Why does an AI that acts before you ask feel helpful in one app but pushy in another?

This looks at why the same idea, an AI that acts before you ask it to, can feel helpful in one product and pushy in another, and which design choices make the difference. In the corpus, intrusiveness rarely comes from proactivity itself. It comes from proactivity without what researchers call civility: knowing when to interrupt, how much to take over, and when to stay out of the way. The Intelligence-Adaptivity-Civility framework makes the point directly. An agent can be smart and adaptive and still be socially blind, cutting in at bad moments or overriding the direction the user has set How can proactive agents avoid feeling intrusive to users?. There's a twist here: most assistants aren't too pushy by default. They're too passive, because training that rewards only the next reply removes initiative. Initiative can be trained back in, from almost none to about 74% of cases with reinforcement learning, so the open design question is how to add it without making the agent annoying Why do AI agents fail to take initiative?.

The clearest evidence comes from a study of programmers working with proactive coding agents. The agents cut the effort programmers had to put in, but they also broke up their workflow. What reduced the disruption was not making the agent less proactive. It was making the agent visible: presence indicators and a shared view of what the agent was working on Do proactive AI coding agents actually disrupt programmer workflow?. Being interrupted by something you didn't know was there feels like an intrusion. Being interrupted by a collaborator you can see working feels more like teamwork. Microsoft's agent design principles say much the same thing in their own words: agents should be 'accessible yet invisible until the right moment,' with uncertainty shown openly and the user kept in control How should AI agents balance helpfulness with staying out of the way?.

There's also a deeper reason timing is hard to get right in AI. In ordinary software, the context is fixed and readable: you can see the screen and you know what state you're in. In AI products, the context keeps shifting. It includes the prompt, the conversation history, retrieved data, and hidden state, none of which users can hold in their heads How does AI context differ from conventional software context?. So when an agent acts unprompted, users often can't tell why it acted now. That gap gets wider because much of an agent's behavior is set by system prompts users can never see Who actually controls what AI systems say?. A proactive move that comes from rules you can't see is much easier to experience as intrusive than one whose reasoning you can follow.

The setting matters too. A review of nine deployed agents found that general productivity tools like ChatGPT and Claude did best on interaction principles. Companion and high-stakes advisory agents did worst, especially on preventing dependency and making the agent's intent visible Do productivity agents actually protect users better than emotional ones?. A proactive nudge in a work tool feels like help. The same nudge from a companion app, whose goals are unclear, can feel like manipulation. The way an agent presents itself socially raises the stakes. Just one strong cue, like a voice, is enough to make people treat an AI as a social actor Do more social cues always make AI feel more present?. Acting on its own is one of the design features that lead people to think an AI is conscious What design features make users perceive AI as conscious?. The more person-like an agent seems, the more its unprompted actions are read as the choices of someone with intentions, and judged that way.

The unexpected lesson: making an agent proactive is less about deciding when it should speak up and more about letting users see why it is speaking up. The corpus has strong material on visibility, timing, and context, but little direct research comparing proactive designs head to head across domains. If you want to go deeper, start with the civility framework and the coding-agent study, which together come closest to answering the question.