AI features fail less often because of the model than because of the interface. A good model behind a poor design creates distrust, waiting frustration and results nobody uses. Designing AI features differs from classic software in one fundamental way: the output is not predictable. The same input can produce different results, some of them wrong. Good UX takes that seriously instead of hiding it.
Setting expectations
Users need to understand what the feature can and cannot do. An empty input field saying «Ask me anything» invites questions the product cannot answer, and disappoints. Concrete examples that show what the feature is for work better: «Summarise this ticket», «Draft a reply in the tone of our last emails». The empty state is the most important moment of an AI feature because it shapes the expectations users bring.
Designing the wait
Language models take time, especially for longer answers. A spinner circling for ten seconds feels longer than an answer that starts appearing after one second. Showing the answer progressively as it is generated shortens perceived waiting considerably. Where that is not possible, for structured results, intermediate steps help: «Searching documents», «Writing answer». Long tasks belong in the background, with a notification when they are done.
Results as drafts
Most AI results are drafts, not finished products. The interface should express that: text that can be edited directly, suggestions that can be accepted or discarded individually, an option to regenerate. An AI feature whose result can only be taken or deleted as a whole forces users into an all-or-nothing decision and gets used less.
Sources and traceability
When an answer is based on documents or data, these should be visible. A link to the underlying document, record or passage lets users check the answer and reveals when the model wrote without a basis. Traceability is the most effective lever for trust — far more effective than claims like «our AI is very accurate».
Communicating errors honestly
AI features make mistakes, and the interface should expect that. A subtle, permanent note that results should be checked is more honest than a one-time warning nobody reads. When the system is uncertain or finds no relevant information, it should say so. An answer like «I found nothing about this in your documents» is better than a plausible invention, and users appreciate that honesty.
Enabling feedback
A simple rating — helpful or not, with an optional comment — gives the product team valuable signals about where the feature fails. This feedback belongs in the analysis and in the test set used to check new versions. It must be stored in compliance with data protection rules, because it often contains content from the original request.
Labelling AI
Since 2 August 2026 the EU AI Act requires that people can recognise when they are interacting with an AI system unless it is obvious. Labelling obligations apply to synthetically generated content. Labelling need not be a nuisance: a small, consistent symbol for AI-generated content and a clear name for the feature meet the requirement and help users judge results correctly.
Leaving control with the user
Users should decide whether and when AI is used. Automatic suggestions that pop up unasked quickly feel intrusive. Features that access other people's data should make this transparent. And anyone who does not want to use an AI feature should be able to switch it off without the product becoming worse to use.
A short checklist
- Does the empty state show what the feature is for?
- Does the answer start appearing quickly, or does the task run in the background?
- Can results be edited, partially accepted and regenerated?
- Are sources visible, and does the system say when it finds nothing?
- Is the AI interaction labelled?
- Can users give feedback and switch the feature off?
Innopulse designs AI features in its own products according to these principles and helps teams build the interface as carefully as the integration behind it.