Should AI chat swap typing for buttons and menus, or just add friction to make you think harder mid-conversation?

This explores two moves that look like opposites: replacing typed conversation with task-specific interfaces to remove effort, and adding deliberate friction inside a single chat turn to make users do more work. The corpus suggests they aren't really rivals. Both start from the same complaint, that the plain chat box puts user effort in the wrong place, and each answers that complaint in a different way.

The task-oriented camp says chat makes people spell out their intent slowly in typed prompts. Can task-oriented UI patterns reduce the friction of AI chat interfaces? names five places where buttons, canvases and visual controls can replace that typing: capturing intent, shaping the output, refining it, running agentic tasks, and fitting the AI into existing workflows. Is chat the wrong default interface for AI products? goes further and calls chat-by-default a step backwards. For goal-directed work, it brings back a layer of instructions that direct manipulation (clicking on things) was invented to remove. The evidence partly backs this up. Interfaces generated on the fly beat text replies in over 70% of cases Do generated interfaces outperform text-based chat for most tasks?. But the gains come at a price. Generated widgets add clarity but are rigid and hard to change mid-task Do generated analysis UIs really work better than chat?. And in one study, letting the AI handle a task cut clicks and scrolling without making the task any faster Does chat delegation actually save time on task completion?. Less effort and less time turn out to be separate measures.

The friction camp starts from the other end. Chen and Gero reviewed 73 designs and found three ways to slow a chat turn down productively: asking the user to contribute something first, holding back part of the AI's help, and changing the shape of the AI's answer How can designers slow down AI chatbot interactions productively?. When tested with users, these interventions reliably raised workload, time spent, and the sense that the work was the user's own. Memory improved only when the friction matched what the user was trying to do at that moment Does friction in AI chatbots actually improve what users remember?.

What connects the two camps is Medeiros's "value source" question: where does the value of this task come from? If the value is the output (a dashboard, a filtered list, a booked flight), friction is pure cost and a task interface wins. If the value is the user's own understanding or ownership (learning, writing, deciding), removing effort removes the point, and friction inside the turn is the right tool. So the real design choice is less about how much effort to ask for and more about where that effort should go. A related finding: the shape of a conversation alone (how long the turns are, how it moves back and forth) predicts user satisfaction almost as well as reading the actual text Can conversation shape predict whether it will work?. Both approaches are really ways of reshaping that structure.

A third option sidesteps the choice: change the conversation itself rather than leaving it or slowing it down. An AI that offers useful information before being asked can cut the number of turns by up to 60% Could proactive dialogue make conversations dramatically more efficient?. A system that runs delegated tasks in the background lets the conversation keep going while the work gets done Can frontends handle delegation while staying conversationally engaged?. These suggest the familiar chat format isn't fixed, and some of its friction is accidental, a by-product of strict back-and-forth turn-taking rather than a design decision.