2025
2025
01
01
Systems
Systems
Trust
Trust
Most AI products fail at the translation layer.
Most AI products fail at the translation layer.
Most AI products fail at the translation layer.
There is a moment in every AI product where something breaks. Not technically. The model works. The API responds. The data is accurate. But the user stops. Hesitates. Closes the tab. Calls support. Or worse: continues without understanding what just happened.
There is a moment in every AI product where something breaks. Not technically. The model works. The API responds. The data is accurate. But the user stops. Hesitates. Closes the tab. Calls support. Or worse: continues without understanding what just happened.
There is a moment in every AI product where something breaks. Not technically. The model works. The API responds. The data is accurate. But the user stops. Hesitates. Closes the tab. Calls support. Or worse: continues without understanding what just happened.
That moment has a name. It's the translation layer.
That moment has a name. It's the translation layer.
That moment has a name. It's the translation layer.
What the translation layer actually is
What the translation layer actually is
What the translation layer actually is
The translation layer is not a screen. It's not a component. It's the gap between what an AI system does and what a human being understands about what just happened.
The translation layer is not a screen. It's not a component. It's the gap between what an AI system does and what a human being understands about what just happened.
The translation layer is not a screen. It's not a component. It's the gap between what an AI system does and what a human being understands about what just happened.
Every AI product has one. Most of them fail there.
Every AI product has one. Most of them fail there.
Every AI product has one. Most of them fail there.
When a model returns a confidence score, someone has to decide what that means for the user. When an AI takes an action autonomously, someone has to decide how much of that process the user sees. When the system is wrong, someone has to decide how that failure gets communicated without destroying trust in everything that came before it.
When a model returns a confidence score, someone has to decide what that means for the user. When an AI takes an action autonomously, someone has to decide how much of that process the user sees. When the system is wrong, someone has to decide how that failure gets communicated without destroying trust in everything that came before it.
When a model returns a confidence score, someone has to decide what that means for the user. When an AI takes an action autonomously, someone has to decide how much of that process the user sees. When the system is wrong, someone has to decide how that failure gets communicated without destroying trust in everything that came before it.
These are not engineering decisions. They are not product decisions. They are design decisions. And in most AI products, nobody is making them deliberately.
These are not engineering decisions. They are not product decisions. They are design decisions. And in most AI products, nobody is making them deliberately.
These are not engineering decisions. They are not product decisions. They are design decisions. And in most AI products, nobody is making them deliberately.
Why this keeps happening
Why this keeps happening
Why this keeps happening
It looks like a recommendation with no explanation. The system chose something, but the user doesn't know why, and doesn't know if they should trust it.
It looks like a recommendation with no explanation. The system chose something, but the user doesn't know why, and doesn't know if they should trust it.
It looks like a recommendation with no explanation. The system chose something, but the user doesn't know why, and doesn't know if they should trust it.
It looks like an autonomous action that happened while the user wasn't looking. Technically correct. Experientially alarming.
It looks like an autonomous action that happened while the user wasn't looking. Technically correct. Experientially alarming.
It looks like an autonomous action that happened while the user wasn't looking. Technically correct. Experientially alarming.
It looks like an error message that tells you what went wrong at the infrastructure level, when all the user needed to know was what to do next.
It looks like an error message that tells you what went wrong at the infrastructure level, when all the user needed to know was what to do next.
It looks like an error message that tells you what went wrong at the infrastructure level, when all the user needed to know was what to do next.
It looks like a confidence indicator that means something precise to an engineer and nothing at all to the person making a decision based on it.
It looks like a confidence indicator that means something precise to an engineer and nothing at all to the person making a decision based on it.
It looks like a confidence indicator that means something precise to an engineer and nothing at all to the person making a decision based on it.
Most AI products treat this as a detail. It's not. It's the product.
Most AI products treat this as a detail. It's not. It's the product.
Most AI products treat this as a detail. It's not. It's the product.
