AI Automation7 min read
What is a smart UI? Interfaces for AI that can be wrong
A smart UI, or intelligent user interface, adapts to the user and to what a model predicts. What the term means, what else shares the name, and six design patterns.
A smart UI, short for intelligent user interface, is an interface that adapts to the person in front of it and to what a model predicts they need, instead of showing everyone the same fixed screens. Today the phrase mostly means the interface wrapped around an AI feature: a search box that answers, a form that fills itself in, an assistant that takes actions.
That definition is the easy part. The useful test of a smart UI is not how it looks when the model is right. It is what the screen does when the model is wrong.
This article covers both: what the term means, the other things that share the name, and the design patterns that decide whether people keep using an AI feature after its first mistake.
Other things called Smart UI
The name is crowded. If you were looking for one of these, here is where each one lives.
A research field. Intelligent user interfaces have been studied since the early 1990s, with their own ACM conference, IUI, first held in 1993. The idea is older than the current wave of AI by three decades.
A testing product. SmartUI is the visual regression tool from TestMu AI, formerly LambdaTest. It compares screenshots of a build against a baseline to catch unintended changes.
A component library. Smart UI is a set of web components: grids, charts, schedulers and form controls.
An anti-pattern. In Domain-Driven Design, Eric Evans used "Smart UI" for an application that keeps its business logic in the interface layer. In that vocabulary it is a warning, not a compliment.
A company. Smart-UI is a software development agency headquartered in Merida, Spain, with an engineering centre in Kharkiv, Ukraine. Founded in 2018, Smart-UI has around 20 engineers and has delivered more than 315 projects. That is us. We have carried the name since 2018, and the meaning this article is about arrived later, with the AI features clients now ask us to build.
Why a smart interface is harder than an ordinary one
An ordinary interface is deterministic. The same click gives the same result, and when something fails it fails loudly, with an error.
A model is different in three ways. It is wrong some share of the time. Nobody can say in advance which times. And a wrong answer arrives looking exactly like a right one: same font, same confidence, same position on the screen.
The third point is the design problem. The model cannot tell the user when to doubt it. The interface is the only place where the difference between "this is certain" and "this is a guess" can be shown, and most interfaces show nothing.
Six patterns that decide it
1. Say how sure the system is
In words and ranges, not decimals. "Likely match" and "about three to five days" are honest. "87.3% confidence" claims a precision the model does not have, and most people cannot act on it anyway.
Show uncertainty only where it changes what the user should do. A warning on every answer is the same as a warning on none.
2. Show where the answer came from
A link to the source document, the date of the data, the record the number was read from. This turns "trust it or reject it" into a five-second check, and it is the cheapest pattern on this list.
3. Make the correction one step
Edit in place. Undo. "Not what I meant." The real cost of an AI mistake is the time it takes to fix, and if fixing a wrong answer takes longer than doing the task by hand, the feature is a net loss however often it is right.
4. Suggest freely, act carefully
Anything reversible can happen without asking. Anything that is not, a payment, a sent message, a deleted record, gets a confirmation that shows exactly what is about to happen. A draft is a suggestion. A send is an action. The interface should make it impossible to confuse the two.
5. Design the "I do not know"
Most AI features are designed for the answer and not at all for its absence. A model that is forced to answer always will, so the interface needs a real exit for the cases where it should not: a plain search, an ordinary form, a way to reach a person.
This is the pattern teams skip, and it is the one that prevents the worst failures.
6. Let correction teach, and let the user see it
If someone corrects the system, it should not make the same mistake for them tomorrow. It should also be visible that it remembered, and possible to make it forget. Personalisation nobody can inspect feels like the product behaving strangely.
Two reference works go much further and both are free: Microsoft's Guidelines for Human-AI Interaction and Google's People + AI Guidebook. The six above are the subset that, in our experience, decides most projects.
You do not need a language model to have this problem
We designed a booking platform where listings, prices and availability arrive live from an inventory system that somebody else maintains. There is no language model anywhere in it. The hard screens were exactly the ones described above.
A price before it is known. A visitor enters dates. The exact price needs a call to the external system. The tempting shortcut is to show each listing's "from" price in the meantime and leave it there. But a "from" price shown to someone who has already given dates is a confident answer to a question they did not ask. The right behaviour is to label an estimate as an estimate and replace it the moment the real figure exists.
A category the feed got wrong. A feed can label a parking space as an apartment. In a list that is a small oddity. If the interface builds its headline "from" price for a whole city out of the cheapest listing, one mislabelled row sets the price of the city. Any number promoted to a headline needs a sanity check that a number in a list does not.
Availability that changes while someone is looking. A calendar is a prediction about the next minute. Checking again at the step before payment, and saying plainly that something changed, is better than failing at the payment itself.
None of that is artificial intelligence. All of it is a system presenting an uncertain answer with full confidence, which is the whole problem a smart UI exists to solve. Adding a model makes it more frequent. It does not make it new.
Five questions before an AI feature ships
- What does the screen show when the model is wrong, and how would the user know?
- How many steps does it take to correct it?
- Which actions can it take without asking, and is every one of them reversible?
- What happens when it has no good answer?
- Where do the corrections go, and who reads them?
If the fifth question has no answer, the feature cannot improve. The mistakes are being made and nobody is collecting them. It is also the usual reason an AI product that works for the first ten customers stops working at a hundred, which is the subject of scaling an AI MVP that is still delivered by hand.
Where we come in
The model is rarely what makes an AI feature fail. Models are good, and getting better without your help. What fails is the layer between the model and the person: nothing shows doubt, nothing can be undone, and the first confident mistake is the last time that customer trusts the feature.
That layer is design and engineering work, and it is what we do on the AI automation and UI/UX design sides of the company. If you have an AI feature that demos well and gets abandoned in its second week, show it to us. The cause is usually on this page.