You make the demo work.
You show the AI doing something impressive. The output is good. The room is impressed. Investors are interested. The product feels like the future.
Then users get access.
They try it a few times, get mixed results, and quietly stop using it. Not with a complaint. Not with a support ticket. Just with absence.
This is the most common failure mode in AI products right now. And it's almost never about the model.
The trust problem
When a new user opens your AI product, they're asking a question that most AI interfaces don't answer: *can I trust this?*
Not in the abstract — not "is AI generally trustworthy" — but concretely: *can I trust this output, in this situation, enough to act on it?*
Most AI products don't help users answer that question. They present outputs confidently and expect users to figure out the rest.
Users don't figure out the rest. They hedge. They second-guess. They copy-paste the output with a caveat. Eventually, they stop using the tool for anything that matters.
This is a design problem. And there are three distinct versions of it.
Problem one: the blank canvas
Open a powerful, general-purpose AI tool for the first time. What do you type?
Most users don't know. Not because the product is bad — because the product can do too many things and hasn't told them where to start.
The blank canvas problem is the onboarding failure of AI products. A blinking cursor and a text field is not enough orientation for a new user. They need to understand what the product is specifically good at, what a useful starting point looks like, and how sophisticated their input needs to be to get a good output.
The fix is not a feature tour. It's opinionated defaults: show users what a good prompt looks like, what the output looks like when it's working, and give them a starting point they can actually modify. Guide the first session. Let them explore from there.
Problem two: the black box
Your AI returns an output. The user reads it. Now what?
If the user can't assess whether the output is correct — can't see how it got there, can't verify it against anything, can't understand why it made the choices it made — they have a decision: trust it blindly or ignore it.
Neither option builds a habit.
The black box problem isn't solved by making the AI more explainable in a technical sense. It's solved by designing the output interface to give users the signals they need to make a judgement call.
What did the AI use to produce this? What's it confident about and where is it less sure? What would a different approach to the same question have returned? How should the user verify this if they need to?
Most AI interfaces give users none of this. The output is presented as a finished object, with no seams.
A document produced by a human has revision history, authorship, sources. A decision made by a human comes with reasoning, context, the person's track record.
An AI output often comes with nothing. It looks authoritative. It might be wrong.
Problem three: false confidence
This is a specific version of the black box problem, and it's worth naming separately.
Most AI interfaces present every output with identical visual confidence. The font is the same. The formatting is the same. The clean, professional layout signals: this is correct.
But AI outputs are not equally reliable. A model that's excellent at summarising documents might be unreliable at arithmetic. A model that's brilliant at generating code might hallucinate API references. A model that's good at first drafts might be bad at citing sources.
The interface rarely reflects this. A confident wrong answer looks identical to a confident right answer.
The fix: design the interface to represent uncertainty where it exists. Show the user that this type of output should be verified. Create visual distinction between high-confidence and lower-confidence outputs. Give users the language to think about what the AI is and isn't good at.
This is not about making the product look worse. It's about making it trustworthy. Users who understand when to trust a tool use it more, not less.
Why this work doesn't get done
AI teams are usually focused on the model. The UX is treated as a wrapper — something that displays the outputs in a clean interface.
This framing misses the point. The UX isn't displaying the outputs — it's determining whether users understand them, trust them, and act on them. It's the interface between a powerful technology and the people who are supposed to use it.
Getting that interface wrong is one of the main reasons technically impressive AI products get quietly abandoned.
The model is not the bottleneck. The trust is.
Daniel Tkačenko
Founder NotTooMuch