It learns your company by doing its job.
Not a model trained on your data, a governed RAG (a knowledge base your agents retrieve from) that fills itself out of the work your team is already doing, and answers only to the people allowed to ask.
One library, many shelves. The shelf is the access boundary.
Everything your company knows about itself goes into one governed library, divided into shelves. Which shelf a document lands on decides who its answers can ever reach.
A knowledge base nobody maintains is a knowledge base nobody trusts.
So none of these are a manual re-index. They are the work your team is doing anyway, captured where it happens.
A folder you already have
Put the files in a SharePoint folder and point Studio at it. The folder’s own permissions decide who the answers reach, drop a document in the team folder and the team can ask about it; drop it in the restricted one and only those people ever see it in an answer. Turn on “Keeping it current” and the base re-checks the folder for you.
Livere-checking is a per-base toggle, off until you turn it onYour meetings, the day they happen
Governed meetings are transcribed, masked on your soil, summarised with their decisions and owners, and kept as context. Nobody maintains it, and weeks later the answer still names the meeting it came from.
Live“Remember this”
Telling the assistant something will write it down, as a governed, attributed, versioned entry. Every fact will carry who taught it and when, so an answer can say “per the CEO in June” rather than presenting a guess as canon, and a correction supersedes the old fact without erasing the record.
PlannedYour documentation, on change
The documents that describe how your company works will re-ingest when they change, so the library tracks them instead of drifting from them.
PlannedDrop a folder. Point at it. Done.
Standing up a governed RAG is not an integration project. It is a folder you already have, one screen in Studio, and the agents you choose to attach it to.
A SharePoint folder you already have, policies, contracts, standards, product docs. Nobody converts anything, and nobody uploads a copy somewhere else.
Point at the folder, say who can read it, and pick how its documents are cut. One screen, no integration project, no data pipeline to own.
The help desk gets the product base, compliance gets the standards, everyone gets the handbook. Each agent then answers from what it is attached to, and from nothing else.
A real first ingest reads out like this: 34 files read, 1,208 chunks indexed, 2 skips named, the skips named, not swallowed. sample figures
Already built a RAG? Keep it.
Plenty of teams have already stood up an index, a graph in Neo4j or Graphiti, a vector store, a search cluster someone tuned for a year. Replacing it should not be the price of governance.
Connect it as a source
Coming soon: register the index you already run as a governed source through the same connector lane as every other system, configure, probe, preview what it would reach, approve.
PlannedIt answers under the same rules
What comes back will ride the spine everything else rides: masked before any model, cited so an answer can be checked, reachable only by the people entitled to it, and on the ledger.
PlannedTo be precise about what this is: your index becomes a source ZeroH can read under governance. It is not us moving our own retrieval onto a graph stack, for a standards library we measured that as the slower, pricier and less auditable choice, and we published the comparison.
How a document gets cut decides what a citation can say.
You choose this per knowledge base, when you create it, because it governs how every document in that base is cut, and the citation inherits it. A handbook and a standards library want different answers here.
Handbooks, policies, decks, product docs
Cut along the document’s own headings and paragraphs, so an answer cites the section a reader would have scrolled to.
Regulation, standards, contracts
Cut along the clause grammar, so the clause number and page survive into the citation. This is the difference between “it’s in this PDF somewhere” and an answer an auditor can check.
For anything a regulator reads, that second column is the whole game: an answer that lands on the clause and the page is one your compliance team can check, and “it is in this PDF somewhere” is not.
A RAG that improves as your team uses it.
Retrieval is not one shot at a vector index. Keyword and meaning are searched together, the candidates are re-scored against the whole question before any of them reach the model, and the workforce keeps tuning what it grounds on as the work moves, while every change to how it answers stays something a person ratified.
Grounded, and cited
Every claim carries the document it came from. Beyond what the base holds, the honest answer is that it does not know, not a plausible guess.
LiveReranked on every query
Keyword and vector retrieval run together, and the candidates are re-scored against the whole question before the best of them reach the model.
LiveMasked before any model
The base holds the real content on your soil; what crosses to the model is tokenized. The reader sees the real names, rehydrated on your servers.
LiveRefused when it should be
Retrieval only reaches the bases the person asking is entitled to read. No entitled source, no answer, never a summary that leaks what it summarises.
LiveYour documents are half of it. The other half is what is happening now.
A project is not a folder, it is a team, its meetings, its decisions and the promises people made to each other. In ZeroH that project is the Teams team you already have, and the agents work inside it.
What’s waiting on who?
Four open items on this project, each from the meeting that agreed it:
A week where nothing moved is skipped with a named reason, never padded.
Draft the newsletter about this project.
Drafted from this project’s own meetings and files:
It writes from what the team actually said, and cites the meeting each claim came from.
And when someone needs the story of that project told to people outside it, the same context becomes the artefact, in the pane where the work happens.
The same projects you have in Teams are in the pane’s picker. Choosing one gives the assistant that project’s context, its meetings, its decisions, its files, and nothing from anyone else’s.
LiveSay what it is for. “A status deck for the stakeholders” resolves against the project you picked, not a blank template.
Liveresolving a deck from the picked project ships todayYour wordmark, palette and type come from the brand kit in Studio, so the deck lands in your own branding rather than a vendor’s.
In buildmatching a template you dropped in the channel is the piece in buildThe thread comes with you too: start in Teams, open the pane in the next app, and the same masked conversation resumes, same context, same governance, different window. Live
The folder’s permission is the boundary.
Whoever can open the folder is whoever the answers reach. There is no second permissions model to keep in sync, and no way for a document to answer someone who could not have opened it.
Entitled asker
Gets the answer, with the document and section it came from. The citation is the fact of record, not a flourish.
LiveEveryone else
Gets an honest refusal, no sources they can read, so no answer. Not a hedge, not a summary that leaks the thing it is summarising.
LiveSome shelves stay dark until you say otherwise.
Salaries, invoices, board material. The assistant can know the shelf exists and still not be able to read the numbers on it, so it tells you, instead of guessing around them.
Compute on the real numbers, on the record
The assistant pulls the terms it needs, finds the amounts masked, and asks before it guesses. One granted lift later it reasons on real figures. The reveal itself is evidence.
LiveThe AI asks to see, and you can cut it off anywhere
When the model needs a real figure, it will ask for the shortest useful window, never longer than your DPO allows, and the reveal will be visible and revocable from any surface.
Plannedsession unmask, the DPO ceiling and the audited reveal are live; the just-in-time ask and the revoke-from-anywhere nudge are the new piecesA library is only worth what your team gets out of it.
The cross-source brief
Ask what a colleague worked on this week and get one weighted note across mail, meetings, commits, files and chat, drawn only from what you are already allowed to see.
Read-only, and only yours. The fan-out runs under your own delegated identity, so it can only ever show what you could already open yourself, and every read is a ledger event, visible to your DPO.
Planneddelegated mail, meeting, commit and file reads are live; the weighted cross-source brief and the person-resolve pick card are the new workThe week, for your content team
Ask marketing’s agent what was interesting here last week and it answers from what actually happened, the meetings, the decisions, the work that landed, instead of from a blank page.
Livemeeting-notes recency ranking is a queued improvement