Nobody Has Fully Figured This Out Yet

Every designer I know is either working on an AI product right now or just finished one. The category exploded and the design thinking around it has not caught up. Teams are shipping chat windows into products that were never built for conversation, slapping a prompt input onto a dashboard, and calling it an AI feature. Users open it, type something, get a wall of text back, and never touch it again.

That is not an AI problem. That is a design problem.

I have been watching this pattern repeat across products since early 2023. The engineering teams move fast, which is right for this moment. But the interface gets treated like a skin over the model. Just make it look clean. Add a loading spinner. Ship it. And then the activation numbers come back flat and everyone scratches their head.

The Chat UI Default Is Overrated

The first mistake most teams make is copying ChatGPT. Not the thinking behind it. Just the visual structure. A text input at the bottom, a scrolling conversation above it, maybe some suggested prompts floating around. It works for ChatGPT because ChatGPT is a general purpose tool. Users arrive with no context and need an open canvas.

Your product is not ChatGPT. Your users are not arriving with no context. They have a job they are trying to finish. They are inside a workflow. They have data on screen and a goal in their head. A chat interface designed for open exploration is the wrong shape for that moment.

The question you should be asking is not what the interface looks like. The question is what the user is trying to accomplish when they reach for the AI feature, and whether the interface removes friction from that specific moment or adds more of it.

Most interfaces add more of it. The user has to translate their goal into a prompt, which is its own skill that most people have not developed. They get a response that is technically correct but formatted for reading, not for acting. Then they have to figure out what to do with that output inside the product they were already using. Three steps when there should be one.

Prompt Design Is a UX Problem, Not a Copy Problem

Teams bring in a copywriter to write the suggested prompts. Sometimes they write them themselves in a Notion doc and hand them to an engineer. The prompts go live and feel generic because they are generic. They were written without watching a single real user try to use the feature.

Prompt design is user research made visible. The prompts you surface to users are a hypothesis about what they actually want. If your AI feature sits inside a project management product and your suggested prompts are things like 'Summarize this project' or 'What are the next steps', you are guessing. Maybe that is right. Probably it is not specific enough to be useful.

Watch someone use your product for thirty minutes without an AI feature. Write down every moment where they stop, search for something, open a second tab, or retype the same thing. Those moments are your prompt inventory. The AI feature should speak directly to those friction points, not offer a general purpose question box.

At Kraftelite we have worked on AI product interfaces where the biggest unlock was not the model or the output format. It was rewriting the three suggested prompts that appeared on first load. Usage went up meaningfully after that change. Not because the interface looked different. Because it finally said something that matched what users were actually thinking.

Output Design Is Where Most Products Fall Apart

The model gives you an answer. Now what does the interface do with it?

Most interfaces dump the text into the chat window and stop there. The output is unstructured, full length, and sitting inside a conversation container that the user now has to scroll through. If the answer is long, they skim it. If they skim it, they miss things. If they miss things, they trust it less next time.

Output design means deciding what form the response should take based on what the user needs to do next. A summary they need to share should look shareable. A list of action items should look like something they can act on immediately. Data analysis should not come back as a paragraph when a simple breakdown would land faster.

This requires designers to work with the people who write the prompts going into the model, not just the output coming out of it. The system prompt shapes the response format. If you want structured output, you ask for structured output. If you want the model to always return three options instead of one answer, you build that into the request. Designers need to be in that conversation, not handed the output at the end and asked to style it.

Nobody talks about this enough. The interface and the model are not separate problems.

Trust Is Fragile and the Interface Is Responsible for It

Users are not naturally trusting of AI output. They have seen hallucinations. They have copied something that turned out to be wrong. They have a baseline level of skepticism that you are working against before they even type their first prompt.

The interface either builds that trust or erodes it. A response that arrives with no indication of where it came from, no way to verify the logic, and no path to dig deeper erodes trust. A response that shows its sources, lets the user see what data was used, and gives them a way to push back or refine the answer builds it.

Confidence indicators matter here. Not fake confidence bars that designers add to make the interface feel more polished. Real signals. 'This answer is based on the last 90 days of your data' is a trust signal. 'Here are the three documents this pulled from' is a trust signal. An avatar and a spinner are not.

Designing for AI means designing for a relationship that changes over time. The user's trust in the feature is not fixed at first use. It grows or shrinks based on every interaction. Every good response is a deposit. Every wrong or confusing response is a withdrawal. The interface is the account ledger.

Onboarding an AI Feature Is Different From Onboarding a Regular Feature

Regular feature onboarding is about discovery and education. Here is what this does. Here is where to find it. Here is how to use it.

AI feature onboarding is about calibration. The user needs to learn what the feature is good at, where it falls short, and how to get better results over time. That is a harder thing to teach in a tooltip or a modal.

The best AI onboarding I have seen works by doing, not explaining. It takes the user's real data, runs a useful action in front of them without them having to ask, and shows them an output that is immediately relevant. The user thinks 'oh, that is actually useful' before they have typed a single character. That moment is worth more than any onboarding copy you could write.

At Kraftelite we think about this as earning the first moment of genuine surprise. If you can get a user to feel genuinely surprised by what the AI did on their behalf in the first two minutes, they will come back to try it again. If the first interaction is neutral or slightly confusing, you have lost the window.

New Interaction Patterns Are Being Invented Right Now

Chat is one pattern. It is not the only one. Inline suggestions inside a text editor, like what Notion AI does, is a pattern. Proactive surface, where the AI brings something to your attention without you asking, is a pattern. Side panel assistants that react to what is on screen are a pattern. Ambient AI that runs in the background and surfaces a result when it is ready is a pattern.

Each of these requires completely different design thinking. The mental model the user needs to have, the moment of friction you are removing, the trust signals that matter, the way output should be formatted. All different.

Most product teams pick one pattern at the start of a project and never revisit it. The chat window gets built. The chat window ships. Six months later the team is wondering why engagement is low on a feature that was expensive to build. Sometimes the answer is that the pattern was wrong for the job. Not the model. Not the prompts. The fundamental shape of the interaction.

This is genuinely hard design work. There is no established playbook. Designers who are doing it well are running fast experiments, watching real users closely, and changing their minds often. Designers who are doing it poorly are treating AI UI like any other UI and wondering why the results feel flat.

What Good AI Interface Design Actually Looks Like

It is fast. The interface does not make the user wait in a way that feels like waiting. Streaming responses, skeleton states that are honest about what is loading, progress that communicates something real rather than just spinning.

It is specific. The feature knows what job it is for and stays in that lane. It does not try to be everything. A focused AI feature that does one thing well gets used. A broad AI assistant that does many things adequately gets ignored.

It is recoverable. When the output is wrong or unhelpful, the user can push back, refine, or try a different angle without starting over. The interface treats the first response as an opening, not a final answer.

And it earns trust by being honest about what it does not know. An AI interface that admits uncertainty is more trustworthy than one that answers everything with equal confidence. Designing that humility into the interface is one of the more interesting challenges in this space right now.

We are at the start of figuring out how AI should feel to use. The teams that get there first are the ones treating it as a genuine design problem, not an engineering output with a UI layer on top. At Kraftelite, this is where we spend a lot of our thinking right now, because the products that get this right in the next two years are going to look very different from the ones that shipped fast and designed late.