What failing at the translation layer looks like
What failing at the translation layer looks like
What failing at the translation layer looks like
The teams building AI products are, understandably, obsessed with what the system can do. The capability is the hard part. Getting the model to work, to scale, to produce reliable outputs: that's where the energy goes. Design gets brought in later, often to make the interface look clean and the onboarding feel smooth.
The teams building AI products are, understandably, obsessed with what the system can do. The capability is the hard part. Getting the model to work, to scale, to produce reliable outputs: that's where the energy goes. Design gets brought in later, often to make the interface look clean and the onboarding feel smooth.
The teams building AI products are, understandably, obsessed with what the system can do. The capability is the hard part. Getting the model to work, to scale, to produce reliable outputs: that's where the energy goes. Design gets brought in later, often to make the interface look clean and the onboarding feel smooth.
But the translation layer isn't a surface problem. You can't solve it with better typography or a cleaner dashboard. It lives in the logic of how information is presented, when it's revealed, what's shown versus what's hidden, and how the system communicates its own limitations.
But the translation layer isn't a surface problem. You can't solve it with better typography or a cleaner dashboard. It lives in the logic of how information is presented, when it's revealed, what's shown versus what's hidden, and how the system communicates its own limitations.
But the translation layer isn't a surface problem. You can't solve it with better typography or a cleaner dashboard. It lives in the logic of how information is presented, when it's revealed, what's shown versus what's hidden, and how the system communicates its own limitations.
In each case, the system worked. The product failed.
In each case, the system worked. The product failed.
In each case, the system worked. The product failed.
What the translation layer requires
What the translation layer requires
What the translation layer requires
Getting this right requires a specific kind of design thinking that most product teams aren't applying yet.
Getting this right requires a specific kind of design thinking that most product teams aren't applying yet.
Getting this right requires a specific kind of design thinking that most product teams aren't applying yet.
It requires designing for what the user needs to know, not what the system knows. These are rarely the same thing. An AI system might have 40 data points informing a decision. The user needs two or three of them, framed correctly, at the right moment.
It requires designing for what the user needs to know, not what the system knows. These are rarely the same thing. An AI system might have 40 data points informing a decision. The user needs two or three of them, framed correctly, at the right moment.
It requires designing for what the user needs to know, not what the system knows. These are rarely the same thing. An AI system might have 40 data points informing a decision. The user needs two or three of them, framed correctly, at the right moment.
It requires designing for uncertainty. Most UX patterns assume deterministic systems: you click, something happens, the result is shown. AI systems don't work that way. The result depends on context, on inputs the user didn't consciously provide, on model behaviors that shift over time. The interface has to hold that uncertainty without amplifying it into anxiety.
It requires designing for uncertainty. Most UX patterns assume deterministic systems: you click, something happens, the result is shown. AI systems don't work that way. The result depends on context, on inputs the user didn't consciously provide, on model behaviors that shift over time. The interface has to hold that uncertainty without amplifying it into anxiety.
It requires designing for uncertainty. Most UX patterns assume deterministic systems: you click, something happens, the result is shown. AI systems don't work that way. The result depends on context, on inputs the user didn't consciously provide, on model behaviors that shift over time. The interface has to hold that uncertainty without amplifying it into anxiety.
It requires designing trust progressively. Users don't trust AI systems by default. They extend trust incrementally, based on small signals that compound over time.
It requires designing trust progressively. Users don't trust AI systems by default. They extend trust incrementally, based on small signals that compound over time.
It requires designing trust progressively. Users don't trust AI systems by default. They extend trust incrementally, based on small signals that compound over time.
Every interaction at the translation layer is either building that trust or eroding it. There is no neutral.
Every interaction at the translation layer is either building that trust or eroding it. There is no neutral.
Every interaction at the translation layer is either building that trust or eroding it. There is no neutral.
What most teams miss
What most teams miss
What most teams miss
The translation layer is not just about what you show. It's about what you deliberately don't show.
