2026
2026
09
09
Behavior
Behavior
Agency
Agency
Designing for the second decision.
Designing for the second decision.
Designing for the second decision.
Most design work is front-loaded. The research, the wireframes, the prototypes, the usability tests: they are almost universally focused on the beginning of the user's journey. The first impression. The first interaction. The moment of activation.
Most design work is front-loaded. The research, the wireframes, the prototypes, the usability tests: they are almost universally focused on the beginning of the user's journey. The first impression. The first interaction. The moment of activation.
Most design work is front-loaded. The research, the wireframes, the prototypes, the usability tests: they are almost universally focused on the beginning of the user's journey. The first impression. The first interaction. The moment of activation.
This is understandable. First impressions matter. Activation matters. But in AI products, the design decision that determines whether a user stays or leaves, whether they trust the product or abandon it, whether they integrate it into how they actually work or relegate it to occasional curiosity, almost never happens first.
This is understandable. First impressions matter. Activation matters. But in AI products, the design decision that determines whether a user stays or leaves, whether they trust the product or abandon it, whether they integrate it into how they actually work or relegate it to occasional curiosity, almost never happens first.
This is understandable. First impressions matter. Activation matters. But in AI products, the design decision that determines whether a user stays or leaves, whether they trust the product or abandon it, whether they integrate it into how they actually work or relegate it to occasional curiosity, almost never happens first.
It happens second.
It happens second.
It happens second.
Why design is obsessed with the first decision
Why design is obsessed with the first decision
Why design is obsessed with the first decision
The first decision is legible. You can observe it. You can test it. You can measure it. Click-through rates, conversion rates, time to first action: these are all measurements of the first decision, and they're relatively easy to collect and act on.
The first decision is legible. You can observe it. You can test it. You can measure it. Click-through rates, conversion rates, time to first action: these are all measurements of the first decision, and they're relatively easy to collect and act on.
The first decision is legible. You can observe it. You can test it. You can measure it. Click-through rates, conversion rates, time to first action: these are all measurements of the first decision, and they're relatively easy to collect and act on.
The first decision is also controllable. The designer shapes the environment in which it happens. The layout, the copy, the hierarchy, the friction: all of these are levers that directly influence whether the user takes the first step.
The first decision is also controllable. The designer shapes the environment in which it happens. The layout, the copy, the hierarchy, the friction: all of these are levers that directly influence whether the user takes the first step.
The first decision is also controllable. The designer shapes the environment in which it happens. The layout, the copy, the hierarchy, the friction: all of these are levers that directly influence whether the user takes the first step.
And the first decision is narratively satisfying. It's the moment the user says yes. The moment the product earns its first proof of value. In a world where products compete for attention and acquisition costs are high, getting to yes quickly feels like the whole game.
And the first decision is narratively satisfying. It's the moment the user says yes. The moment the product earns its first proof of value. In a world where products compete for attention and acquisition costs are high, getting to yes quickly feels like the whole game.
And the first decision is narratively satisfying. It's the moment the user says yes. The moment the product earns its first proof of value. In a world where products compete for attention and acquisition costs are high, getting to yes quickly feels like the whole game.
But it isn't. The first decision gets the user into the product. The second decision determines whether they stay.
But it isn't. The first decision gets the user into the product. The second decision determines whether they stay.
But it isn't. The first decision gets the user into the product. The second decision determines whether they stay.
Design that optimizes exclusively for the first decision produces products with strong acquisition metrics and weak retention. The user said yes. They just didn't find a reason to say yes again.
Design that optimizes exclusively for the first decision produces products with strong acquisition metrics and weak retention. The user said yes. They just didn't find a reason to say yes again.
Design that optimizes exclusively for the first decision produces products with strong acquisition metrics and weak retention. The user said yes. They just didn't find a reason to say yes again.
What the second decision actually is
What the second decision actually is
What the second decision actually is
In a traditional product, the second decision is usually straightforward. The user has completed an action, the system has responded, and the next step is obvious. The flow continues. The second decision is just the next step in a designed sequence.
In a traditional product, the second decision is usually straightforward. The user has completed an action, the system has responded, and the next step is obvious. The flow continues. The second decision is just the next step in a designed sequence.
