Why Chat Isn’t Enough: Reimagining the AI User Journey

Chat interfaces have become the default interaction model for AI. It makes sense—after all, large language models are trained on text, so chatting with them feels intuitive and natural. But just because chat is common doesn’t mean it’s optimal.

As we build more AI-powered products, I believe we’ll see the rise of hybrid interfaces—combinations of conversational input and custom UI components that help users clarify intent, supply richer context, and get better results.

The Problem with Pure Chat

Open-ended chat can be inefficient and prone to failure:

• Users often don’t know how to phrase prompts correctly.
• Models rely heavily on nuanced instructions to perform well.
• When the model needs more context, it typically responds with a follow-up question—which can lead to a back-and-forth that quickly becomes frustrating.

For example, let’s take a look at OpenAI's DeepResearch. The agent will take the user’s query and ask a series of follow-up questions to clarify the task. In the current experience, users need to write long, structured replies in a single message, and hope that the agent can associate the question response to the right question. This is not a great a great experience, and can create confusion for users.

What Hybrid Interfaces Can Solve

Imagine a user chatting with an AI assistant to complete a task—say, creating a project brief. The assistant’s task is triggered via chat, and then generates a dynamic form to capture specific details like deadlines, target audience, or tone. The user clicks and interacts with the application instead of typing paragraphs.

This model improves:

• Context capture: Structured inputs reduce ambiguity.
• Reliability: Shared responsibility between model inference and user-driven, guided structure.
• Efficiency: Fewer miscommunications, more effective task outcome.
• Cost Implications: Reduction of inference cost

Why This Matters

Even the most advanced models today don’t always reason in ways humans expect. When they go off-course, there isn’t always a clear failsafe. Embedding users into task execution allows users to mitigate the risk of models going off in the wrong direction.