User guide
Ask AI assistant
A chat box that lives inside the search palette: ask a question, get an answer written from this guide, with the page it came from cited underneath. It is a preview - honest about what it knows and quiet about what it does not.
Opening the assistant
Press ⌘K (Ctrl+K on Windows and Linux) on any page and choose the Ask AI tab. From the Search tab, any question that returned results also offers an Ask AI about … row - choosing it hands your exact words to the assistant. Each conversation continues with follow-up questions; opening the tab starts a fresh one, and recent conversations are listed so you can resume one instead.
Answers from the guide, with receipts
The assistant answers only from the product documentation - this guide - and lists the pages it used as clickable citations under every answer. If the docs do not cover your question, it says exactly that rather than guessing: a wrong answer that looks right is worse than a shrug. That rule is the whole design - citations are not decoration, they are the receipt.
What it can see, stage by stage
The preview rolls out in stages, and all of them obey one rule: the assistant sees exactly what you can see, and nothing you cannot - it reads through the same rules every page enforces, never around them. The first two stages are read-only. Stage 3 adds drafts, and the rule holds there too: it can prepare a write but never perform one - a draft becomes data only when you press confirm on its card.
- Stage 1 - the guide. Answers come from this documentation only, cited by page path. A question about your own work is met with a plain statement of the limit.
- Stage 2 - your work, as you. The assistant gains read-only tools over the
rows your account can already open: search across work and discussions, one item with its
history and recent messages, your open queue, the organization's workflow words, and the
recent activity feed. Answers then carry two kinds of receipts: doc chips name their
page; work chips show the reference itself -
ALP-1- and jump straight to that item. A reference the assistant invents earns no chip: receipts appear only for rows a tool actually fetched for you. - Stage 3 - proposes, you confirm. The assistant can draft new work items and discussion comments. Every draft is a card: you see the exact fields, and the card offers Confirm and Reject - nothing else writes. Confirming executes as you, through the same form the website uses, so you get precisely what you would have typed; the trail records it honestly as agent-drafted and confirmed by you. A draft is a request, not a checkpoint: if the world moved while it waited - the project closed, the parent was deleted, your membership changed - confirming fails loudly and writes nothing. Rejecting is silent: a declined draft simply dies.
If the assistant answers that it cannot see your workspace, your tree is still on stage 1 - stages 2 and 3 turn on tree by tree, stage 3 only ever on top of stage 2, and the preview decides the order.
Pauses, never losses
Thinking costs money that the service absorbs, so each organization tree gets a daily and a monthly allowance of questions - the current numbers sit in the usage limits table. Reaching either pause stops new questions until the period resumes: threads, past answers, and citations stay readable, and a refused question is never even saved, so you can simply ask it again later. On stage 2, one question still spends one slot however many times the assistant consults its tools - the loop is bounded, so a working question can never spend your day twice. On stage 3, deciding a draft is always free: pressing Confirm or Reject contacts no model at all, so it never spends a question slot - the pause governs asking, not your own decisions.
Preview availability and privacy
The assistant ships organization tree by organization tree during the preview; if the tab is missing for you, your tree is not enabled yet, and stage 2 turns on tree by tree on top of that. While you ask, your question and the material used to answer it are sent to the model provider: always the public documentation excerpts, and on stage 2 also the workspace rows the tools fetched - each one something you can already see, clipped to what the answer needs. On stage 3, the fields of a draft are part of that same trip: to propose a work item, the model composes its fields, so what the draft would contain travels to the provider as a tool call like every other answer material. The draft itself is stored on our side as your pending card; it becomes real data only when you confirm, and confirming sends nothing to the provider at all. Nothing is sent that your own pages would not show you. Answers are stored as your conversation threads - private to you and the organization they belong to - so you can resume them.