In a traditional product, the second decision is usually straightforward. The user has completed an action, the system has responded, and the next step is obvious. The flow continues. The second decision is just the next step in a designed sequence.
In an AI product, the second decision is genuinely complex. The system has produced an output. That output exists on a spectrum from exactly right to completely wrong, with most outputs living somewhere in the ambiguous middle. The user has to evaluate it, form a judgment about its quality and relevance, and decide what to do next.
In an AI product, the second decision is genuinely complex. The system has produced an output. That output exists on a spectrum from exactly right to completely wrong, with most outputs living somewhere in the ambiguous middle. The user has to evaluate it, form a judgment about its quality and relevance, and decide what to do next.
In an AI product, the second decision is genuinely complex. The system has produced an output. That output exists on a spectrum from exactly right to completely wrong, with most outputs living somewhere in the ambiguous middle. The user has to evaluate it, form a judgment about its quality and relevance, and decide what to do next.
That decision is not designed in most AI products. The output gets a container. The output gets typography and spacing and perhaps a copy action. But the moment of evaluation, the moment when the user looks at what the system produced and decides whether to act on it, modify it, or reject it, that moment is treated as the user's problem, not the designer's.
That decision is not designed in most AI products. The output gets a container. The output gets typography and spacing and perhaps a copy action. But the moment of evaluation, the moment when the user looks at what the system produced and decides whether to act on it, modify it, or reject it, that moment is treated as the user's problem, not the designer's.
That decision is not designed in most AI products. The output gets a container. The output gets typography and spacing and perhaps a copy action. But the moment of evaluation, the moment when the user looks at what the system produced and decides whether to act on it, modify it, or reject it, that moment is treated as the user's problem, not the designer's.
It isn't. It's one of the most consequential design moments in the product.
It isn't. It's one of the most consequential design moments in the product.
It isn't. It's one of the most consequential design moments in the product.
The second decision is where the user's mental model of the system gets updated. Where trust is confirmed or questioned. Where the relationship between the user and the product either deepens or stalls. Getting it right requires designing not just the output, but the entire context in which the output is received and evaluated.
The second decision is where the user's mental model of the system gets updated. Where trust is confirmed or questioned. Where the relationship between the user and the product either deepens or stalls. Getting it right requires designing not just the output, but the entire context in which the output is received and evaluated.
The second decision is where the user's mental model of the system gets updated. Where trust is confirmed or questioned. Where the relationship between the user and the product either deepens or stalls. Getting it right requires designing not just the output, but the entire context in which the output is received and evaluated.
The second decision is the moment the product proves it was worth the first one. Most AI products have no design for it at all.
The second decision is the moment the product proves it was worth the first one. Most AI products have no design for it at all.
The second decision is the moment the product proves it was worth the first one. Most AI products have no design for it at all.
Three paths, three different design problems
Three paths, three different design problems
Three paths, three different design problems
When a user receives an AI output, there are three things they can do. They can act on it directly. They can modify it before acting. They can reject it and try again.
When a user receives an AI output, there are three things they can do. They can act on it directly. They can modify it before acting. They can reject it and try again.
When a user receives an AI output, there are three things they can do. They can act on it directly. They can modify it before acting. They can reject it and try again.
Each of these paths is a different design problem, and each requires different things from the interface.
Each of these paths is a different design problem, and each requires different things from the interface.
Each of these paths is a different design problem, and each requires different things from the interface.
Acting directly requires confidence. The user has evaluated the output and decided it's good enough to use as is. The design challenge here is making that evaluation easy. Not by hiding the output's limitations, but by giving the user the signals they need to assess quality quickly and accurately. An output that takes significant cognitive effort to evaluate will be acted on less often, regardless of its actual quality, because the effort itself becomes a form of friction.
Acting directly requires confidence. The user has evaluated the output and decided it's good enough to use as is. The design challenge here is making that evaluation easy. Not by hiding the output's limitations, but by giving the user the signals they need to assess quality quickly and accurately. An output that takes significant cognitive effort to evaluate will be acted on less often, regardless of its actual quality, because the effort itself becomes a form of friction.
