AI
The New Triangle: User, Model, and UI
For the first time, three actors operate simultaneously: the user, the model, and the UI. The model is an active participant, not a passive bystander. It orchestrates the conversation and reaches for UI components as capabilities. This changes the fundamental relationship from the web apps of the last 20 years. The design challenge is learning to work with this triangle instead of fighting it.
Design for Capabilities, Not Your Entire Product
The biggest mistake Beavers sees is companies trying to squeeze their entire dashboard into an MCP app. It never works well. The framing is wrong. She advises asking what users are trying to accomplish, not what the product does. Start with the essential capabilities. OpenAI’s submission guidelines say the same thing: see your app as capabilities the model can orchestrate, not a mini version of your product.
The Model as a 10x Product Multiplier
Models excel at three things: intent detection, generation, and extraction. The Cottages app lets users type “a vacation home at the beach for five people, three dogs, and two cats” without manual filtering. The Excalidraw app uses the LLM to draft content, then lets the user iterate. A phone buyback app extracts the phone’s condition from a photo. These capabilities solve user problems that were previously impossible.
Prepare for Composability
Your app is no longer a walled garden. Beavers urged builders to prepare for a larger ecosystem where the model composes different systems. A luxury brand could surface a suitcase with an island background if the conversation already mentions a vacation in Greece. An analytics company should return enough data for the model to handle follow-up queries or push data into another system.
UI and Model Are One Experience
The UI and model must be complementary and synchronized. ChatGPT often renders a widget and then summarizes the same information below. Beavers called this “stupid” but common. To fix it, give the widget a detailed meta description and use the tool output only to inform the model. Synchronization requires using navigate functions and OpenAI’s set widget state to persist the user’s location back to the model.
Notable Quotes
“Don’t just throw away everything that you know about good product and UX design just because we’re in this new brave new world of models.” Erica Beavers · ▶ Watch (9:07)
“The model is not going to compensate for bad UX design. It’s actually going to amplify it.” Erica Beavers · ▶ Watch (9:22)
“See your app as capabilities that the model can orchestrate and not a mini version of your product.” Erica Beavers · ▶ Watch (11:14)
“You can throw everything away that I just said and just create something really dumb and viral.” Erica Beavers · ▶ Watch (24:22)
Key Takeaways
- Ship capabilities, not entire products, into MCP apps.
- Use the model for intent detection, generation, and extraction.
- Keep the UI and model synchronized to avoid redundant information.
About the Speaker(s)
Erica Beavers is a co-founder of Alpic, a company that helps businesses expose their products to AI via MCP. She previously built Streamroot, a video delivery startup acquired by a major US telco. Her focus is on making agent-native apps easy.