Stripe Chief Design Officer Katie Dill on How to Use AI to Scale Intent, Quality, and Artistry

Stripe Chief Design Officer @lil_dill shares: How can we use AI to scale intent, quality, and artistry?

# Three warning signs

1. LLMs naturally gravitate toward the mean. Large models are good at giving the "most likely answer"—what has been popular in the past, what has worked before—but they are not good at giving original solutions that belong to your brand and your users' context. She showed a completely characterless website and had the audience guess whether it was selling software or Korean BBQ, confirming that what AI produces by default is exactly this kind of "personality-free" thing. 2. The temptation of "a sense of completion." AI makes things appear "finished" extremely quickly. She mocked herself with a microwave frozen burrito: 90 seconds from hungry to lunch; you know it tastes bad but you compromise because of speed. Likewise, type one sentence and you get an interface with rounded corners and shadows. "Looking polished" does not equal "actually solving the problem, actually being differentiated." 3. Work becomes disposable. Output is too easy, so responsibility and long-term thinking (who will maintain it?) are easily set aside.

The result of these three combined is "zombie UI": monotonous, hollow, and something no one truly cares about. And we spend half our waking hours staring at screens; what users want is software that treats them seriously, even down to details as small as a calendar tab showing the correct date.

# Four practical suggestions from Dill

1. You must have your own point of view. Without a point of view, AI will supply one for you, and that answer is very likely generic and backward-looking. In an era when the power to build is distributed and everyone contributes to products, teams must first align on "who we are and what users care about."

She gave an internal design review at Stripe as an example. A colleague asked: For things made by AI, what should the quality standard be? The answer to that question is simple: users do not care how something was made, only whether it is good, so the standard should look only at the final output. That review ultimately produced 17 improvement points, such as adjusting the frost effect a bit more, softening the edges a bit more, making the bubbles a bit smaller. "Pepsi bubbling" has since become a byword at Stripe for extreme polish. The goal is not perfection, but going one layer deeper than what users can perceive; users will eventually feel that craftsmanship.

How do you cultivate a point of view? Practice awareness. Pay attention to what users truly need, not just what they say; notice what is good and what is mediocre in life; draw references from art and science, and gradually build your own taste and standards.

2. Encode standards into the machine. The reality is that designers will not be present at every decision point. Interfaces are generated in rooms without designers, agents automatically fix problems in the middle of the night, and generative UI assembles in real time. So the object of the design system has changed: in the past it was the interface; now it is the system itself.

She used Gutenberg as an analogy. Gutenberg made 290 character variants, including different widths, ligatures, and abbreviations, in order to make machine typesetting achieve the quality of hand typesetting and avoid ugly white gaps when text was justified. Old design systems scaled consistency; new systems must scale intent.

Stripe's practice also took detours. At first they used MCP to let AI read design documents, and the results were very poor: three people would get three different results from the same prompt. Later they iterated into a CLI built on top of the design system: it runs where developers work and injects documentation at the right moment to avoid context rot; the content it provides was upgraded from atomic components to complete templates and flows, so the system knows what the product should look like and how it should be assembled.

But she also quoted Christopher Alexander as a reminder: "A system can satisfy all the rules and still be dead." This point is only the passing line.

3. Refuse to treat "done" as "good." In the past, quality filtering was built into the process: only 1 of 20 ideas could be greenlit, and it was trimmed and polished all the way through. Now 20 ideas can be built in a week. The benefit is that you can skip theory meetings and look directly at real things, but filtering has been postponed until after building, and that is exactly the moment when it is hardest to say no.

So the role of the "editor" has become more important than ever. Who is reviewing the product end to end to see whether it is a complete whole? Teams must break the habit of thinking "built equals done, done equals good."

The editor's core method is to experience it personally like a user: Does it really solve the problem? Does it fit the user's way of thinking? Is it still coherent when placed in the full user journey? Many things look good on their own, but look disjointed when put together. The editor must be responsible for every decision and every pixel.

She quoted Nabil Khureshi's essay on art: Great art lies in unexpected details and deep themes that run throughout, and this is exactly where AI falls short. What is most offensive about AI slop is the feeling that "details don't matter." The cup is green, but it would be fine if it were blue; users can see that attitude at a glance.

The case came from the opening animation of this conference. Designer Stefan used AI to iterate starting from a 3D scene model. The first version was pretty cool, but when examined at the pixel level, something always felt off. He made 56 versions before getting a final product with real vitality. Her key insight is: without AI, he might never have tried such a complex scene and so many perspectives; AI expanded the space of possibilities, but getting it right in one pass is an illusion. Human judgment is still what gates the work.

Today's interface paradigms—chat boxes, charts, command lines—are by no means the endpoint of interaction design. We should invent new interfaces and new aesthetics. Best practices have not yet been written, and now we have the tools to write them, just as multitouch gave rise to entirely new interactions and synthesizers gave rise to entirely new sounds.

Stripe uses AI to amplify its creative team. The cover of the new book "Built to Grow" began with a stunning base artwork by a human marbling artist, then AI finely adjusted the colors and lines—essentially scaling up good human judgment.

There are three ways to make AI a stronger creative partner: improve the input by including brand beliefs, aesthetic standards, and source materials in the prompt, rather than just "make me a BBQ restaurant website"; pressure the output—do not fall into the burrito trap, always push one step further; and use adversarial agents to help find flaws.

The real difficulty is culture. Following the cookie-cutter path is safe and comfortable; finding something unique that improves the status quo is much harder. Leaders should leave room for teams to explore and protect the strange. AI lowers the cost of creation, and the budget saved should be spent on making truly special things. If you only use AI to do existing things faster, you miss the most interesting part. AI can raise the floor, but the ceiling depends on people holding it up.