Proactive Research: The Feature That Makes Dots Feel Like Employees

Ask most AI products what they did today and the honest answer is: nothing, until you typed something. That's the gap OpenAI is trying to close with proactive research, one of the quieter features buried in the dots launch on September 29, 2026. A dot doesn't sit idle between your messages. It keeps scanning the tools you've connected it to, on its own initiative, looking for information that might matter to whatever you've asked it to care about.

It's a small phrase in the announcement — "proactive research" — doing a lot of work. It's also the single feature most responsible for the sense, in early coverage, that dots feel less like a chatbot with a memory and more like a coworker who happens to check Slack before you're awake.

What "proactive" actually means here

The mechanics are narrower than the framing suggests, and that narrowness is deliberate. Per OpenAI's own description, relayed by VentureBeat, proactive research is explicitly read-only — a dot can look at connected applications for information relevant to its assigned context, but it cannot send a message, modify a file, or take any action while doing so. It doesn't reply to the customer, comment on the ticket, or edit the doc. It watches, synthesizes, and waits to tell you what it found — or waits for you to ask.

That constraint is the whole design. An agent that acts autonomously in the background is a liability nobody's insurance policy is ready for; an agent that only reads in the background is closer to a smart notification system with taste. OpenAI is betting that most of the value in "always-on" comes from noticing, not from doing — and that the doing part should still require either an explicit ask or a pass through Custom Rules and Auto-Review, the permission layer covered in the previous piece in this series.

It's also why proactive research pairs naturally with the use cases OpenAI actually demoed. A developer's dot watching customer feedback channels doesn't file a fix on its own — it notices a pattern, drafts a diagnosis, and has a pull request ready by the time a human looks. A product team's dot doesn't rewrite the launch deck unprompted — it notices the product changed and flags that the deck is now wrong. The action stays gated; only the noticing runs continuously.

Why "feels like an employee" isn't just marketing

There's a real cognitive shift that happens when software checks in on things you didn't ask it to check. A calendar reminder is proactive in the trivial sense — it fires on a timer you set. A dot doing proactive research is proactive in a harder sense: it's forming a judgment, unprompted, about what's relevant, based on context you gave it once and a set of tools it's watching continuously.

That's closer to what a competent junior employee does than what a tool does. You don't tell a good coworker to re-check the customer feedback dashboard every morning; they just do, because they've internalized what you care about. OpenAI is explicitly chasing that dynamic — the announcement's own line describes dots as agents that "keep working after an employee closes the chat window — monitoring projects, using software, responding to changing information." Proactive research is the mechanism that makes the "monitoring" half of that sentence true rather than aspirational.

The read-only boundary is the load-bearing wall of the whole feature. Take it away and proactive research becomes an agent quietly making decisions you never saw it consider. Keep it, and it's closer to a research assistant with unusually good attention.

It's worth naming the tradeoff plainly, though, because OpenAI's own materials don't dwell on it: a dot that's always watching is also a dot that's always generating — inferences, summaries, half-formed conclusions about what matters — even when nobody asked. Some of that will be useful. Some of it will be noise dressed up as insight, and the Activity View exists partly to let you audit which is which, a topic the next article in this series touches from the data-handling side.

The training question proactive research raises

Because a dot's background scanning generates a lot of ambient content — half-formed notes, things it noticed but didn't act on — OpenAI had to make a specific call about what happens to that material. Per the company's own privacy framing, dots' proactive research and private notes to themselves are explicitly excluded from being used to improve OpenAI's models, separate from the broader workspace-level training settings that govern everything else a dot touches. That's a meaningful line to draw: the exhaust of constant background thinking isn't feeding the next model release. Whether that holds up as scrutiny increases is the subject of the next piece in this series.

How this compares to the competition

Neither of dots' September rivals ships an equivalent. Grok Bot observes user actions and turns them into repeatable routines — a memory of how you work, not a standing watch over what's changing. Meta's Muse is closer in spirit — it can make unprompted suggestions and keeps working after the app closes — but Meta's public materials don't describe a formally read-only "proactive research" mode with its own training-data carve-out; the emphasis there is on the Sentinel approval layer governing outbound actions, not on a distinct background-research capability.

That gap matters for anyone deciding which of these three products to actually trust with ambient access to their tools. A feature with a named boundary — read-only, training-excluded, visible in an Activity View — is easier to reason about than a general promise that the agent "keeps working." OpenAI built proactive research to be legible on purpose. Whether it stays that way once specialist dots and enterprise deployments multiply the number of things being proactively researched is the open question worth watching past launch day.

For teams already comfortable installing composable capabilities into an agent they run and inspect themselves — the model Claude Code Agent Skills follows — proactive research is a useful contrast case: it's powerful precisely because it's narrow, and narrow, inspectable scopes are the property worth demanding from any agent that gets to watch your tools all day.

Part 21 of 30 in the ChatGPT Dots series. Previously: What a Dot Cannot Do Yet: The Honest Limits List. Next: Data, Training, and Dots: What OpenAI Will and Will Not Use. Browse more builder insights or explore AI skills at aiskill.market.