Your people already work inside an AI assistant. The only open question is whether your documents reach it under your rules or under nobody's.
A private server, in a cloud account you own, that lets the AI assistant your people already use search, read and work with your own documents and operational data. It runs no language model of its own. It retrieves, scopes, computes, scaffolds and records, and it reports what is missing rather than filling the gap.
We build two things, and they answer the same question in opposite directions: where does the model sit relative to your data? Everything else follows from that answer, so it belongs at the top of the page rather than buried in a comparison grid. If the first cell describes your constraint, the rest of this section is not for you, and ninety seconds is the right amount of time to find that out.
The whole environment runs in a cloud account you own: chat over frontier models, an optional knowledge base, a content policy made mandatory by cloud identity policy, and business modules on their own isolated origins. No third-party assistant vendor is in the path at all. If your security owner needs content never to reach one, this is the product.
Your staff keep the assistant they use today. The gateway brings that assistant to your files under your rules: your sign-in, your access policy, your confirmed facts, your calculations. In exchange, what a tool returns travels to your assistant's vendor. That is the trade, and it is the whole trade.
The deciding question is not budget or size. It is whether a third party may see document content at all. If the answer is no, you need the platform, and leading with the gateway in that room is the wrong move. We wrote the decision out in full, including the cases where we would talk you out of this product: platform or gateway, an honest way to choose.
The alternative this product actually competes with is not another vendor. It is the thing your people are doing right now: pasting confidential files into an assistant one at a time, with no access rules, no citations and no record. Measured against that, a governed connector is strictly better.
In the vendors' sense of the word, this is a connector: your assistant discovers it, signs in to it and calls its tools. The difference is everything sitting behind it. It is not a knowledge graph, not a chatbot, and not a module of the platform.
Approved content copied one way out of your document libraries into storage you own, parsed, embedded and indexed there.
Which tools, records and classes of data a person reaches is decided server-side on every call, not in the assistant.
Confirmed facts, kept apart from whatever the documents merely say, each attributed to the person who confirmed it.
Canonical tables, answers in SQL over every row, workflow gates on entry criteria, and refusal rather than estimation.
The mirror is read-only, granted one library at a time, and revocable by you in one action. How it works takes each of these apart.
The tools are grouped in families, and they change as a client's process changes. These three rules do not.
This suits an established, document-heavy organization whose people already use a commercial AI assistant. The buyer is usually the owner-operator or the operations lead who wants more throughput from the same team. The prerequisites are real, and a project that starts without them stalls in the first fortnight. They are checked in the assessment rather than assumed.
| Prerequisite | What it actually means | What happens without it |
|---|---|---|
| An owner for the source content | One named person who can say what belongs in a library and what does not. | Nobody can approve a library, so nothing gets indexed. |
| A folder taxonomy, or the will to make one | The mirror preserves your tree exactly, so a file's address in the mirror is its path in the library. | Classification leans on folder conventions, so scoping becomes unreliable. |
| Named roles | Who reaches which tools, records and classes of data, written down before anyone connects. | There is no default role by design, so an unrecognized person gets nothing. |
| An approved assistant | One your organization has sanctioned, on a plan that supports authenticated custom connectors. | There is nothing to connect the gateway to. |
| Business rules explicit enough to test | A calculation or a gate stated precisely enough that code can refuse when it is not met. | Deterministic tools cannot be written, and you are back to asking a model. |
Judgment, not measurement: these are the conditions the one real build needed, generalised.
An organization that has banned external AI assistants, or whose security owner needs data never to reach a third party, is not a buyer for this product. It needs the platform instead, and we will say so in the first meeting. See also where AI does not belong, the same discipline applied to the work rather than the architecture.
This product cannot win an argument about the data boundary, so it does not try. What it wins on is action, your own rules, and numbers that reconcile.
Everything at rest stays in your account: the mirror of the approved libraries, the index, the confirmed facts, the canonical tables and anything generated. What a tool returns travels to your assistant's vendor, because that is where the assistant reads it, and that traffic is governed by your existing agreement with that vendor.
The full breakdown, line by line, is on access and the data boundary. We make no residency claim for this product at all.
No. Your document libraries stay the front door for filing, and the assistant is the one your people already use. A person adds a connector once, signs in with their corporate account, then asks questions in plain language. Nobody types a tool name. If you later change assistant, the protocol is an open standard, so that is a reconnection rather than a migration.
The server invents nothing. Outward-facing artifacts accept confirmed facts only, computed figures come from SQL over complete data with the query attached, and anything missing appears as a gap. What your assistant then writes in prose is outside our control, which is why every passage carries a citation. We offer no faithfulness or groundedness guarantee, because the server never sees the answer.
Tens of dollars a month on your own cloud bill in the one build so far, and nothing per seat, because the answering model is not in that account. Cost tracks document churn rather than how much your people use it. The engagement is scoped in the strategic assessment; there is no published price for this product.
Six pages, written to be read by a sceptic. Take them in order, or start with the one that answers your own question first.
Where the gateway trades a boundary for reach, the platform does the opposite.
Neither product is allowed to borrow the other's claim. This is where the word is defined.
A bounded piece of work that decides which of the two products fits, what it would govern, and what it would cost to run. You keep the deliverable whether or not you proceed.