Conversations is a chat surface, and it looks like every other one you have used. The difference is entirely in what sits behind it: your organisation's live records, rather than a model's recollection of the internet.
Ask what your pipeline looks like and it reads your deals. Ask who is on leave on Thursday and it reads the leave table. The answers are exactly as current as the data, because they are the data.
Starting
The left rail holds four starters, and they are the four questions people actually ask:
- What is my agenda today
- Who is waiting on a reply
- How is the business doing
- What did I miss
Each one seeds a fresh conversation. They are ordinary prompts rather than special features, so you can edit any of them into the question you meant before sending it.
There is nothing privileged about starting from a starter. Typing your own question works identically and costs the same.
What it can actually answer
Real answers come from your own records: your schedule, your pipeline and where deals are at risk, revenue and overdue invoices, who is on leave, your tasks, and what is waiting on your approval.
Where it does not have the data, it says so rather than producing something plausible. That is the important behaviour and the reason it is worth trusting with a business question at all. A confident wrong number is worse than no number, because you will act on it.
Narrowing with @-tokens
Typing @ in the composer opens a short menu of scope tokens. There are six, and they save a great deal of rephrasing.
| Token | Narrows to |
|---|---|
@org |
The whole organisation |
@team |
Your team or department |
@mine |
Only your own records |
@pipeline: |
One named pipeline |
@hub: |
One HQ |
@period: |
A time range, such as 30d, mtd or ytd |
The first three stand alone. The last three take a value after the colon.
They matter most on the questions where the default scope is ambiguous. "How are we doing" means something different to a regional manager and a salesperson, and @mine @period:mtd removes the argument before it starts.
The four modes
Below the composer sits a mode selector. It occupies the position where a model picker would be in most tools, and it deliberately is not one: which model runs is an administrator's decision, not a per-question one. What you choose here is what you want done.
Ask
The default. A question against your data.
Research
Forces a web search and requires cited sources. This has enough workflow attached that it has its own piece.
Organize
Swaps the composer for the capture bar, so you can pin a record, a link or a note without leaving the conversation.
The case is narrower than it sounds and genuinely useful: you are mid-conversation, something worth keeping surfaces, and you want it kept without losing your place.
Automate
Takes you to AI Agents, where recurring work is configured. It is a doorway rather than a mode, and it exists because "can you do this every Monday" is a question people naturally ask in a chat window, and the honest answer is that it belongs on a different screen.
Asking what it can do
Typing a question about the assistant's own capabilities returns a written, curated answer rather than a generated one. Five things, in plain language: answer, plan, organise, research, act.
That answer is fixed on purpose. A model describing its own capabilities is the single least reliable thing it does, and an inaccurate list is worse than none, because people will believe it.
Memory, and where it differs
In the Conversations tab, chats are saved automatically. You will find them again.
In the command palette, they are not. A quiet offer to remember appears after the first answer, and nothing is stored unless you take it.
That split is deliberate. The palette is for a quick question you do not want cluttering a history; the tab is a place you came to on purpose.
A saved conversation is titled from your first line, so ask something specific and it will be findable later. "Quick question" makes a poor title and there will be nine of them by Thursday.
The history
Saved conversations group into Pinned, Today, This week and Older.
Hovering a row reveals three actions: pin it to the top, archive it out of the list, or delete it. Delete asks once before it goes.
Pin the two or three you are actively working through. Archive the rest rather than deleting them: search still reaches archived conversations, and the answer you needed six weeks ago has a habit of being the one you need again.
What it costs
Every question spends credits, because every question is a language model reading and writing. Roughly one credit per question is the working figure, and short factual questions cost less than long analytical ones.
The starters cost the same as typing the question yourself. There is no cheaper path and no premium one.
What does not cost credits: reading old conversations, searching your history, pinning, archiving, and everything else on the surface. Only the asking.
Getting better answers
Be specific about the window. "How are we doing" and "how did we do this quarter against last" produce very different answers, and only one of them is actionable.
Name the thing. Companies, people and deals can be referenced directly, and doing so removes all the guesswork about which one you meant.
Ask follow-ups rather than restating. The conversation carries its context. Narrowing across three short questions is faster and more accurate than one long one, and each of the three is cheaper.
Treat a surprising number as a question, not a fact. Ask where it came from. It will tell you, and a good proportion of the time the answer is that you asked about something slightly different from what you meant.
The honest limits
It reads, it does not verify. If a record is wrong, the answer built on it is wrong, confidently.
It does not know what is not written down. The reason a deal actually stalled is usually in somebody's head, not in the notes.
It is not a reporting tool. For anything you want to look at repeatedly, ask for the chart and save it. A number in a chat message is a number you will re-ask for next week.
Did this answer your question?
No, ask a person