The shape of it
Not a paragraph of prose with links underneath. Every claim carries how far the evidence goes, every number carries where it came from and when it was read, and the footer says how many of the retrieved documents you were actually allowed to see.
Four things it has to do
Drive, Slack, HubSpot first — the three that answer a whole department's questions end to end. Everything else through MCP, wrapped and permission-mapped rather than rewritten.
Every figure carries its source and the time it was read. Every claim says whether it was measured, contested, or modelled — and those are different words on purpose.
Permissions come from your own systems and are checked before retrieval, not filtered off the screen afterwards. The footer says how many documents you were allowed to see.
Frontier models today. Models we train and host inside your network later — that part is honest roadmap, not a shipping feature.
One assistant for the whole company is the thing that fails
A single general assistant pointed at every system has no idea what a good answer looks like in finance, so it produces a confident, plausible, unusable one. A department wants a master of that department: something that knows the vocabulary, knows which system wins when two disagree, and knows what it is not allowed to guess. Each one has a named human owner and a set of questions they sign off.
Gross margin after freight. Which ledger line is authoritative when the CRM disagrees.
Refuses toProjecting an unclosed period. Reporting a figure that does not reconcile to the GL.
Stage definitions, the discount floor, and when a deal is genuinely committed.
Refuses toCalling a deal won before the order form is signed. Counting a signal as a fact.
The estate, who owns which system, and what an access request actually needs.
Refuses toGuessing at an access decision. Answering about a system it cannot read.
Attribution windows, channel definitions, and which number the board sees.
Refuses toAttributing revenue across a window the data cannot support.
Policy, process and where the current version of a document lives.
Refuses toInferring headcount from Slack. Answering anything about an individual's record.
The product's real failure modes, and which answer is current rather than archived.
Refuses toQuoting a deprecated runbook without saying it is deprecated.
Your boundary is the product, not a policy
Every other vendor in this category needs a copy of your corpus in their cloud, because that is what their architecture requires. The reason we can offer something else is that the reasoning can run where the data already lives.
The same engine that runs the product people use today, pointed at your systems. Your documents are read to answer your question and are never used to train anything.
Smaller models we train on your language and host on your hardware, so a regulated buyer can run the whole thing with no egress. No training has started. We are saying this is where we intend to go, not that you can buy it — and no tuned model would ship without beating the model it learned from on your own questions.
Four done properly beat forty done thinly
Connector count is this category’s vanity metric. A native connector is not a week of work and then free — it is a permission model to mirror, a rate limit to respect, and a schema that changes without notice. So we are building the smallest set that answers one department end to end, and taking the rest through MCP, wrapped and permission-mapped rather than rewritten.
Three permission models that are well documented and genuinely different from each other. Enough to answer a sales or marketing question end to end.
Where the finance questions live. Importing existing metric definitions matters more than any two document connectors.
Jira, Confluence, SharePoint, Monday and the long tail — most already have a maintained MCP server. We wrap and certify rather than write.
The answer is a question away from being the tool
Every enterprise search product ends at the answer. Australis is a build engine first — it already turns a sentence into working, verified software. So the company whose systems it understands is the company it can build the internal tool for: the reconciliation screen, the weekly brief, the thing somebody currently maintains in a spreadsheet. That is not a services engagement. It is the product we already ship, pointed at what it just read.
Trial it on a question you already have
Pick the question your team asks every week and keeps answering by hand — the month-end variance, the pipeline that never reconciles — and we will build it around that one, in one department, in weeks rather than quarters. Pricing is set with our early partners, on numbers we have measured rather than guessed.
