AI & Development
Designing a good AI chat experience goes well beyond the prompt. Learn the UX patterns, error states, and trust signals that separate professional AI
Conversational AI interfaces have become a dominant UI pattern across products: customer support, search, documentation, productivity tools, and consumer applications all now offer chat-based interactions with AI. But the early wave of AI chat features was largely undifferentiated - a text input, a streaming response, and minimal attention to the design details that make the experience feel trustworthy and useful. As the market matures, the design quality of the conversational interface is increasingly a differentiator.
Users form incorrect mental models of AI capabilities and then feel betrayed when the reality does not match. The most common mismatch: users expect the AI to remember previous sessions, understand context it was never given, or have access to information it does not have. Setting expectations before the first interaction - through onboarding text, capability descriptions, or starter prompts - reduces frustration and builds appropriate trust.
Effective expectation setting is specific and honest: "I can help you find information in our product documentation. I don't have access to your account data or order history." This is more useful than vague capability descriptions that overpromise and underdeliver.
LLM response latency is noticeably longer than users expect from typical app interactions. A 2-3 second wait for the first token to appear feels abandoned in the context of a conversation. Streaming - showing tokens as they are generated - dramatically improves perceived latency even when wall-clock time is the same. The user sees immediate activity and understands that something is happening.
For tasks that require longer processing - complex analysis, multi-step retrieval, large document processing - communicating the expected duration ("Analyzing your document, this will take about 30 seconds") manages expectations and prevents users from assuming the application is broken. Progress indicators that reflect actual progress rather than a spinning indicator build confidence during longer waits.
AI systems are uncertain in ways that traditional software is not, and designing for this uncertainty is critical. When the model is uncertain, the interface should reflect that uncertainty - not paper over it with confident-sounding language. "Based on the information I found, the return window is 30 days, but I'd recommend confirming this with customer support" is more trustworthy than "Your return window is 30 days" when the model is not certain.
Providing an easy path to escalation - a "Talk to a human" button, a link to mock documentation, a contact form - gives users recourse when the AI's response does not fully meet their needs. Users who have a clear escalation path are more tolerant of AI limitations than users who feel trapped in a chat interface with no alternative.
Long paragraphs of text are hard to scan in a chat interface. Responses that use headers, bullet points, numbered steps, and short paragraphs are significantly more usable than wall-of-text responses to complex questions. Configuring the model to use formatting - and rendering Markdown in the chat UI - improves readability substantially.
At the same time, over-formatting simple responses creates friction. A conversational reply to a simple question should not have headers and bullet points. Matching the response format to the nature of the question - prose for conversational queries, structured formatting for how-to questions or comparisons - requires prompt design that teaches the model to make this distinction.
For AI applications that make factual claims - support agents, documentation assistants, research tools - showing the sources behind the response dramatically increases trust and reduces the frustration of hallucinated facts. "Based on the product documentation (section 3.2)" gives the user a way to verify the claim and signals that the system is grounded in authoritative information rather than generating from training data.
Source attribution also shifts the verification burden appropriately: when sources are shown, users know how to check a claim they are uncertain about. When there are no sources, users cannot know whether to trust the response or how to verify it. For business-critical information, the difference matters.
Users increasingly expect AI applications to remember them across sessions - their preferences, their history, their context. Designing how the system uses persistent memory, and giving users visibility into and control over what is remembered, is both a feature and a trust concern. Users who cannot see what the AI knows about them and cannot correct or delete that information are rightly uncomfortable. Transparency about memory, and easy controls to manage it, are table stakes for applications that use persistent context.