Introduction
AI UX patterns are reusable design solutions for problems that only AI products have: answers that can be wrong, actions the AI takes for a user and usage that runs out. The eight patterns below cover trust, control and usage limits, with examples from Gemini, ChatGPT, Cursor and Midjourney that you can check yourself.
Key Takeaways
-
AI products need patterns for three problems standard UX does not cover: unreliable answers, actions taken on a user’s behalf and usage that runs out.
-
Every pattern below is shown in a real product, so teams can see it working before they design their own.
-
Usage limits deserve the same design attention as trust, because a limit that stops a task or asks for an upgrade too early can end a user’s first session with a feature.
-
Start with the patterns that match what your AI does: act, answer, remember or meter usage.
What Are AI UX Patterns?
AI UX patterns are reusable solutions for designing AI features. They differ from standard UX patterns because AI outputs can differ from one request to the next, the AI can act for the user and usage can be limited. Teams also call them AI design patterns or AI UI patterns, and generative AI design patterns when the feature creates text, images or code. Once your team settles on the patterns it needs, they belong in your design system as a shared AI UX pattern library, so every AI feature handles these moments the same way.
Microsoft’s 18 Guidelines for Human-AI Interaction group advice by when it applies: upon initial interaction, during interaction, when the system is wrong and over time (Microsoft Research, 2019). The guidelines were published in 2019, and none of the 18 covers usage limits or cost, which is why this list adds three usage patterns.
The 8 AI UX Patterns at a Glance
|
Pattern |
Problem it solves |
Real example |
|---|---|---|
|
1. Show sources so users can verify answers |
Users cannot tell where an answer came from |
Gemini |
|
2. Flag claims that may be wrong |
Users cannot tell which parts of an answer to check |
Gemini |
|
3. Ask before the AI acts |
The AI takes actions users did not approve |
ChatGPT |
|
4. Make every AI action easy to undo |
AI mistakes take manual work to reverse |
Cursor |
|
5. Let users see and control what the AI remembers |
Users do not know what the AI keeps about them |
ChatGPT |
|
6. Show the cost before an action runs |
Users are surprised by how fast they use their allowance |
OpenAI’s app guidelines |
|
7. Give users options when they hit a limit |
Limits end in a dead end |
OpenAI’s app guidelines |
|
8. Keep users working after a limit is reached |
Limits stop work halfway |
Midjourney |
Product examples in this article were checked against each vendor’s own documentation on 16 November 2026. Product behavior changes often, so check the linked source before you copy a pattern.
Patterns for Trust
Trust patterns help users decide whether an AI answer is right before they act on it. In a 2026 survey of 107 senior product leaders, 44% said trusting AI output is where users struggle most with AI features (Userflow, 2026).

