Unique patterns in AI design
AI design may not have invented many new components yet, but it is reshaping familiar patterns around guidance, oversight, and trust.
I haven’t seen many entirely new atomic-level components emerge in AI design yet. But I have noticed a set of recurring patterns across AI products. Most of them solve the same few problems: helping users get started, making the system’s work visible, and building enough trust for people to let AI do more.
Wayfinders
AI products are becoming more familiar, but most users are still learning how to use them well. They are not thinking in prompts, agents, or models. They are trying to get something done.
That makes wayfinding one of the most important areas of AI product design. Good AI interfaces don’t expect users to write the perfect prompt. They help users understand where to begin, what context to provide, and how to move from vague intent to useful output.
Initial CTA
The input box has become the default starting point for AI products. It is flexible, familiar, and low-friction. But it also exposes the hardest part of AI interaction: most people do not know how to describe what they want.
A short prompt rarely contains enough nuance. Even experienced users often need several rounds to get a strong result. That creates user frustration and system cost, since vague prompts force the model to guess.
A better pattern is to keep the input box at the center, but surround it with scaffolding: suggestions, templates, attachments, modes, prompt enhancers, and example galleries. These supports shift the work from prompt engineering to selection and refinement.
The input still captures intent, but the interface helps carry the burden. That makes the system more forgiving and increases the chance that the first output is useful.
Follow-up
Follow-ups help users continue after the first response. They clarify intent, extend the interaction, and turn a one-off answer into a workflow.
Good follow-ups do not simply ask, “What next?” They anticipate useful next steps based on the user’s goal and the output that was just created.
Common forms include clarifying questions, deeper probes, comparisons, action nudges, and export options. The point is to help users refine what they want without making them start over.
AI interactions are rarely one-shot. Users often discover what they want only after seeing the first result. Follow-ups make that discovery feel natural.
Example gallery
One of the hardest parts of opening a new AI tool is knowing what to do first. Galleries solve this by showing what is possible.
A good gallery is not just a showcase. It is onboarding, inspiration, and instruction at the same time. It helps users understand what the product can make, what inputs work, and what quality to expect.
Galleries can be curated by the platform, generated by the community, or dynamically personalized. The strongest ones balance aspiration with accessibility: they show the product at its best while making users feel like they can create something too.
Governors
As AI systems become more capable, designers need better ways to show what the system is doing, what it is about to do, and where human oversight is needed.
This matters less in simple chat. It matters much more when AI starts calling tools, running workflows, making decisions, or producing large amounts of information.
Governors preserve user agency. They help users understand the AI’s process, intervene when needed, and decide when to trust the output.
Action plan
Action plans are checkpoints. Before the AI begins a complex or compute-heavy task, it explains what it intends to do. The user can confirm, adjust, or redirect before the system goes too far.
Some plans are advisory: the AI shows its approach, but continues without approval. Others are contractual: the user must approve before execution begins. This is especially useful in coding tools, agents, automation, and presentation generators, where a wrong assumption can be expensive.
The higher the risk or cost, the more important the action plan becomes. If AI is about to build an app, generate a deck, run an automation, or change something across tools, users should get a chance to confirm the direction.
The trade-off is speed. Confirmation adds friction, but it also prevents the AI from confidently running in the wrong direction.
Stream of thought
A stream of thought shows the visible trace of the AI’s work: the plan it formed, tools it used, code it ran, sources it checked, or decisions it made.
This makes the system more legible. Users can see whether the AI is still working, why it produced a certain answer, and when they may need to step in.
Most products use a similar pattern: a bounded area, often collapsed by default, that shows the AI’s process during or after the task.
The design question is how much to expose. Too little detail feels opaque. Too much becomes noise. The right level depends on the task. A short chat may need very little. A long-running research, coding, or agentic workflow likely needs more.
The more time, money, or trust the user is investing, the more observable the AI’s work should be.
Citations
Citations connect AI-generated answers back to source material: PDFs, transcripts, web pages, internal documents, databases, or search results.
Their job is verification. They help users understand where an answer came from and whether they should trust it.
But citations can also create false authority. A cited answer can still misrepresent a source, rely on weak evidence, or draw the wrong conclusion. That means citations should not be decorative. They should help users inspect the evidence.
Useful patterns include inline highlights, direct quotations, source metadata, deep links, and lightweight reference lists. Good citation design moves users from passive trust to active verification.
Verification
As AI becomes more autonomous, verification becomes more important. When the system can act on behalf of users, a wrong decision can cause real damage: lost work, wasted money, reputational harm, or exposed information.
Not every task needs approval. Searching, summarizing, or drafting usually does not need a hard stop. But verification should be required when the consequences of a mistake are meaningful.
The simplest pattern is a go/no-go moment. The AI shows a plan, sample, or proposed action, and the user approves or stops it.
Verification should be lightweight. It should not ask users to redo the AI’s work. It should give them just enough information to make a confident decision.
Trust builders
Trust is one of the most important layers of AI product design. Users need to believe the system is accurate, ethical, and safe enough to rely on.
This is not only a UI problem. It is a product philosophy problem. Teams have to decide what defaults they are willing to defend, how much control users should have, and when the product should make its values visible.
Consent
Consent matters when AI captures, analyzes, or reuses data. Transcription existed before generative AI, but the stakes feel different now. Models can learn from voices, generate from conversations, and reuse personal or professional content in new ways.
Consent matters across personal data, organizational data, and other people’s data. The last category is especially hard because AI often captures people who never directly agreed to use the tool.
Consent patterns include opt-in disclosure, silent-by-default capture, post-hoc alerts, and explicit permission for training. Each one sends a different message about the product’s relationship with users and the people around them.
Consent is not just legal coverage. It defines the trust boundary of the system.
Data ownership
Web platforms have long justified data collection through personalization. But personalization often becomes extraction. Platforms collect what people click, read, search, write, and engage with. They are not just observing behavior. They are capturing intent.
AI makes this more complicated. The system may now capture what users ask, upload, create, edit, and co-produce with a model. That raises a harder question: what belongs to the user, what belongs to the platform, and what happens to everything in between?
Many AI companies have responded with data retention and training controls. Most are still opt-out by default. Some, like Figma, have moved toward more user-friendly defaults by turning model-training access off.
Other companies go further and promise that user data will not be used for training at all. If a product takes this route, the interface should say so clearly. Otherwise, users may mistake the absence of a setting for the absence of control.
Wrapping up
These are only a few AI design patterns I’ve been thinking about lately. As workflows become more complex, the patterns will become more specific. But the core challenge will stay the same: help users understand what AI can do, give them control over what happens next, and build enough trust for them to keep going.
Useful links
https://www.smashingmagazine.com/2025/07/design-patterns-ai-interfaces/