How to design for the future of LLMs

Most AI tools today rely on open-ended chat interfaces that leave users overwhelmed or confused. Maggie Appleton, lead designer at Elicit, argues for a different approach: building focused, task-specific tools powered by AI behind the scenes. This article explores how vertical design, real coding skills, and a shift toward sense-making can unlock AI’s true potential.

Why chat-based AI isn't enough

Most AI interfaces today are little more than dressed-up chat boxes. You type something in, get a response, and hope it’s useful. According to Maggie Appleton, lead designer at Elicit, this model is fundamentally broken.

“If you tell a user they can do anything, they will do nothing,” she says. “They’ll type something like ‘what’s the capital of Uganda’ and then close the tab.”

Instead of chasing a general-purpose assistant that tries to do everything, Appleton and the Elicit team took a different path. They built a focused tool that helps researchers review thousands of scientific papers in minutes instead of months. It’s not about one super-smart model. It’s about breaking big tasks into small, checkable steps.

Appleton draws a distinction between "horizontal" tools that try to serve everyone and "vertical" ones that serve a specific purpose. ChatGPT is horizontal. Your dentist’s office software is vertical. It’s built for a clear task, a clear user, and a clear outcome.

“This is where AI will really shine,” she says. “In complex workflows where it quietly helps people move faster by suggesting patterns, surfacing data, and reducing grunt work.”

Designing with AI means working differently

At Elicit, AI doesn’t handle everything in one go. Instead, the system breaks research down into a series of small decisions. For example, one model might ask: “Does this paragraph mention childhood malnutrition?” Another then extracts the finding. A third rates how confident it is in that extraction. Each step is transparent and easy to verify.

That design philosophy shapes how the team builds, too. Appleton doesn’t just sketch ideas in Figma. She writes production code.

“You can’t fake AI behavior in a design tool,” she explains. “You need to work with the real thing. The actual responses matter.”

Her advice to designers is direct: learn JavaScript, HTML, CSS, and React. Use tools like GitHub Copilot. Don’t just imagine how AI products might behave. Build them and find out.

While many teams are focused on generation—AI that creates text, images, or code—Appleton believes the bigger opportunity is in sense-making. Tools that help people understand and navigate existing information will matter more than those that flood us with even more content.

“We don’t need more noise,” she says. “We need ways to find the signal.”

The future belongs to focused, useful tools

Elicit doesn’t generate summaries. It helps researchers follow their existing process faster, with smart assistance at each step. That kind of utility is what Appleton believes the best AI tools will offer: not novelty, but clarity, speed, and confidence.

Looking ahead, she sees basic development work becoming highly automated. More people will build software without writing much code. But that won’t necessarily lead to better design.

“Founders won’t hire designers if they think they can prompt their way to a working app,” she says. “Designers will become more valuable at the extremes. They will either focus on strategy and research or on exploring new interface patterns. The routine stuff, like buttons and menus, is already being automated.”

Ultimately, the best AI interfaces won’t feel like AI at all. They will just feel like tools that work.

“You don’t tell users something is built in Python,” Appleton says. “Why would they care that it’s AI? They care about getting something done and trusting the result.”

As AI capabilities grow, the designers who understand this will shape the future. The rest will still be making chat boxes that no one uses.

Listen to the full episode here.

Source: Dive club podcast