Do people quit using an AI feature because it's in the wrong spot, or because explaining what they want is just too much work?

This explores whether people drop AI features because the feature is in the wrong place, or because typing out what they want is too much work. First, a limit: the collection has no study that measures abandonment directly, and nothing that tests placement against articulation head to head. What it does have is a strong argument that the two problems are linked. Many 'placement' failures turn out to be articulation failures. The AI was put in a chat box, and the chat box makes users describe their goal in words.

The articulation side is the best documented. The 'gulf of envisioning' work argues that users often can't say what they want up front, because intent takes shape during the interaction (Why can't users articulate what they want from AI?). Models respond to prompts but don't ask questions back, so the user is left to work out their own requirements. The proposed fix is useful: let the model offer options, which turns 'imagine what you want' into 'choose from these', and choosing is much easier. Friedman makes a similar point from the design side. Typed prompts are slow, and patterns such as visual intent maps, style 'lenses' and canvas editing let people steer the AI without writing paragraphs (Can task-oriented UI patterns reduce the friction of AI chat interfaces?).

Placement shows up mostly as a question of which interface a task gets, more than where a button sits. Medeiros argues that defaulting to chat is a step backwards for structured, goal-driven work. It brings back the 'describe it in words' layer that point-and-click interfaces were built to remove. Medeiros offers two tests for choosing between talking and clicking: how fast the task needs to be, and where its value comes from (Is chat the wrong default interface for AI products?). One of Friedman's five levers is literally 'workflow embedding', which means placing the AI inside the tool where the work already happens. In the collection's framing, then, good placement is a way to reduce articulation friction, because the surrounding context does part of the explaining for you.

A less obvious point: some of what looks like the user 'not knowing how to ask' is a design problem. A chat interface invites people to use their everyday conversational skills, but the AI isn't actually communicating the way a person does. Interactions break down in ways that feel like user error but start in the design (Why do users fail with AI interfaces designed like conversations?). On top of that, AI context keeps shifting (prompt, history, retrieved data, hidden state). Users can't build a stable mental model of it the way they can with a normal app, so the same request can work one day and fail the next (How does AI context differ from conventional software context?).

The Cocoa study comes closest to evidence on what keeps people engaged (Can users steer AI agents better by revising plans mid-execution?). Users could assign each step of a plan to themselves or to the agent, and revise the plan while it ran. That gave them more control than a chat baseline without making the tool harder to use. With only 16 lab participants it is suggestive, not proof. It points to a reframing of the question: retention may depend less on placement versus articulation and more on whether the interface lets intent develop gradually. The question users face is whether they need to know what they want before they start.