evaluate what was produced, and decide what happens next. The design challenge is making that re-entry feel natural. The user shouldn't have to reconstruct context from scratch. They shouldn't have to figure out where the system stopped and where they need to start. The interface should orient them immediately.
These are different problems because they require different things from the user. Handing off requires trust. Re-entry requires comprehension. Design that handles one well often fails at the other.
These are different problems because they require different things from the user. Handing off requires trust. Re-entry requires comprehension. Design that handles one well often fails at the other.
These are different problems because they require different things from the user. Handing off requires trust. Re-entry requires comprehension. Design that handles one well often fails at the other.
Most AI products design the output carefully and ignore the transitions around it. But users don't experience outputs in isolation. They experience the moment of receiving them, and that moment is as designed as anything else on the screen.
Most AI products design the output carefully and ignore the transitions around it. But users don't experience outputs in isolation. They experience the moment of receiving them, and that moment is as designed as anything else on the screen.
Most AI products design the output carefully and ignore the transitions around it. But users don't experience outputs in isolation. They experience the moment of receiving them, and that moment is as designed as anything else on the screen.
What happens when the handoff is badly designed
What happens when the handoff is badly designed
What happens when the handoff is badly designed
The symptoms are familiar, even if the cause isn't always identified correctly.
The symptoms are familiar, even if the cause isn't always identified correctly.
The symptoms are familiar, even if the cause isn't always identified correctly.
Users who interrupt the system before it finishes, not because they're impatient, but because they're not sure it understood them and they'd rather start over than discover the mistake at the end.
Users who interrupt the system before it finishes, not because they're impatient, but because they're not sure it understood them and they'd rather start over than discover the mistake at the end.
Users who interrupt the system before it finishes, not because they're impatient, but because they're not sure it understood them and they'd rather start over than discover the mistake at the end.
Users who re-read the system's output multiple times before acting on it, not because the output is complex, but because they're not sure how to evaluate it or how much confidence to place in it.
Users who re-read the system's output multiple times before acting on it, not because the output is complex, but because they're not sure how to evaluate it or how much confidence to place in it.
Users who re-read the system's output multiple times before acting on it, not because the output is complex, but because they're not sure how to evaluate it or how much confidence to place in it.
Users who undo the system's actions immediately after they happen, not because the actions were wrong, but because the transition was so abrupt that the user didn't feel they had consented to it.
Users who undo the system's actions immediately after they happen, not because the actions were wrong, but because the transition was so abrupt that the user didn't feel they had consented to it.
Users who undo the system's actions immediately after they happen, not because the actions were wrong, but because the transition was so abrupt that the user didn't feel they had consented to it.
Users who stop using autonomous features entirely and revert to manual control, not because the AI performed badly, but because the handoff made them feel like they had lost control of something they didn't mean to surrender.
Users who stop using autonomous features entirely and revert to manual control, not because the AI performed badly, but because the handoff made them feel like they had lost control of something they didn't mean to surrender.
Users who stop using autonomous features entirely and revert to manual control, not because the AI performed badly, but because the handoff made them feel like they had lost control of something they didn't mean to surrender.
In each case, the problem isn't the AI's capability. It's the transition into and out of it.
In each case, the problem isn't the AI's capability. It's the transition into and out of it.
In each case, the problem isn't the AI's capability. It's the transition into and out of it.
Bad handoffs don't just create friction. They create doubt. And doubt, once established, attaches itself to everything the system does afterward.
Bad handoffs don't just create friction. They create doubt. And doubt, once established, attaches itself to everything the system does afterward.
Bad handoffs don't just create friction. They create doubt. And doubt, once established, attaches itself to everything the system does afterward.
The signals that make a handoff legible
The signals that make a handoff legible
The signals that make a handoff legible
A well-designed handoff communicates three things: that the transfer is happening, what is being transferred, and what comes next.
A well-designed handoff communicates three things: that the transfer is happening, what is being transferred, and what comes next.
A well-designed handoff communicates three things: that the transfer is happening, what is being transferred, and what comes next.
That the transfer is happening means the user knows, without having to infer it, that control has moved. Not through an explicit notification every time, but through signals consistent enough that the user has learned to read them. A change in the interface's state. A shift in the visual focus. A moment of stillness before the system responds. These signals don't have to be loud. They have to be reliable.
