What someone's trying to get done may decide whether they chat with an AI, direct it, team up with it, or use a custom-built tool.

This explores whether what a person is trying to get done determines how they end up working with an AI: chatting, giving orders, collaborating, or using a purpose-built interface. The corpus has no single study that watches users pick a mode based on their goal. It does have several frameworks that tie goals and modes together, plus evidence that the default mode, chat, often doesn't match what people are actually trying to do.

The most direct framing comes from a taxonomy of five interaction modes: collaborative, directive, assistive, cooperative and explorative How should we categorize different types of human-AI interaction?. The modes are sorted by two things: who starts the exchange, and whether the human and the model work on the same thing together or on separate pieces. Goals are part of the definition here. If you're exploring, you lean toward open-ended back-and-forth. If you know exactly what you want, you give directions. A broader review of 291 papers backs this up by listing user purposes as one of six core design dimensions, alongside how much control users have and how engaged they are How should designers structure human-GenAI interaction systems?. In other words, the field treats purpose as something that shapes the interaction itself, not just as background.

The less obvious finding is that many goals fit chat badly, and people respond to better-fitting modes when they get them. Friedman argues that chatbots make users spell out their intent slowly, one typed message at a time. For many tasks, better tools are ones that capture intent visually, offer "lenses" for shaping the output, or let users edit directly on a canvas Can task-oriented UI patterns reduce the friction of AI chat interfaces?. When an LLM generates a task-specific interface such as a dashboard or small tool instead of a block of text, users prefer it in over 70% of cases, especially for structured, information-heavy work Do generated interfaces outperform text-based chat for most tasks?. A separate analysis of 66 tools found twelve recurring, combinable interaction patterns, such as sliders, selections and constraints that sit alongside the prompt How do interaction signals reshape what generative AI actually receives?. The gap between goal and mode is often bridged by these signals rather than by a better prompt.

There's a twist: the user may not need to choose at all. Poppins watches a person's screen, works out their immediate goal, and steers its output toward it. Users preferred its outputs over a standard LLM's 66–86% of the time Can systems infer what users want from watching their screen?. This flips the question. Instead of the goal shaping the user's choice of mode, the system reads the goal and adapts. The type of goal also affects how well things go. Productivity agents followed good interaction principles far better than companion or high-stakes advisory agents. The weakest areas were exactly where emotional or risky goals need the most care: preventing dependency and showing the user what the agent is trying to do Do productivity agents actually protect users better than emotional ones?.

The gap to know about: these papers describe and classify, but none follows real users to show which goal leads to which mode in practice. What they do suggest is that chat is the default, not a choice users make based on their goals, and that interfaces that infer intent or build themselves around the task may end up deciding the mode for the user.