Most work tools open on a dashboard that tells you how much there is. Counts, charts, a burndown. None of it answers the question you actually have at nine in the morning, which is: what needs me, and what should I do first.
Home now opens on a box you can type a question into, and a short list of things that genuinely need you, each with a reason attached. It is the same work graph underneath. The difference is that you ask it something instead of reading it.
Ask about your chat and your tasks
The box at the top takes plain words. "What needs me today?" "Catch me up on chat." "Brief me on my top task." "What changed this week?" It reads across your tasks and your channels, so a question that spans both has one answer instead of two searches.
Answers stream back. While it works it shows what it is doing, one quiet line at a time: checking your tasks, searching the workspace. You can see the steps it took rather than waiting at a spinner and hoping. Task references in the answer are links, so the reply is a place to start working, not a dead end you have to re-search.
It does nothing until you say so
The assistant can create a task, comment on one, post a chat message or hand a task to an agent. It never does any of those on its own. When it wants to write something it stops and shows you a card.
Nothing is written until you press Confirm. Cancel throws it away. It proposes at most three of these in a single reply, so you are never handed a wall of decisions.
It also cannot do anything you could not do yourself. It acts with your permissions, not with elevated ones. Guests can ask questions and read the answers, and cannot make changes at all. The page says so plainly rather than letting a guest discover it at the moment they press a button.
Undo, and its honest limits
After you confirm a created task or a comment, you get fifteen minutes to undo it. One press and it is gone again.
Two things that undo does not do, because pretending otherwise would be worse than saying it:
- A posted chat message and a task handed to an agent cannot be undone. A message other people have already read, and a run an agent has already started, are not ours to retract. Those two get no undo button at all rather than a button that fails.
- Undo is refused if the thing changed in the meantime.If someone edited the task or replied to the comment after it was created, undoing would quietly destroy their work, so it stops and tells you what changed.
Needs you, with the reason attached
Under the box is a short list of what is actually waiting on you. The point is the second half of each row: not just what it is, but why it is in front of you.
It shows five rows and offers the rest behind a count, so the list stays a list rather than becoming another inbox to triage. Underneath it, three to five next steps are built from those same rows. Picking one fills the ask box with a question about that specific item, which means the common case is two clicks and no typing.
It learns which kinds of things you act on
People use the same page differently. One person opens agent questions first and ignores the catch-up cards. Another never touches a review row but always opens the chat summary. Home notices which kinds you open and moves those up for you.
What it stores is deliberately thin. It records the kind of thing, for example "overdue" or "brief task", and a count of how often that kind was shown and opened. It does not store task names, message text, or anything you wrote. The kinds are a fixed list, so anything that is not on it is dropped rather than recorded.
The reordering is also deliberately limited. Agent questions and overdue work keep their place at the top whatever your habits are, because the cost of burying those is high and the benefit of personalising them is nil. Learning only reorders within a group, and it never hides anything. With little history, the order is the normal one.
The counts do not pile up forever either. The daily record of which kinds were shown and opened is kept for up to about ninety days, and a job deletes anything older than that.
What you need to use it
The two halves have different requirements, and it is worth being exact about which is which.
Needs you and Next steps work on every plan
Including Free. They read your own work graph, so there is nothing to switch on and nothing to pay for.The ask box needs Pro, and a model
It needs the Pro plan and an AI model configured for the instance by whoever runs it. On a Free or Starter workspace it says the assistant is part of Pro and links you to the plans. If the plan is fine but no model is set up, it says that instead.There is a fair-use limit on questions
Thirty questions an hour, per person per workspace. If you hit it, it says so and asks you to try again shortly.
Self-hosted instances read their licence rather than a per-workspace plan, and an active trial counts as Pro while it lasts.
Why a box and not another dashboard
A dashboard answers the questions its author thought of, in the order its author chose. That works until your question is slightly different from the one on offer, which is most mornings.
A box plus a reasoned list is a smaller promise. The list says here are the four things with your name on them and why. The box handles everything else, and shows its working so you can tell whether to believe it. When it wants to change something it stops and asks, which is the part that makes the rest usable.