Acting directly requires confidence. The user has evaluated the output and decided it's good enough to use as is. The design challenge here is making that evaluation easy. Not by hiding the output's limitations, but by giving the user the signals they need to assess quality quickly and accurately. An output that takes significant cognitive effort to evaluate will be acted on less often, regardless of its actual quality, because the effort itself becomes a form of friction.
Modifying requires access and orientation. The user has decided the output is close but not quite right. They need to be able to change it without feeling like they're fighting the system. The design challenge is making modification feel like a natural extension of the interaction, not a workaround. The interface should make it obvious what can be changed, how to change it, and what effect the change will have on the output.
Modifying requires access and orientation. The user has decided the output is close but not quite right. They need to be able to change it without feeling like they're fighting the system. The design challenge is making modification feel like a natural extension of the interaction, not a workaround. The interface should make it obvious what can be changed, how to change it, and what effect the change will have on the output.
Modifying requires access and orientation. The user has decided the output is close but not quite right. They need to be able to change it without feeling like they're fighting the system. The design challenge is making modification feel like a natural extension of the interaction, not a workaround. The interface should make it obvious what can be changed, how to change it, and what effect the change will have on the output.
Rejecting and restarting requires recovery without punishment. The user has decided the output doesn't serve their need. They need to try again, which means the system needs to understand what went wrong and give them a path to a better result. The design challenge is making rejection feel productive rather than wasteful. An interface that treats rejection as a dead end, sending the user back to a blank input with no memory of what just failed, is an interface that trains users to lower their standards rather than improve their requests.
Rejecting and restarting requires recovery without punishment. The user has decided the output doesn't serve their need. They need to try again, which means the system needs to understand what went wrong and give them a path to a better result. The design challenge is making rejection feel productive rather than wasteful. An interface that treats rejection as a dead end, sending the user back to a blank input with no memory of what just failed, is an interface that trains users to lower their standards rather than improve their requests.
Rejecting and restarting requires recovery without punishment. The user has decided the output doesn't serve their need. They need to try again, which means the system needs to understand what went wrong and give them a path to a better result. The design challenge is making rejection feel productive rather than wasteful. An interface that treats rejection as a dead end, sending the user back to a blank input with no memory of what just failed, is an interface that trains users to lower their standards rather than improve their requests.
Designing for all three paths, with the same care given to each, is what separates AI products that improve with use from AI products that plateau after the first session.
Designing for all three paths, with the same care given to each, is what separates AI products that improve with use from AI products that plateau after the first session.
Designing for all three paths, with the same care given to each, is what separates AI products that improve with use from AI products that plateau after the first session.
What the second decision reveals about the product
What the second decision reveals about the product
What the second decision reveals about the product
The second decision is a diagnostic. How users make it, and which path they most often take, reveals more about the product's actual quality than almost any other signal.
The second decision is a diagnostic. How users make it, and which path they most often take, reveals more about the product's actual quality than almost any other signal.
The second decision is a diagnostic. How users make it, and which path they most often take, reveals more about the product's actual quality than almost any other signal.
A product where users consistently act directly on outputs is a product with well-calibrated output quality and clear evaluation signals. Users trust it enough to act, and the trust is warranted.
A product where users consistently act directly on outputs is a product with well-calibrated output quality and clear evaluation signals. Users trust it enough to act, and the trust is warranted.
A product where users consistently act directly on outputs is a product with well-calibrated output quality and clear evaluation signals. Users trust it enough to act, and the trust is warranted.
A product where users consistently modify outputs is a product that's close but not quite right. The system understands the domain but not the individual user well enough yet. The modification behavior is valuable signal: it tells the product team exactly where the gap between the system's output and the user's need is located.
A product where users consistently modify outputs is a product that's close but not quite right. The system understands the domain but not the individual user well enough yet. The modification behavior is valuable signal: it tells the product team exactly where the gap between the system's output and the user's need is located.
A product where users consistently modify outputs is a product that's close but not quite right. The system understands the domain but not the individual user well enough yet. The modification behavior is valuable signal: it tells the product team exactly where the gap between the system's output and the user's need is located.
