Blog · Chapter 05 · UI development
Interfaces people trust and can correct
Chat is not the only answer. Choosing the right interaction pattern and designing for trust, streaming and correction.
The interface decides whether an AI application gets used. Chapter 5 argues that chat is only one option, and often not the best one, and that trust has to be designed in.
Pick the interaction pattern that fits the task
| Pattern | Best for |
|---|---|
| Embedded AI features | Workflows that already have an application: a "summarise" or "draft reply" button where people work |
| Review inbox | High-volume processing, where AI prepares cases and people approve, edit or reject them |
| Copilot sidebar | Assistance next to an existing tool, with context from the current screen |
| Chat | Exploration and open questions over documents and data |
| Voice | Phone channels and hands-free situations |
A claims team that processes hundreds of emails a day is better served by a review inbox than by a chat window.
Design for trust
Users trust an AI system that shows its work and lets them fix it:
- provenance: citations and links to the source documents;
- uncertainty: confidence indicators and highlighted fields that need attention;
- editability: every AI output can be changed before it is used;
- feedback: thumbs, corrections and reasons, captured as data for improvement;
- graceful failure: clear messages and a manual path when the AI cannot help.
Watch out for automation bias. When the system is right most of the time, reviewers stop checking. Highlight the uncertain parts and measure review time and override rates.
Streaming and long-running work are architecture
LLM responses take seconds and agent runs can take minutes. Stream tokens to the UI (server-sent events or WebSockets), show progress for multi-step runs, and make runs resumable, with state owned by the backend. A user who closes the tab should not lose the result. Plan this from the start; it is hard to add later.
Prototype fast, then build for production
Python UI tools such as Streamlit or Gradio are good for prototypes and internal tools. For customer-facing or high-usage applications, plan a production frontend with proper authentication, accessibility, state management and design.
Key takeaways
- Pick the interface that fits the task, not the trend.
- Design for trust: sources, uncertainty, easy correction.
- Streaming and resumable runs are architecture, not polish.