AI-ready Design Systems for GenAI experiences

What needs to stay consistent when AI generates the interface?

I have never been particularly fond of design systems.

I understand why they exist, especially in large digital organizations where multiple teams are working across the same experience or a multi-product portfolio. They create consistency, reduce repeated work, and help companies develop digital experiences more efficiently.

Also, people like predictability. We like products that behave in familiar ways. We do not want to relearn how to submit a form or find our account settings every time we use a different product. So, it makes sense that design systems exist.

But the same standardization that created consistency also turned a lot of digital experiences vanilla. Designers started searching for existing components and patterns instead of asking whether there might be a better way. Again, I understand why. Not every button needs to be reinvented and most organizations cannot afford to explore every interaction from scratch.

And this got me thinking about what we are already seeing in chat experiences. The answers and follow-up questions respond to what someone has said, rather than following a script someone wrote in advance. Someone still designs how the experience should work, but they have not written every exchange. So how do we make sure the next interaction is meaningful and feels like part of the same experience?

AI is changing the equation.

I have been experimenting with tools such as Codex and Claude Code, and I am still impressed by how quickly an idea can become something tangible. You can describe what you want, provide some context, and soon have something that looks surprisingly close to what we could put into production.

This creates an obvious problem. Place that output next to the organization’s existing products and it often feels off. It might use the wrong patterns, ignore the brand guidelines, or use a component in completely the wrong context.

If AI has access to the components, code, content rules, and principles of an organization, it has a better chance of producing something that feels like it belongs there. That is useful, but it is only the beginning.

This process still feels familiar. Someone prompts an idea, AI generates an interface, a person reviews it and eventually a version is released. The process has become faster, but the interface still exists before the user gets there.

What happens when it does not?

The interface may not exist until you need it.

Today, most digital experiences have a predefined structure. The product team has already decided what information is visible, which steps people need to follow and how the experience should behave.

We can personalize some of it. Recommendations change, products appear in different order, and the offer reflects something the product knows about us. But the interface itself often remains the same.

AI introduces the possibility of going further. The product could generate the interface according to what someone is trying to achieve and what it knows about their context.

Let us imagine two people asking for a financial application to help understand their investment options. One understands financial products and wants to compare risk, fees, and historical performance. The other has never invested before and does not know where to start.

The first person might receive a detailed comparison table with filters. The second might receive a guided experience that explains the available options and asks a few questions before showing anything else. The experience would be generated when the person needs it.

This is what makes hyper-personalization more interesting to me. The product could change how someone interacts with it according to what they are trying to do.

But there is a distinction here. Getting a different answer does not necessarily mean the interface has changed. And getting a different interface does not mean it is more useful or more personal. That depends on whether the change helps the person.

Different cannot mean random.

If every user can receive a different interface, how do they learn how the product works? How do we make sure it remains accessible? How do we evaluate the variations? And how does a company stop the experience from turning into a random collection of screens that happen to work?

People still value familiarity. A product that changes every time you use it might be personalized, but it could also be completely confusing. Not everything should be adapted.

In the financial example, the explanations and comparison tools might change, while the way someone reviews and confirms a decision stays familiar. There may be parts of the product where generating a different interface adds no value at all.

And what happens when the person misunderstands the product? Asking a simple question does not necessarily mean someone is a beginner. Can they switch to the detailed comparison? Can they return to a view they already understand?

The system needs context, but people also need a way to correct its interpretation of that context. Otherwise, personalization could become another thing they have to fight to get their work done.

Figuring out what can change, what should stay familiar and how people regain control will be one of the more interesting design challenges ahead of us.

From components to rules for interaction

And this is where I think design systems have become even more important.

A generated interface does not require every component to be invented again. AI could choose from existing components whose behavior has already been defined and tested. The combination can change while the parts retain familiar behaviors.

But knowing that a component exists is different from knowing how to use it.

AI needs to understand what a component is for, when it should appear, how it works with other components and when it should not be used. It needs guidance on what can change and what must remain consistent. That includes ways for people to correct a misunderstanding, undo a change or return to a familiar view.

Some of this knowledge already involves design-system documentation. Some is spread across Figma comments, tickets, research, accessibility guidelines, and people’s heads. An AI-ready design system would need to bring more of it together in a way that both people and AI can use.

The design system can describe how an experience can be put together without prescribing every screen someone might need. In the financial example, it can explain when a guided sequence is appropriate, how it connects to a detailed comparison and how someone can move between the two without losing what they have already entered.

But documenting a rule does not guarantee that it will be followed. The product still needs checks on what gets generated and a usable fallback when the result does not meet those requirements. Giving AI better context is part of the work. Checking what it does with that context is another part.

Designing the possible experience

If more interfaces work this way, the work of designers changes too.

We may not design every screen that every person will see. Instead, we might design which parts of the experience are fixed, which can adapt and how far that adaptation can go.

We will still need to understand what people are trying to achieve and define how the product behaves, communicates, and earns trust. But we may increasingly be designing a range of experiences, including what happens when the system gets something wrong.

That sounds exciting, but also a bit uncomfortable. Designers are used to reviewing what goes into production. What happens when nobody has seen the exact interface before the user does?

I do not think the answer is to let AI figure it out. We need to become much clearer about the context, intent and constraints that should guide it, and keep checking whether the resulting experience helps someone complete their task.

This is where I see AI-ready design systems becoming part of the foundation. The same systems that made digital products more predictable could give AI enough structure to make them more personal, without making them completely unfamiliar.

The next interface may not exist until someone needs it. But the rules that make it feel like part of the product need to exist long before that.

We have been exploring AI-ready design systems at Nagarro. You can read more about that work here: AI-ready design systems at Nagarro.