The default paradigm for AI products is people finding AI: think of a question, open a chat window, type, wait.
The path looks short, but most users never finish it. They don't know what to ask, what data the system holds, or what kind of deliverable a sentence can buy. A chat box is a blank page, and for most people a blank page is pressure, not an invitation.
Invert it
The system knows what you are looking at and what you need. So finish the analysis before you open the page.
An example. A user enters a requirement, dates and quantity on a procurement platform and gets a list of supplier options. From the list alone it is still hard to tell which option suits this need, whether the supplier fits, what constraints the delivery terms carry. So we gave every search its own order diagnosis: identify the suppliers behind the top recommendations, fetch supplier, category and delivery data in parallel, assemble a structured judgment, and place it on the results page.
The user asked nothing. The analysis is already there.
Three places to find the scene
A proactive agent can appear in three positions:
- Inside a task. Like the order diagnosis above: the user is doing something, and the AI prepares the judgment that task needs ahead of time. Trigger and reader are the same person; the AI just removes the research step.
- Inside content. Discover industry changes daily, generate insight cards and news, and make the entities in the text clickable and askable. Reading becomes the entrance to AI.
- Inside outreach. Push briefs, event previews and insights by user profile. The most effective position and the most dangerous: a wrong push costs more than no push.
Output is not the measure
The trap proactive agents fall into is proving themselves by volume: how many cards generated today, how many briefs, how many articles.
Those numbers mean nothing. Ten thousand cards overnight that nobody opens equal zero. So we defined a task in the north star as a unit of work consumed by a person: only what gets opened, followed up on, or adopted counts. That one rule turned the evaluation of proactive products from output to consumption, and forces every proactive action to answer: which step did you save the user?
The boundary
Proactive does not mean replacing. The AI finds the scene and prepares the judgment, but a person still decides. A diagnosis can say which supplier fits better; which one to pick, what to quote, how to negotiate belongs to the professional. The goal of a proactive agent is for people to look up less, wait less and copy less, not think less.