Your design system matters more in the AI era, not less

There is a tempting story going around that AI makes design systems irrelevant. If a model can render any interface on demand, why bother maintaining tokens, components, and guidelines. The honest answer is the opposite. The more of your interface is generated, the more your design system decides what good looks like. AI does not invent quality. It amplifies whatever foundation you give it.

Three shifts every SaaS team is facing

1. APIs are the new product surface

AI agents do not click buttons or navigate menus. They call APIs. If your most important actions are only available behind a modal or a multi-step wizard, an agent cannot use them, and increasingly will route around your product entirely. The bar is now clean, complete, well-documented endpoints for every meaningful action a human can take. Your API is no longer a back office, it is the front door for a growing share of your users.

2. Design systems are a force multiplier, not overhead

When AI assembles screens on demand, the components, tokens, and patterns you maintain become the vocabulary the model speaks. A tight system with clear naming, predictable spacing, and a small set of well-tested primitives lets the model produce interfaces that feel cohesive every time. A loose one produces drift, inconsistency, and the slow erosion of trust. The same prompt against a strong system and a weak one yields visibly different products.

3. You now design for two users at once

Every product has two audiences now. The human, who needs trust signals, undo, audit trails, and a clear sense of what is happening on their behalf. The agent, which needs structured data, stable identifiers, idempotent endpoints, and machine-readable error messages. The same workflow often needs both surfaces in parallel: a confirmation screen for the person, a JSON response for the agent. Treating them as equal first-class users from day one is the new default.

Why a strong design system is the highest-leverage investment

Imagine two teams building competing products. Both use the same model to generate parts of the interface. Team A has spent the last year hardening their design system: documented tokens, accessible components, clear states, written guidelines for spacing and density. Team B has shipped quickly and accumulated dozens of one-off styles. Hand the same prompt to both. Team A gets a polished, consistent screen the user trusts immediately. Team B gets something that looks plausible at a glance and starts to feel off the longer you use it. The model is the same. The ceiling is not.

• Tokens that name colour, spacing, radius, and motion in plain English
• A small set of primitives that handle 80 percent of layouts: card, list, table, form, dialog
• Documented states for empty, loading, error, success, and partial data
• Accessibility built in, not bolted on, so generated screens never ship inaccessible defaults
• A short written voice and tone guide so generated copy stays in your brand

What this means for designers

The pixel work shrinks. The judgment work grows. When the model can render a passable screen in seconds, the most valuable thing a designer does is decide what should and should not be generated, what needs a human in the loop, and what the underlying system should make easy by default. Taste, restraint, and a deep understanding of the user's mental model become the moat. The designer's job is to make complex tasks feel obvious, and then to encode that obviousness into the system the model uses.

The keyboard freed us from the typewriter, the plow freed us from the spade. AI frees us from building screens. What we still own is what to build, and why it matters.

— paraphrased from the original talk

Frequently asked questions