A product where users consistently reject and restart is a product with a fundamental alignment problem. The system is not producing outputs that match what users actually need, and users have learned that modification is futile. This is the most serious signal, and it's almost always traceable to a design failure in how the product elicits input from the user, not just how it produces output.
A product where users consistently reject and restart is a product with a fundamental alignment problem. The system is not producing outputs that match what users actually need, and users have learned that modification is futile. This is the most serious signal, and it's almost always traceable to a design failure in how the product elicits input from the user, not just how it produces output.
A product where users consistently reject and restart is a product with a fundamental alignment problem. The system is not producing outputs that match what users actually need, and users have learned that modification is futile. This is the most serious signal, and it's almost always traceable to a design failure in how the product elicits input from the user, not just how it produces output.
A product where users disengage entirely after receiving an output, neither acting, modifying, nor explicitly rejecting, is a product that has failed at the second decision in the most complete way. The user received something, didn't know what to do with it, and left. That disengagement is the hardest signal to read and the most important to design against.
A product where users disengage entirely after receiving an output, neither acting, modifying, nor explicitly rejecting, is a product that has failed at the second decision in the most complete way. The user received something, didn't know what to do with it, and left. That disengagement is the hardest signal to read and the most important to design against.
A product where users disengage entirely after receiving an output, neither acting, modifying, nor explicitly rejecting, is a product that has failed at the second decision in the most complete way. The user received something, didn't know what to do with it, and left. That disengagement is the hardest signal to read and the most important to design against.
The second decision is the product's report card. Reading it requires instrumenting not just what users do first, but what they do with what the system gives them. Most analytics don't capture this. Most products are flying blind at the moment that matters most.
The second decision is the product's report card. Reading it requires instrumenting not just what users do first, but what they do with what the system gives them. Most analytics don't capture this. Most products are flying blind at the moment that matters most.
The second decision is the product's report card. Reading it requires instrumenting not just what users do first, but what they do with what the system gives them. Most analytics don't capture this. Most products are flying blind at the moment that matters most.
Designing for the right second decision, not the fast one
Designing for the right second decision, not the fast one
Designing for the right second decision, not the fast one
There is a temptation in AI product design to optimize for speed at the second decision. To make it as easy as possible for the user to act on the output, because action is a positive signal and friction is a negative one.
There is a temptation in AI product design to optimize for speed at the second decision. To make it as easy as possible for the user to act on the output, because action is a positive signal and friction is a negative one.
There is a temptation in AI product design to optimize for speed at the second decision. To make it as easy as possible for the user to act on the output, because action is a positive signal and friction is a negative one.
This is the wrong optimization.
This is the wrong optimization.
This is the wrong optimization.
The goal is not a fast second decision. It's a good one. A user who acts quickly on a poor output has not had a good experience. They've had an efficient one. The efficiency will not save them from the disappointment that follows when they discover the output didn't serve their need.
The goal is not a fast second decision. It's a good one. A user who acts quickly on a poor output has not had a good experience. They've had an efficient one. The efficiency will not save them from the disappointment that follows when they discover the output didn't serve their need.
The goal is not a fast second decision. It's a good one. A user who acts quickly on a poor output has not had a good experience. They've had an efficient one. The efficiency will not save them from the disappointment that follows when they discover the output didn't serve their need.
Designing for the right second decision means giving users the time, the information, and the tools they need to evaluate the output accurately before they act on it. It means making the cost of modifying or rejecting low enough that users exercise that option when they should, rather than acting on outputs they know are suboptimal because the path to something better feels too long.
Designing for the right second decision means giving users the time, the information, and the tools they need to evaluate the output accurately before they act on it. It means making the cost of modifying or rejecting low enough that users exercise that option when they should, rather than acting on outputs they know are suboptimal because the path to something better feels too long.
Designing for the right second decision means giving users the time, the information, and the tools they need to evaluate the output accurately before they act on it. It means making the cost of modifying or rejecting low enough that users exercise that option when they should, rather than acting on outputs they know are suboptimal because the path to something better feels too long.
It means designing the moment of evaluation as deliberately as the moment of input. What information does the user need to assess this output? Where should that information live? How much of it is necessary before action, and how much can be deferred? What signals indicate that the output is reliable in this context, and how are those signals communicated?