1. Show Sources So Users Can Verify Answers
The pattern: Showing sources is a pattern that places the origin of an answer next to the answer itself, so users can check it in one step.
Why it works:
-
Answers a common first doubt about an AI response: where the information came from.
-
Lets users check a single claim without leaving the product to search for it.
-
Moves the work of proving an answer from the user to the interface.
Real-world example: Google’s Gemini Apps help says Gemini sometimes shows sources and related content within and below its responses (Google, 2026).
Key takeaway: Sources placed beside an AI answer let users verify it in one step.
Go deeper: AI UI trust patterns
2. Flag Claims That May Be Wrong
The pattern: Flagging uncertain claims is a pattern that marks the parts of an answer that may be wrong, so users know where to check first.
Why it works:
-
Directs attention to the one statement that needs checking.
-
Turns a vague worry about accuracy into a specific item a user can verify.
-
Protects confidence in the parts of the answer that hold up.
Real-world example: Gemini’s double-check feature uses Google Search to find content that is likely similar to or likely different from statements in a response (Google, 2026). Microsoft’s HAX Toolkit uses this feature as an example of making clear how well a system can do what it can do (Microsoft, 2026).
Key takeaway: Marking the claims that may be wrong keeps users’ trust in the rest of the answer.
Go deeper: AI explainability UX patterns
Patterns for Control
Control patterns keep users in charge when the AI takes action or remembers information about them.
3. Ask Before the AI Acts
The pattern: Asking before acting is a pattern that pauses the AI before an action that is hard to reverse and asks the user to confirm it.
Why it works:
-
Removes the fear that the AI will spend money or commit to something without permission.
-
Gives users a moment to review the details while the action can still be canceled.
-
Keeps people accountable for the decisions that carry real consequences.
Real-world example: OpenAI’s release notes say ChatGPT Work always asks for confirmation before consequential actions, such as completing a reservation or payment (OpenAI, 2026).
Key takeaway: A confirmation step before high-stakes AI actions keeps users in charge of decisions with real consequences.
Go deeper: Human-in-the-loop AI design
4. Make Every AI Action Easy to Undo
The pattern: Undo for AI actions is a pattern that saves the state before the AI changes something, so users can return to it in one step.
Why it works:
-
Lowers the cost of trying the AI, because a bad result can be reversed.
-
Lets users accept larger AI changes without reviewing every line first.
-
Turns an AI mistake into a one-click fix.
Real-world example: Cursor’s documentation says its agent automatically creates checkpoints before significant changes, and users can click a checkpoint to preview their files and restore them to that state (Cursor, 2026). This matches Microsoft’s guideline to support efficient correction, making it easy to edit, refine or recover when the AI system is wrong (Microsoft Research, 2019).
Key takeaway: One-step undo makes users willing to let the AI attempt larger changes.
Go deeper: Agentic AI guardrails design
5. Let Users See and Control What the AI Remembers
The pattern: Memory control is a pattern that shows users what the AI remembers about them and lets them change or remove it.
Why it works:
-
Answers the privacy worry about what the AI keeps and for how long.
-
Explains why a response feels personal, which makes personalization easier to accept.
-
Gives users an off switch, so sharing details feels reversible.
Real-world example: OpenAI’s Memory FAQ says users can delete individual memories, clear all of them or turn memory off. Users can tap a book icon below a response to see which sources, including memories, personalized it. Temporary Chats do not use or create memories (OpenAI, 2026). This matches Microsoft’s guideline to provide global controls, letting users customize what the AI system monitors and how it behaves (Microsoft Research, 2019).
Key takeaway: Visible, editable memory makes users comfortable sharing the context that improves AI answers.
Patterns for Usage and Limits
A usage limit that stops a task halfway, or an upgrade request that arrives before the user has seen value, can end someone’s first session with an AI feature for good. These three patterns keep limits visible, fair and recoverable.
6. Show the Cost Before an Action Runs
The pattern: Showing the cost before an action runs is a pattern that tells users which options use more of their plan or credits before they choose one.
Why it works:
-
Removes surprise, which is what makes a limit feel unfair.
-
Lets users pick a lighter option when they want to save their allowance.
-
Builds the habit of checking usage before the limit arrives.
Real-world example: OpenAI’s Sign in with ChatGPT UI and UX guidelines tell apps to label options that use a person’s plan or credits faster (OpenAI, 2026).
Key takeaway: A cost label on each option lets users manage their allowance before they run out.
7. Give Users Options When They Hit a Limit
The pattern: Options at the limit is a pattern that replaces a dead-end error with clear next steps when a user reaches a usage limit.
Why it works:
-
Turns the limit into a decision point with a clear way forward.
-
Points users to the one action that resolves the problem.
-
Keeps users in the product at a point where a dead end would end the session.
Real-world example: The same OpenAI guidelines say that when a usage limit is reached, an app should offer options and make Manage usage the primary action (OpenAI, 2026).
Key takeaway: A usage limit with one clear primary action gives users an obvious next step.
8. Keep Users Working After a Limit Is Reached
The pattern: Continued work after a limit is a pattern that lets users keep going at a reduced speed or capacity once they reach a limit.
Why it works:
-
Lets users finish the task they started, which protects the work they have already put in.
-
Keeps upgrading optional, so the limit does not feel like pressure.
-
Shows the product’s value at the limit, which is often when users decide whether it is worth paying for.
Real-world example: Midjourney’s plans page says unlimited image generations with Relax Mode are available on its Standard, Pro and Mega plans (Midjourney, 2026). Its documentation says users who run out of Fast time can buy more Fast hours (Midjourney, 2026). Falling back to a slower mode lets users finish their work instead of being cut off, while keeping the upgrade optional.
Key takeaway: A slower fallback mode lets users finish their work at a limit and keeps upgrading optional.
Which AI UX Patterns to Ship First
The patterns to ship first depend on what your AI feature does.

|
If your AI feature |
Start with |
|---|---|
|
answers questions or summarizes content |
1 and 2 |
|
takes actions for the user |
3 and 4 |
|
remembers users across sessions |
5 |
|
has usage limits, credits or metered plans |
6, 7 and 8 |
Check Your AI Feature Against These Patterns
These are questions to discuss with your team, not a pass or fail score. The last column suggests one signal to track for each pattern once it ships.
|
Pattern |
Question to ask |
Sign it is missing |
What to track |
|---|---|---|---|
|
1 |
Can users see where an answer came from? |
Users copy answers into a search engine to check them |
How often users open a source |
|
2 |
Does the product show which claims may be wrong? |
Every answer is presented with the same confidence |
How often users check or edit a flagged claim |
|
3 |
Does the AI ask before actions that are hard to reverse? |
Users only discover an action after it has happened |
How often users cancel at the confirmation step |
|
4 |
Can users undo what the AI changed in one step? |
Fixing an AI mistake means redoing the work by hand |
How often users undo an AI change |
|
5 |
Can users see and change what the AI remembers? |
Users avoid sharing details because they cannot remove them later |
How often users view, edit or delete memories |
|
6 |
Do users know what an action will use before it runs? |
Users are surprised by how fast they reach a limit |
How often users pick the lighter option after seeing a cost label |
|
7 |
Does a limit come with clear next steps? |
Users meet a dead end with no way forward |
What users do right after reaching a limit |
|
8 |
Can users finish their task when they hit a limit? |
Work is lost or abandoned at the limit |
How many users keep working after reaching a limit |
When Patterns Are Not the Fix
Patterns cannot rescue an AI feature whose answers are poor or that solves the wrong problem. Patterns make a useful AI feature trustworthy and easy to control. They cannot make an unhelpful one useful.
Conclusion
AI features earn repeat use when users can check answers, stay in control and keep working within their limits. Show sources and flag weak claims so answers can be trusted. Ask before acting, make actions easy to undo and let users manage what the AI remembers. Show what an action will use, offer clear options at a limit and let work continue instead of stopping it.
Building these patterns into an existing product is the focus of reloadux’s AI feature experience design work. Teams adding AI to an existing product can start with reloadux’s 2-day trial on one workflow, a focused UX review of one core flow plus one or two redesigned screens. Teams building a new AI product can try one use case in a day, free.
FAQs

Shahmir Farooq
Sr. Communication Designer





