Design for the moment the AI is unsure.
The design of an AI interface is not complete when it shows a good answer. Users need to understand what the system knows, what it did, what it has not checked and what they can safely do next.
Can the user tell what is safe to do next?
Design evidence, correction, authority and finality together. The interface should describe guarantees the underlying system can actually make.
Design the states that fluent text can hide.
Evidence is incomplete
No change proposed
The submission does not establish the requested effective date.
- What the interface reveals
- Show the relevant source passage and the missing field.
- The useful next step
- Request clarification or enter a reviewed value.
Sources disagree
Change held for review
The attachment and account record contain different terms.
- What the interface reveals
- Compare the two values and their source versions.
- The useful next step
- Resolve the difference before approval.
Ready for authorized review
Not saved until acknowledged
The proposed change is visible beside the original.
- What the interface reveals
- Preserve the reviewer, source and validation context.
- The useful next step
- An authorized person may approve the specific change.
The right interface may have no conversation at all.
A structured comparison or review table can expose evidence and action more clearly than a chat transcript. A generic confidence badge is not a substitute for task-specific information.
Watch a person recover from an incorrect suggestion.
- Recognition of incorrect evidence
- Correction, escalation and undo behavior
- Keyboard and assistive-technology access to dynamic states
The complete AI UX and product design argument
For CPO, Head of Design, CTO
The situation behind the request
An AI drafting feature proposes an update to a customer record. The user must decide whether the result is based on the correct source, what changed from the original and whether to approve the write.
Where the obvious solution falls short
A chat-first interface can hide the structure of the task. Generic confidence badges often do not help. Users may accept fluent text as certainty even when the application lacks evidence.
How we would approach the work
We model uncertainty, correction, undo, provenance, permissions and state transitions as part of the core product design. UI patterns connect to system guarantees so copy and controls do not imply a capability the backend lacks.
What the engagement may include
- Task flow and interface states for AI and non-AI paths.
- Cited evidence, compare/review, edit and escalation patterns.
- Design system components for loading, refusal, uncertainty, audit and undo.
- Usability tests paired with model behavior and error conditions.
The choices worth making explicitly
Not every AI interaction should be a conversation. A structured review table, command panel or side-by-side comparison may be more useful than another chatbot.
What we would test before calling it ready
Test whether users can recognize incorrect evidence, inspect source context, recover from bad output and understand when an action becomes final. Accessibility must work with dynamic content.
A sensible first step
Start by designing the error and review states next to the best-case journey. That is where a product earns trust.