UX Design for AI Products: Patterns for Onboarding, Trust, Errors and Agents

The short answer

Good UX design for AI products comes down to five things: set clear expectations about what the AI can and cannot do, show where answers come from, make outputs easy to accept, edit or undo, handle mistakes honestly, and keep people in control when an agent acts for them. Microsoft's 18 Guidelines for Human-AI Interaction, Google's People + AI Guidebook and Apple's guidance on generative AI all point the same way. The best AI products feel less like magic and more like a capable colleague you can check.

Why UX design for AI products is different

Traditional software gives the same output for the same input. AI does not. Apple's Human Interface Guidelines note that with generative AI, small changes to inputs, or even the same input given twice, often produce very different outcomes, and you cannot always predict what people will ask. That changes the designer's job. You are no longer only designing screens; you are designing for uncertainty.

1. Onboarding: set expectations early

An empty prompt box is intimidating. People do not know what to ask, or how good the answer will be. The first two of Microsoft's guidelines are to make clear what the system can do and how well it can do it.

• Suggest starting points. Apple recommends curated suggestions for open-ended features like a search bar or generation prompt.
• State limits up front. If the feature is weak at something, say so, and show how to get good results.
• Use structured inputs where you can. A form with three fields often beats a blank prompt for a specific task.
• Show a real example. A finished output on the first screen teaches more than a paragraph of text.

2. Trust and explainability: show your working

People trust AI output more when they can check it. Microsoft's guidelines include making clear why the system did what it did. In practice:

• Cite sources. Perplexity links every answer to numbered citations that you can hover to preview or click to read.
• Show inputs. Tell people which documents, data or context the AI used.
• Show confidence carefully. A plain label such as "needs review" is more useful than a precise-looking percentage most people cannot interpret.
• Label AI content. Apple advises that you clearly identify when and where you use AI, and never let people think they are dealing with a human when they are not.

3. Outputs: easy to accept, edit or reject

The best AI features produce drafts, not final answers. GitHub Copilot shows code suggestions as greyed text in the editor: press Tab to accept, Esc to dismiss, or accept word by word. That pattern works far beyond code. Microsoft's guidelines call for efficient invocation, dismissal and correction, and Apple says people should be able to dismiss, revert or retry AI changes they do not agree with.

• Make the output editable in place
• Offer regenerate and undo next to every result
• Let people accept part of an output, not just all or nothing
• Keep the original visible when the AI rewrites something

4. Errors and graceful failure

AI will be wrong sometimes. OpenAI's help centre says ChatGPT can produce incorrect or misleading outputs and asks people to verify important information. Design for that from the start:

• Say what went wrong. "I could not find that in your documents" beats a confident wrong answer.
• Offer a way forward. Suggest a better prompt, a manual path or a human contact.
• Keep a non-AI fallback. Apple recommends a non-AI fallback where possible, so people can still finish the task.
• Make feedback easy. A thumbs up or down with an optional reason, placed right on the output.

5. Agents: keep people in control

When AI moves from answering to acting, the stakes rise. An agent that books, sends, buys or deletes needs clear guardrails. OpenAI's launch post for ChatGPT agent described several: it asks permission before actions of consequence, people can interrupt, take over the browser or stop a task at any point, and certain critical tasks such as sending emails need active oversight.

• Show the plan first. List the steps the agent will take before it starts.
• Confirm before anything irreversible. Payments, emails, deletions and bookings need a clear yes.
• Show progress live. A visible log of what the agent is doing builds trust.
• Make stopping easy. One obvious control to pause or take over.
• Summarise at the end. What was done, what was not, and what needs a human.

6. Privacy and data

Tell people what data the AI uses, where it is processed and whether it is used for training. Apple's guidance is to process as much as possible on the device and make sure people know when their information may be sent to a server. Put these controls in settings, and mention them at the moment of first use.

A checklist for AI features

• The feature explains what it can and cannot do on first use
• Open-ended inputs come with suggested starting points
• Answers show sources or inputs where possible
• AI-generated content is labelled
• Every output can be edited, regenerated or undone
• Errors explain what happened and offer a next step
• There is a non-AI path to finish the task
• Agents confirm before irreversible actions and can be stopped
• Data use is explained in plain language

Where to learn more

Microsoft's HAX Toolkit lists all 18 guidelines with examples. The original edition of Google's People + AI Guidebook has chapters on mental models, explainability and trust, feedback and control, and errors and graceful failure. Apple's Human Interface Guidelines have a page on generative AI. All three are free.

We design AI products and the sites that sell them. For Breakout, an inbound AI SDR, we did the product design, website and launch film. For Further AI, we delivered brand, website, motion and a Webflow build in the window before its $25M Series A from a16z. See our Website, Product & App Design page for more.

FAQ

Should every AI product be a chat interface?

No. Chat suits open-ended questions. For specific tasks, buttons, forms and inline suggestions are often faster and clearer. Many good AI products use chat only as one entry point.

How do we show that an AI answer might be wrong?

Show sources, mark low-confidence results in plain language, and keep a short reminder to check important information. Do not hide uncertainty behind a confident tone.

How much should an agent do without asking?

Let it act freely on reversible, low-risk steps, and ask before anything that costs money, contacts someone or cannot be undone.

Where can we find a design partner for an AI product?

Our list of the best design studios for AI startups is a good place to start.

Building an AI feature and not sure how it should behave? Show us the flow, and we'll help you design the moments where trust is won or lost.