AI has made design systems faster to build and, as a result, easier to let drift. A single prompt can now produce foundations and a first set of components, which means anyone in a business can generate something that looks finished. That leaves a harder question about what happens next.
I posed that question to design system leaders through a short survey, then brought some of them together at a facilitated leaders session we hosted in London to discuss it further. The survey results reported that the most common worry was governance keeping pace with AI, followed closely by the role of human judgement. Naming, token architecture and scope also came up.
Our view at YLD, draws on our own work as well as what we have heard and seen across the industry: The teams that do well build it right first, because that is what makes the speed worth having. With that in mind, here are five shifts that show what this looks like in practice.
1. Governance moves from control to guardrails set early
At the leaders session, one group asked whether we have lost the battle for governance. Most tables, however, gave a more careful answer: the battle has moved. A team that holds a hard line risks being bypassed. With design systems now being used to create flows by engineers and product managers, expecting 100% correct usage is unrealistic. A more achievable target is around 75%, with the remaining 25% treated as an opportunity to learn what needs to improve or change within the system. That said, the room also agreed that large organisations still benefit from firm structure. Without it, poor work gets repeated across the business, and AI only speeds that up. We see the same pattern in our client work. Whatever the target, it only holds when the guardrails are in place early. In our AI-era engagements, governance now takes a much larger share of the work, including a written AI policy on what agents may generate freely and what needs review.
Governance now also covers how AI reads the system and the internal tools people build with it. Scope belongs in the same conversation as governance. AI makes it possible to take on work that would never have made it into a sprint. However, without an agreed scope, a system can grow faster than anyone can maintain it, through contributions nobody has checked and usage nobody can see. In our own work, we now put more emphasis than ever on agreeing scope before building starts. Doing so also makes the next shift possible: Once you agree what the system covers, you can give everyone else somewhere to explore outside it. Because buy-in is hardest to win in the product layer, product needs to be part of that conversation from the start.
2. Playgrounds separate exploration from production
Thanks to AI, more roles, from POs to the CEO, can now generate UI and code quickly, and the truth is that they will do so with or without your design system. Rather than clamping down and trying to prevent the unpreventable, the room aligned on creating a playground is a good idea: a separate space where anyone in the business can “have a go”, explore and prototype freely without touching production.
What a prototype is for decides how much governance it needs. A prototype made to learn from needs very little, while one headed for production needs much more. The risk is that quick AI prototypes look finished when they are really just sketches. One team that asked people to justify their ideas saw proposals fall from dozens a quarter to a handful. The playgrounds we have seen work best are the ones where everyone agrees from the start what each prototype is for.
3. Human judgement becomes the deciding factor
"Human in the loop" is a phrase everyone uses, but the room questioned what it really means. Having an actual person in the loop does not guarantee good outputs, or even the right ones. That depends on whether they are the right person, what expertise they bring, and whether they know what they are there to do and what to check for.
A long queue of approvals also invites quick click-through, which doesn’t guarantee that errors are spotted. For that reason, the framing with most support was that AI finds the issues and people solve them. In practice, that means agents sweeping for incorrect component usage, off-brand copy and accessibility failures, with people deciding what to do about them.
This links to one of the worries leaders listed most often in the survey: the role of human judgement and taste. In our experience, a person in the loop works brilliantly when they know what to check. AI has democratised capability so that anyone can now generate a component, and the output can look correct while hiding performance or accessibility problems. It still needs substantial refinement, and it is only as good as the system it works from. Expertise is knowing what to question when something looks right.
4. The team's role evolves from maintaining components to scaling judgement
During the leaders session, the room also asked what a design system team is for if speed and consistency no longer set it apart. Their answers centred on judgement, codified knowledge and a consulting role: scaling design practice and responding to change. We saw the same shift ourselves when we built an internal baseline design system with AI. Generating the components was the easy part. Most of the effort went into checking every contribution and watching for drift, which is exactly the kind of judgement work the room described.
Designers are already connecting to codebases and clearing old design debt, which led the room to debate whether they should upskill in engineering. The conclusion? The room was split. Many of the leaders I surveyed run small teams, so the question is a practical one. The room is right to be wary of investing in skills the tooling might absorb. In our project work, discipline expertise still matters, and we still staff our design systems projects with both a designer and an engineer. Even so, we can see that changing as the lines between roles continue to blur.
A future design systems individual contributor will not look the same as it does today. We expect that role to look more like a Product Builder: a product designer at the centre, with a broader understanding of how things get built added to the craft. We already see designers opening pull requests for small fixes under engineers’ review, clients’ own teams keeping the system moving after handover, and smaller teams covering work and roles that were rarely funded.
5. The design system becomes infrastructure for AI
Teams that have connected AI to their design system described quick gains. A prototype that looks off-brand can match the brand within minutes once the system is exported as markdown. When AI is linked across ticketing, design, documentation and code now flags drift and finds features documented nowhere, and a change log is turning into a decision log that a model can read. This led the room to a broader question: how should AI read and use the system, alongside the people who already do? For token updates, one person suggested a non-LLM decision model such as Jev, which applies a change exactly as specified and uses fewer tokens.
Making naming, tokens and documentation easier for models to read came up in both the survey and the leaders session discussions. In our engagements we now treat it as a phase of its own: connecting design and code, setting up MCP access, packaging the system as an agent skill and detecting drift automatically. As AI lowers the barrier to entry, more people can produce UI and code. The design system becomes the shared source of truth, so the right components, tokens and patterns are used every time, whether a person or a model does the work.
Without a well-structured, well-managed design system, everyone falls back on their own judgement, or the model’s. With one in place, AI can work across the whole organisation to the same standard. For that reason, a design system is becoming essential infrastructure for AI-powered product development at scale.
Build it right
Speed is no longer what separates one design system from another. Anyone can generate a component, and one prompt can produce a first set of foundations. What separates them instead is what makes that speed safe: guardrails and scope set early, structure that models can read as well as people, expertise in the right place, and a team that responds to change. Across all five shifts, this is what we mean by building right, so the speed is worth having.
Which of these five are you working through in your own design system, and which shift is the hardest for your team right now? I would love to hear your thoughts and compare notes, so feel free to get in touch with the YLD team, or reach out to me directly.