That the transfer is happening means the user knows, without having to infer it, that control has moved. Not through an explicit notification every time, but through signals consistent enough that the user has learned to read them. A change in the interface's state. A shift in the visual focus. A moment of stillness before the system responds. These signals don't have to be loud. They have to be reliable.
That the transfer is happening means the user knows, without having to infer it, that control has moved. Not through an explicit notification every time, but through signals consistent enough that the user has learned to read them. A change in the interface's state. A shift in the visual focus. A moment of stillness before the system responds. These signals don't have to be loud. They have to be reliable.
What is being transferred means the user understands the scope of what the system has taken on. Not a technical specification of the task. A human-readable confirmation that the system received the right problem. "I'll draft this based on the three points you highlighted" is a handoff signal. A spinner is not.
What is being transferred means the user understands the scope of what the system has taken on. Not a technical specification of the task. A human-readable confirmation that the system received the right problem. "I'll draft this based on the three points you highlighted" is a handoff signal. A spinner is not.
What is being transferred means the user understands the scope of what the system has taken on. Not a technical specification of the task. A human-readable confirmation that the system received the right problem. "I'll draft this based on the three points you highlighted" is a handoff signal. A spinner is not.
What comes next means the user knows what their role will be when the system hands back. Will they be reviewing? Approving? Editing? Redirecting? The re-entry is smoother when the user has been primed for it during the handoff.
What comes next means the user knows what their role will be when the system hands back. Will they be reviewing? Approving? Editing? Redirecting? The re-entry is smoother when the user has been primed for it during the handoff.
What comes next means the user knows what their role will be when the system hands back. Will they be reviewing? Approving? Editing? Redirecting? The re-entry is smoother when the user has been primed for it during the handoff.
None of these signals require new components or complex interactions. They require intentionality about moments that most design processes treat as infrastructure rather than experience.
None of these signals require new components or complex interactions. They require intentionality about moments that most design processes treat as infrastructure rather than experience.
None of these signals require new components or complex interactions. They require intentionality about moments that most design processes treat as infrastructure rather than experience.
The handoff that can't be undone
The handoff that can't be undone
The handoff that can't be undone
There is a category of handoff that deserves its own treatment: the moment when the AI takes an action that the user cannot reverse.
There is a category of handoff that deserves its own treatment: the moment when the AI takes an action that the user cannot reverse.
There is a category of handoff that deserves its own treatment: the moment when the AI takes an action that the user cannot reverse.
Sends an email. Submits a form. Executes a transaction. Deletes a file. These are not hypothetical scenarios in AI products. They are increasingly common as systems become more agentic, more capable of acting in the world on the user's behalf.
Sends an email. Submits a form. Executes a transaction. Deletes a file. These are not hypothetical scenarios in AI products. They are increasingly common as systems become more agentic, more capable of acting in the world on the user's behalf.
Sends an email. Submits a form. Executes a transaction. Deletes a file. These are not hypothetical scenarios in AI products. They are increasingly common as systems become more agentic, more capable of acting in the world on the user's behalf.
The design stakes at these moments are different from any other handoff. The user doesn't just need to understand that control is transferring. They need to understand exactly what is about to happen, why, and that they have a genuine opportunity to stop it.
The design stakes at these moments are different from any other handoff. The user doesn't just need to understand that control is transferring. They need to understand exactly what is about to happen, why, and that they have a genuine opportunity to stop it.
The design stakes at these moments are different from any other handoff. The user doesn't just need to understand that control is transferring. They need to understand exactly what is about to happen, why, and that they have a genuine opportunity to stop it.
This is not about adding a confirmation dialog to every action. Confirmation fatigue is real, and it produces users who click through confirmations without reading them. It's about designing the moments before the irreversible action in a way that ensures the user arrives at it fully informed, not just technically warned.
This is not about adding a confirmation dialog to every action. Confirmation fatigue is real, and it produces users who click through confirmations without reading them. It's about designing the moments before the irreversible action in a way that ensures the user arrives at it fully informed, not just technically warned.
