AI8 min read
The AI experience gap: why intelligent products still frustrate users
Model quality has improved remarkably. User satisfaction with AI features often has not. The gap is not intelligence. It is the experience wrapped around it.
By Product Practice, Product design & research
Most AI features shipped in the last two years were built on genuinely capable models. Many of them are also quietly ignored by the people they were built for. Usage spikes at launch, then settles at a level nobody puts in the board deck.
When teams investigate, the instinct is to blame the model: not accurate enough, not fast enough, not smart enough. Sometimes that is true. More often, the model is fine and the experience around it leaves users unable to predict, steer or trust what it does.
The blank box problem
A text box with a blinking cursor is the least helpful interface ever designed. It offers no affordances: nothing tells the user what the system can do, what it is good at, or how to ask. People either ask for something trivial and are underwhelmed, or ask for something ambitious and get a confident wrong answer.
The fix is to show possibility. Suggested starting points drawn from the user's actual context. Examples of good requests. Structured inputs for common jobs, with the free box available for everything else. Users should be able to see the shape of what the feature does before they type.
Users need to steer, not just accept
Many AI features present output as a finished thing to accept or reject. That is a binary choice applied to something that is usually mostly useful. The user wants to keep most of it and change a part, and the interface gives them no way to do that except starting again.
- Let people edit output in place rather than regenerating everything.
- Offer targeted adjustments: shorter, more formal, focus on this section.
- Remember corrections within the session so the same mistake does not repeat.
- Make it clear which parts came from the system and which from the user.
Make checking cheap
Users who carry responsibility for the result cannot act on output they cannot verify. If checking an AI answer takes as long as doing the work, the feature saves nothing. Show sources, highlight the parts that matter, and let people jump straight to the evidence.
Design for the wrong answer first
Every probabilistic system will be wrong sometimes, and users tend to meet those cases early because they test the edges. The experience of a wrong answer, how visible it is, how easy to correct, how gracefully the system admits uncertainty, will shape trust far more than the experience of a right one.
Users do not abandon AI features because they are imperfect. They abandon them because the imperfection is hard to see and slow to fix.
Product Practice
Product design & research
The product practice covers discovery, research, interaction design and design systems. It writes about the structural decisions that sit underneath an interface, and what happens when they are made carelessly.
