Designing AI interfaces people can trust
Design AI interfaces around clear permissions, visible progress, useful failure states, and human review. Practical considerations beyond the chat window.
The interface around an AI system shapes how people understand its output and decide whether to act. A useful product makes uncertainty, permissions, and progress understandable at the point where they matter.
Choose an interaction for the task
A conversation is useful for exploring an unclear request. A structured form may be clearer when the system needs a small set of known inputs. A review screen can make comparisons and approvals easier than a long sequence of chat messages.
Start with the decision the user needs to make. Show the relevant inputs and result together, with a clear next action. Avoid making people reconstruct the system state from a conversation history.
Make consequential actions explicit
Separate a draft or recommendation from the action it proposes. Before an agent sends information or changes a record, show what will happen and who or what will be affected. Approval should be understandable without reading implementation details.
Where a result depends on retrieved information, make the sources available for review. Present unresolved uncertainty plainly. A polished interface should not imply that an unverified answer has been checked.
Design the uncomfortable states
Plan for missing data, slow operations, partial completion, and a person changing their mind. Explain what has happened and whether retrying could repeat an action. Keep work available when a recoverable error occurs.
Check keyboard navigation, focus management, labels, and status announcements throughout the journey. Test with people completing the task, including examples where the system gives an incorrect or incomplete answer. The design should help them notice and recover.