The translation layer is not just about what you show. It's about what you deliberately don't show.
The translation layer is not just about what you show. It's about what you deliberately don't show.
Hiding complexity is not the same as removing it. The complexity stays. The infrastructure stays invisible. The model stays invisible. But the user still needs to feel that something real and reliable is working underneath.
Hiding complexity is not the same as removing it. The complexity stays. The infrastructure stays invisible. The model stays invisible. But the user still needs to feel that something real and reliable is working underneath.
Hiding complexity is not the same as removing it. The complexity stays. The infrastructure stays invisible. The model stays invisible. But the user still needs to feel that something real and reliable is working underneath.
That feeling doesn't come from animations or reassuring microcopy. It comes from the architecture of the experience: from the moments where the system explains itself without being asked, from the controls that give users a sense of agency even when they're not using them, from the consistency that makes the product feel predictable even when the outputs aren't.
That feeling doesn't come from animations or reassuring microcopy. It comes from the architecture of the experience: from the moments where the system explains itself without being asked, from the controls that give users a sense of agency even when they're not using them, from the consistency that makes the product feel predictable even when the outputs aren't.
That feeling doesn't come from animations or reassuring microcopy. It comes from the architecture of the experience: from the moments where the system explains itself without being asked, from the controls that give users a sense of agency even when they're not using them, from the consistency that makes the product feel predictable even when the outputs aren't.
This is the hardest design problem in AI products right now. Not because it's technically complex. Because it requires holding two things in tension simultaneously: making the system invisible, and making the user feel in control of it.
This is the hardest design problem in AI products right now. Not because it's technically complex. Because it requires holding two things in tension simultaneously: making the system invisible, and making the user feel in control of it.
This is the hardest design problem in AI products right now. Not because it's technically complex. Because it requires holding two things in tension simultaneously: making the system invisible, and making the user feel in control of it.
The work nobody talks about
The work nobody talks about
The work nobody talks about
There is a version of AI product design that is mostly cosmetic. Clean interfaces on top of powerful models. Good typography. Smooth transitions. The design equivalent of a nice lobby in a building with structural problems.
There is a version of AI product design that is mostly cosmetic. Clean interfaces on top of powerful models. Good typography. Smooth transitions. The design equivalent of a nice lobby in a building with structural problems.
There is a version of AI product design that is mostly cosmetic. Clean interfaces on top of powerful models. Good typography. Smooth transitions. The design equivalent of a nice lobby in a building with structural problems.
And then there is the work of actually designing the translation layer. Of deciding, for every piece of information the system produces, how it gets communicated, when, at what level of detail, with what degree of transparency about how it was generated.
And then there is the work of actually designing the translation layer. Of deciding, for every piece of information the system produces, how it gets communicated, when, at what level of detail, with what degree of transparency about how it was generated.
And then there is the work of actually designing the translation layer. Of deciding, for every piece of information the system produces, how it gets communicated, when, at what level of detail, with what degree of transparency about how it was generated.
That work is not glamorous. It doesn't show up in design awards. But it's the difference between an AI product that users trust and one they abandon after the third session.
That work is not glamorous. It doesn't show up in design awards. But it's the difference between an AI product that users trust and one they abandon after the third session.
That work is not glamorous. It doesn't show up in design awards. But it's the difference between an AI product that users trust and one they abandon after the third session.
Most AI products fail at the translation layer because nobody owns it. Engineering owns the model. Product owns the roadmap. Design gets the screens.
Most AI products fail at the translation layer because nobody owns it. Engineering owns the model. Product owns the roadmap. Design gets the screens.
Most AI products fail at the translation layer because nobody owns it. Engineering owns the model. Product owns the roadmap. Design gets the screens.
The translation layer lives between all three. Until someone treats it as the core design problem, it will keep failing.
The translation layer lives between all three. Until someone treats it as the core design problem, it will keep failing.
The translation layer lives between all three. Until someone treats it as the core design problem, it will keep failing.
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.