AI-ASSISTED DESIGN PRACTICE
I asked Claude Code for an AI-native interface. It built a chatbot.
Alex Thanasenaris - 8 min read
AI Workflow
Claude Code
Interaction Design
Enterprise UX
I've spent most of the last several years in design leadership — setting strategy, reviewing other people's work rather than making my own. Lately I've been rebuilding a more hands-on practice, using Claude and Claude Code to move faster between strategy, design, and working prototypes than a traditional workflow allows, while keeping an eye on how the rest of the field — from research groups to teams shipping agentic products at scale — is converging on the same open questions.
The setup
I was prototyping an HR and payroll admin console — a place where an administrator could describe what they needed in plain language instead of navigating a dozen settings screens for something non-routine, like onboarding a new hire with a custom benefits waiting period. My brief to myself was explicit: no chatbot bolted onto the side of a product. The real design problem was figuring out where natural language should change how the interface works.
"Build the ask surface for an AI-native HR/payroll admin console... Include a chat-style interface for the interaction... When the admin submits their request, show the AI confirming what it did."
Read that again. I asked for something AI-native, explicitly not a chatbot, then described the exact shape of a chatbot in the same breath.
What came back
A chat window. Bubbles on the right for what I typed, bubbles on the left for the reply. I asked it to set up a new hire's benefits with a non-standard waiting period, and got: "Done! I've set that up for you." No preview. No effective date. No specific field. Nothing to review before it executed.
What the rest of the field has been finding
I didn't land on this in isolation. Nielsen Norman Group has described the shift from command-based to intent-based interaction as the first genuinely new UI paradigm in six decades, and a growing body of "agent UX" writing argues that agent interfaces aren't chatbot skins with an AI label attached — they're a distinct design discipline built around transparency and override controls.
What I changed, and why it mattered
A structured confirmation replaced the chat reply. Disambiguation became a real state. The traditional interface stopped being decorative. An audit trail appeared.
The actual takeaway
"AI-native" doesn't mean putting a chat box on a product. It means deciding, deliberately, where language is the right tool for expressing intent and where structure is what earns trust — then building both on purpose.
SOURCES REFERENCED
Nielsen Norman Group — on command-based vs. intent-based interaction
Agent UX writing — on agent interfaces as a distinct design discipline
Academic research — on confirmation timing in multi-step agentic tasks