AI chat and assistant UX checklist

NN/g recommends informative opening messages and suggested prompts that show the chatbot's capabilities and scope. An empty box gives no hints. NN/g: 10 Guidelines for Designing Your Site's AI Chatbots

NN/g found prompt suggestions work when they are contextually relevant, personalised, and matched to the task and the person's experience level. NN/g: Prompt Suggestions

Clickable suggestions cut typing and make the next step obvious. NN/g: 10 Guidelines for Designing Your Site's AI Chatbots

NN/g recommends consolidating AI chat with other chat features; multiple bots confuse people about which one to use. NN/g: 10 Guidelines for Designing Your Site's AI Chatbots

Losing the conversation on navigation forces people to start again. NN/g: 10 Guidelines for Designing Your Site's AI Chatbots

W3C: placeholder text is not a replacement for labels, screen readers do not treat it as one, and browsers usually show it below minimum contrast. W3C WAI Forms Tutorial: Instructions

People paste long text into assistants; a single-line box hides most of what they are sending.

A rejected file after a long upload is a dead end unless the message says what would work.

NN/g lists voice input as an option that suits accessibility needs and personal preference. NN/g: 10 Guidelines for Designing Your Site's AI Chatbots

Retyping a long prompt after a network error is the most avoidable frustration in chat.

Apple suggests specific messages such as “Summarizing key themes from your notes” over a vague “Processing…”. Apple HIG: Generative AI

People often see within a few lines that the answer is off track. Stopping should keep what was generated so far.

NN/g recommends keeping the scroll position at the start of a response so people read it from the beginning. NN/g: 10 Guidelines for Designing Your Site's AI Chatbots

Status changes should be announced without moving focus (WCAG 4.1.3). Announcing each token as it streams makes the page unusable with a screen reader. WCAG 2.2: Understanding Status Messages

NN/g recommends progressive disclosure to keep the conversation short and scannable. NN/g: 10 Guidelines for Designing Your Site's AI Chatbots

NN/g recommends images, not only links or text descriptions, for products and instructions. NN/g: 10 Guidelines for Designing Your Site's AI Chatbots

Apple recommends surfacing controls such as Edit, Undo, Retry or Adjust near generated content. Apple HIG: Generative AI

Apple suggests offering alternative versions where choice helps bridge the gap between the model's interpretation and what someone wants. Apple HIG: Generative AI

NN/g recommends placing sources directly next to the specific claim they support. NN/g: Explainable AI in Chat Interfaces

Checking a claim should take one click. NN/g suggests deep links to the referenced passage or a preview of the source. NN/g: AI Chatbots Discourage Error Checking

NN/g advises against vague link labels such as “Source”, and recommends styling citations distinctly from the answer. NN/g: Explainable AI in Chat Interfaces

NN/g recommends plain-language disclaimers paired with an action, placed near the input box and included in onboarding. NN/g: Explainable AI in Chat Interfaces

Microsoft's G15: enable people to give feedback during regular interaction. Microsoft HAX: Guidelines for Human-AI Interaction

PAIR recommends feedback questions at a level of detail the team can actually use to improve the model. PAIR: Feedback + Control

PAIR suggests being specific about timing (now or later) and scope (just you or everyone) when acknowledging feedback. PAIR: Feedback + Control

People rate an answer without expecting their whole chat to be read. Apple asks for explicit permission when sensitive data is used for model improvement. Apple HIG: Generative AI

Microsoft's G12: remember recent interactions and let people refer back to them. Microsoft HAX: Guidelines for Human-AI Interaction

Microsoft's G17 calls for global controls over what the AI monitors and how it behaves; long-term memory is exactly that. Microsoft HAX: Guidelines for Human-AI Interaction

Without this, personalised answers look like the model guessed private details, which feels intrusive.

Delete should mean what people think it means, or say what it doesn't do.

People ask sensitive questions they don't want kept. Apple asks for the minimum data needed and a clear way to opt out. Apple HIG: Generative AI

NN/g recommends letting people email, download or share chat content worth keeping. NN/g: 10 Guidelines for Designing Your Site's AI Chatbots

NN/g recommends letting people resize the chat window. NN/g: 10 Guidelines for Designing Your Site's AI Chatbots