Designing UX for AI: Trust, Explainability, and Fallback States

If you treat an AI feature like a standard API call, the user experience will be abysmal.

When a user clicks a button to load a standard dashboard, the server responds in 200 milliseconds. When a user asks an LLM a complex question, the model might take 4 seconds to think before it returns the final answer.

If you show a user a spinning loading wheel for 4 seconds, they will assume your product is broken and abandon the session.

AI UX requires a completely new set of design patterns focused on managing latency, building trust, and handling non-deterministic failures. Here are the core principles Product Managers must enforce when designing AI interactions.

1. Latency UX: The Streaming Illusion

You cannot eliminate LLM latency, but you can manipulate the user's perception of it.

• Streaming Tokens: Never wait for the entire response to generate before showing it to the user. You must stream the tokens (words) to the UI as they are generated. This gives the illusion of an immediate response, even if the entire answer takes 10 seconds to finish typing out.
• The "Thinking" State: If you are using a multi-step RAG pipeline (e.g., retrieving documents, then passing them to the LLM), the system will have a delay before the first token streams. You must design a dynamic "Thinking" state. Show exactly what the system is doing: "1. Searching database... 2. Reading 5 documents... 3. Synthesizing answer..." Transparency reduces frustration.

2. Explainability: The Citation Requirement

Generative AI suffers from a massive trust deficit. Users inherently assume the AI might be hallucinating.

If your AI outputs a fact, it must explicitly prove where it got that fact.

• The Citation Pattern: If your product uses RAG, the UI must include inline footnotes or hover-cards that link directly back to the source document. If the AI says, "The PTO policy allows 15 days," the user should be able to click a superscript number next to that sentence and see the exact PDF highlighting that rule.
• Confidence Scores: For critical enterprise workflows (like legal or medical AI), display a confidence score. If the model is 60% confident, color the text yellow to signal caution.

3. Designing Graceful Fallback States

Standard software either works, or it throws an Error 404. AI software can fail silently by confidently giving a wrong answer.

You must design "Fallback States" for when the AI is unsure.

• The Deflection Pattern: If the AI's confidence score is low, or if the RAG system finds no relevant documents, do not let the AI guess. The UI must explicitly say: "I could not find a definitive answer in your uploaded documents. Here are three related topics I did find, or you can click here to escalate to human support."
• The Human-in-the-Loop Handoff: For complex workflows (like AI drafting an email to a client), design the UI so the AI never sends the email automatically. It should populate a draft state that requires a physical human click to "Approve and Send."

4. Steering and "Editability"

When the AI gives a bad output, the user’s immediate instinct is to edit it.

• Prompt Steering UI: Do not make the user rewrite their entire prompt if the output is slightly off. Provide UI controls (sliders or toggle buttons) that adjust the tone ("Make it more formal," "Make it shorter") without requiring them to type a new prompt.
• Direct Text Editing: The generated output should always be a mutable text box, not static text. Let the user delete the paragraph they hate without regenerating the entire response.

Designing AI is not about hiding the machine behind a magical curtain. It is about exposing the machine's reasoning so the user feels like a confident collaborator, rather than a suspicious observer.

---

External References

• PracHub: AI PRD UX Section
• PromptLayer: LLM PM Competency

Related Reading

• AI for User Research: From Transcripts to Insights in Minutes
• How to Conduct User Interviews That Reveal Real Insights
• The Mom Test for Product Managers
• Writing PRDs for AI Products: The New Framework
• RAG Systems: What Every PM Building AI Products Must Know

Elevate Your PM Career

Are you ready to test your product sense and see where you stand in the AI era? Take the ORLOG PM Assessment to get your personalized growth roadmap and discover your PM archetype.

---

FAQ

Why is a spinning loader bad for AI products?

A spinning loader sets the expectation of a deterministic, fast database query. When an LLM takes 5-10 seconds to generate an answer, the user assumes the application has timed out. Streaming text changes the paradigm from "loading data" to "having a conversation," which has much higher latency tolerance.

What is 'Prompt Injection' and how does UX prevent it?

Prompt injection is when a user maliciously tells the AI to ignore its instructions. UX cannot entirely prevent it (that requires backend guardrails), but good UX restricts the user's input. Instead of a wide-open text box, use drop-downs or strict character limits to constrain what the user can ask.

Should we give our AI persona a human name?

Generally, no. Giving an AI a human name (like "Ask Dave") creates a false expectation of human-level empathy and reasoning. Name it descriptively (e.g., "Financial Analyst Assistant") to correctly set the expectation of its capabilities.