It means designing the moment of evaluation as deliberately as the moment of input. What information does the user need to assess this output? Where should that information live? How much of it is necessary before action, and how much can be deferred? What signals indicate that the output is reliable in this context, and how are those signals communicated?
It means designing the moment of evaluation as deliberately as the moment of input. What information does the user need to assess this output? Where should that information live? How much of it is necessary before action, and how much can be deferred? What signals indicate that the output is reliable in this context, and how are those signals communicated?
An interface that makes the right second decision easy is an interface that has internalized a truth most design processes ignore: the value of an AI product is not in the outputs it produces. It's in the decisions users make because of them.
An interface that makes the right second decision easy is an interface that has internalized a truth most design processes ignore: the value of an AI product is not in the outputs it produces. It's in the decisions users make because of them.
An interface that makes the right second decision easy is an interface that has internalized a truth most design processes ignore: the value of an AI product is not in the outputs it produces. It's in the decisions users make because of them.
The second decision as a relationship signal
The second decision as a relationship signal
The second decision as a relationship signal
Over time, the pattern of second decisions a user makes tells a story about their relationship with the product.
Over time, the pattern of second decisions a user makes tells a story about their relationship with the product.
Over time, the pattern of second decisions a user makes tells a story about their relationship with the product.
A user who moves from frequent modification toward frequent direct action is a user whose trust in the system is growing. The system is learning their preferences, and they are learning the system's capabilities. The second decisions are getting easier because the relationship is maturing.
A user who moves from frequent modification toward frequent direct action is a user whose trust in the system is growing. The system is learning their preferences, and they are learning the system's capabilities. The second decisions are getting easier because the relationship is maturing.
A user who moves from frequent modification toward frequent direct action is a user whose trust in the system is growing. The system is learning their preferences, and they are learning the system's capabilities. The second decisions are getting easier because the relationship is maturing.
A user who moves in the opposite direction, from direct action toward increasing modification and rejection, is a user whose trust is eroding. Something has changed, in the system's behavior, in the user's needs, or in the gap between them. The second decisions are getting harder because the relationship is breaking down.
A user who moves in the opposite direction, from direct action toward increasing modification and rejection, is a user whose trust is eroding. Something has changed, in the system's behavior, in the user's needs, or in the gap between them. The second decisions are getting harder because the relationship is breaking down.
A user who moves in the opposite direction, from direct action toward increasing modification and rejection, is a user whose trust is eroding. Something has changed, in the system's behavior, in the user's needs, or in the gap between them. The second decisions are getting harder because the relationship is breaking down.
A user whose second decisions are consistently fast and confident is a user who has found their working mode with the product. They know what to expect, they know how to evaluate it, and they know what to do with it. That fluency is the product of design: not of a single interaction, but of every interaction that built toward it.
A user whose second decisions are consistently fast and confident is a user who has found their working mode with the product. They know what to expect, they know how to evaluate it, and they know what to do with it. That fluency is the product of design: not of a single interaction, but of every interaction that built toward it.
A user whose second decisions are consistently fast and confident is a user who has found their working mode with the product. They know what to expect, they know how to evaluate it, and they know what to do with it. That fluency is the product of design: not of a single interaction, but of every interaction that built toward it.
Designing for the second decision, done well, produces users who reach that fluency. Who develop a working relationship with the system that feels natural and productive. Who don't think about the product when they're using it because the second decision has become instinctive.
Designing for the second decision, done well, produces users who reach that fluency. Who develop a working relationship with the system that feels natural and productive. Who don't think about the product when they're using it because the second decision has become instinctive.
Designing for the second decision, done well, produces users who reach that fluency. Who develop a working relationship with the system that feels natural and productive. Who don't think about the product when they're using it because the second decision has become instinctive.
That instinct is the goal. Not the first yes, but the hundredth effortless yes. The second decision, designed well, is what gets you there.
That instinct is the goal. Not the first yes, but the hundredth effortless yes. The second decision, designed well, is what gets you there.
That instinct is the goal. Not the first yes, but the hundredth effortless yes. The second decision, designed well, is what gets you there.
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.