This is not about adding a confirmation dialog to every action. Confirmation fatigue is real, and it produces users who click through confirmations without reading them. It's about designing the moments before the irreversible action in a way that ensures the user arrives at it fully informed, not just technically warned.
That means surfacing the right information at the right moment before the point of no return. It means making the scope of the action concrete, not abstract. It means giving the user a path to modify, not just confirm or cancel.
That means surfacing the right information at the right moment before the point of no return. It means making the scope of the action concrete, not abstract. It means giving the user a path to modify, not just confirm or cancel.
That means surfacing the right information at the right moment before the point of no return. It means making the scope of the action concrete, not abstract. It means giving the user a path to modify, not just confirm or cancel.
The irreversible handoff is where the entire trust architecture of an AI product is tested. A system that handles these moments well earns a level of trust that almost nothing else can produce. A system that handles them badly loses it in a way that almost nothing else can recover.
The irreversible handoff is where the entire trust architecture of an AI product is tested. A system that handles these moments well earns a level of trust that almost nothing else can produce. A system that handles them badly loses it in a way that almost nothing else can recover.
The irreversible handoff is where the entire trust architecture of an AI product is tested. A system that handles these moments well earns a level of trust that almost nothing else can produce. A system that handles them badly loses it in a way that almost nothing else can recover.
The handoff as a moment of truth
The handoff as a moment of truth
The handoff as a moment of truth
Every interaction in an AI product contributes to the user's model of what the system is and how it behaves. But handoffs are disproportionately influential.
Every interaction in an AI product contributes to the user's model of what the system is and how it behaves. But handoffs are disproportionately influential.
Every interaction in an AI product contributes to the user's model of what the system is and how it behaves. But handoffs are disproportionately influential.
They're the moments when the user is most aware of the system's agency. When they're most conscious of the boundary between their intentions and the system's actions. When the gap between what they meant and what the system understood is most likely to become visible.
They're the moments when the user is most aware of the system's agency. When they're most conscious of the boundary between their intentions and the system's actions. When the gap between what they meant and what the system understood is most likely to become visible.
They're the moments when the user is most aware of the system's agency. When they're most conscious of the boundary between their intentions and the system's actions. When the gap between what they meant and what the system understood is most likely to become visible.
A product that handles handoffs well teaches its users to trust it incrementally. Each smooth transfer reinforces the belief that the system is working with the user, not around them. Over time, that belief becomes assumption. The handoffs become invisible because they've been validated enough times to stop requiring conscious attention.
A product that handles handoffs well teaches its users to trust it incrementally. Each smooth transfer reinforces the belief that the system is working with the user, not around them. Over time, that belief becomes assumption. The handoffs become invisible because they've been validated enough times to stop requiring conscious attention.
A product that handles handoffs well teaches its users to trust it incrementally. Each smooth transfer reinforces the belief that the system is working with the user, not around them. Over time, that belief becomes assumption. The handoffs become invisible because they've been validated enough times to stop requiring conscious attention.
A product that handles handoffs badly does the opposite. Each awkward transition is a reminder that the user is dealing with a system that doesn't fully understand them, operating in ways they can't fully predict. Trust erodes not in a single dramatic failure, but in a long sequence of small moments where the transfer didn't feel right.
A product that handles handoffs badly does the opposite. Each awkward transition is a reminder that the user is dealing with a system that doesn't fully understand them, operating in ways they can't fully predict. Trust erodes not in a single dramatic failure, but in a long sequence of small moments where the transfer didn't feel right.
A product that handles handoffs badly does the opposite. Each awkward transition is a reminder that the user is dealing with a system that doesn't fully understand them, operating in ways they can't fully predict. Trust erodes not in a single dramatic failure, but in a long sequence of small moments where the transfer didn't feel right.
The handoff is not a technical event. It's a relationship moment. It's the product demonstrating, again and again, whether it deserves the agency the user is extending to it. Design it accordingly.
The handoff is not a technical event. It's a relationship moment. It's the product demonstrating, again and again, whether it deserves the agency the user is extending to it. Design it accordingly.
The handoff is not a technical event. It's a relationship moment. It's the product demonstrating, again and again, whether it deserves the agency the user is extending to it. Design it accordingly.
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.