Tasks
Company knowledge
Search and reason over your team's Brain — the assistant looks up what your organisation already knows before answering, and tells you which pages it used.
Your Brain is the team’s knowledge base: policies, people, customers, decisions, uploaded material turned into pages. The assistant searches it before answering questions about your organisation, so answers come from what your team wrote down rather than from what a model assumes about companies in general.
Where it runs: in a task, and in AI staff runs for staff connected to a brain.
What the assistant can do
| It can | What that looks like |
|---|---|
| Search the brain | Ranked pages with snippets — the fastest way to find what exists |
| Answer from the brain | A grounded answer combining several pages, with the pages it cited and the gaps it found |
| Read one page | The full content of a page by name |
The activity line on the reply lists the pages it looked at, so you can open them and check the answer against the source.
The same three moves are available directly in Brain: add knowledge, search it, or ask it to reason across pages.
Gaps are part of the answer
When the assistant answers from the brain it also reports what the brain did not have. That is not a failure message — it is the most valuable output the tool produces. A repeated gap (“nothing about parental leave”) is a to-do for whoever maintains the knowledge base.
Grounding, and why it matters
An AI staff member that depends on retrieval is explicitly told: if the search returns nothing, say so rather than answering from memory. This is not caution for its own sake. Asked about a policy with an empty brain, an ungrounded model will invent an entitlement complete with a clause number and an effective date, and it will sound exactly like a real answer.
The one carve-out is material you supply. If you paste a contract clause and ask what it means, the assistant works on it directly — it does not refuse on the grounds that the text was not retrieved.
Getting good answers
Ask the question, not for the document. “What is our refund window for annual plans?” searches better than “find the refunds page”.
Fill the gaps it reports. The loop is: ask, see the gap, write the page, ask again. The brain gets useful by being used.
Check the citations on anything consequential. They are listed for exactly that reason.