A few months ago, most AI conversations in design still focused on productivity: faster wireframes, generated interfaces, better copy, instant prototypes.
AI systems increasingly remember context, use tools, collaborate with other agents, operate across workflows, consume design systems, write code, and make decisions without requiring a human to specify every step.
The interface is no longer necessarily the product. Sometimes it is simply one layer between human intent and an intelligent system underneath.
Agentic systems are becoming more sophisticated, but products do not necessarily need to become more visibly agentic.
A system might use several specialized agents to research, compare information, validate an answer, and execute an action. The person using it probably does not care. They care that the work gets done.
For designers, this creates an interesting inversion: as the architecture becomes more complicated, the experience may need to become simpler.
The internal coordination belongs to the system. What belongs to the user is what matters: what is happening, what requires attention, what can happen autonomously, what still needs approval, and what happened if something went wrong.
The future may contain far more agents than agent interfaces.
Design systems were built to help teams create coherent products at scale.
Now they have another audience: machines.
Components, tokens, accessibility rules, interaction patterns, documentation, and code can increasingly become context that AI systems use to produce work consistent with an existing product.
Figma's MCP tooling makes this shift tangible. AI agents can extract variables, components, and layout information from Figma, generate code from selected designs, and increasingly write native content back to the canvas. In one recent Coinbase case study, Code Connect improved design-system adherence while reducing token costs by an average of 22.5 percent.
It is no longer only a library of reusable interface parts.
It can become a machine-readable expression of how the product should work.
A visually complete system with weak semantics, unclear relationships, or poorly documented rationale may be significantly less useful in an AI-driven workflow.
The familiar sequence—
design → handoff → development
—also starts to look increasingly artificial.
The more interesting source of truth may become the intent connecting all three.
That is a deeper design problem than simply deciding what a screen should contain.
For decades, designers could assume that eventually there would be a screen.
We could inspect it, prototype it, test it, approve it, and ship it.
That assumption is becoming less reliable.
Agents can increasingly interact directly with the capabilities underneath an interface. Other systems can generate interface structures dynamically according to context. Some work may be completed without presenting a conventional interface at all.
It makes consistency harder to define.
If the interface can adapt, what must remain stable?
Accessibility. Permissions. Consequences. Brand. Learnability. Human control. Trust.
These become invariants: things the system may design around, but should never casually rewrite.
Instead of defining exactly what every experience should look like, design systems may increasingly define the conditions within which experiences are allowed to change.
AI products want attention.
AI buttons. AI panels. AI suggestions. AI assistants attached to almost everything.
We are still in the stage where products feel the need to demonstrate intelligence.
Mature AI should not constantly interrupt someone simply because it has something it could say or do.
Sometimes the intelligent action is to prepare quietly.
Sometimes it is to surface one exception.
Sometimes it is to wait.
Sometimes it is to do nothing.
As AI becomes infrastructure rather than novelty, restraint may become one of the clearest signals of product maturity.
The best AI experience may eventually be the one we barely notice.
Taken together, these shifts point toward a larger change in the discipline.
UX has traditionally spent enormous effort defining interaction: what appears, where it appears, what happens after someone clicks, and what comes next.
Intelligent systems make some of those details increasingly dynamic.
So design moves upward.
Screens → systems → behaviour → rules → outcomes.
That does not mean screens disappear. It means they become one possible expression of something deeper.
Designers increasingly need to define intent, context, permissions, boundaries, consequences, escalation, trust, what the system should optimize for, and what it must never compromise.
That is a much larger design space than the interface alone.
The question used to be:
How will AI change the interfaces we design?
I am increasingly interested in a different one:
What do we design when intelligence becomes part of the medium itself?
The answer is starting to emerge.
We design the conditions under which intelligent systems behave well.
We decide where autonomy is useful and where it becomes dangerous. We decide what deserves attention and what should remain invisible. We define what can adapt and what must remain stable.
Most importantly, we make complexity understandable without exposing complexity simply because it exists.
AI is not making UX less important.
It is pushing UX beyond the screen—and forcing the discipline to become as systemic as we have always said it was.
This article considers how AI is changing the scope of UX design. In Designing AI Agents: Trust, Autonomy, and Human Control, I take the next step and examine how permissions, autonomy, feedback, and recovery can become concrete design decisions. Read the two together: this article frames the broader shift, while the companion explores what responsible implementation requires.
