Anyone who answers to auditors, boards or clients needs a sentence SazAI Corpus knows how to say: I don’t know, it isn’t in the material. The agent reads the same vault as you, from inside, and every answer arrives with the address it came from.

Ask the way you speak

One question in plain English, and the answer comes back with the address it came from

You ask the way you would ask someone on your team: which slides mention gross margin, what the contract says about termination, which sheet holds the cost per unit. No syntax, no search field, in plain English. The agent answers from what the engine already read and measured on entry, and brings the address along: the slide, the note, the section, the sheet. The conversation stays with you, in your vault, and nothing it answers is built from another company’s data.

Which slides mention gross margin?
Two: slide 12 (Q3 Results) and slide 34 (Appendix). On 12, the speaker note carries the consolidated quarter figure.sources: slide 12 · slide 12 note · slide 34
And in the Portuguese version, what did the title of 12 become?
"Resultados Q3: Margem Bruta Consolidada." The translation sits side by side with the original in your vault.source: pt-BR translation · slide 12
Two questions, two answers, and each one with the address it came from.

A generic chatbot gives you an answer; an agent with source gives you a claim you can check. The source is an address in your vault, and the click opens slide 12 with its notes and comments around it. That is the difference between an impression and a fact.

The conversation is with the AI

The engine is what touches the file, and that division is where everything starts

The layer that talks understands what you want and chooses the operation; the one that touches the file is a deterministic engine that executes a specification over a known structure. That is why the worst possible error from the AI stays an error of conversation, and shows up in the receipt instead of in the file.

Translation is the one point where the model’s text actually reaches the document, and there the boundary is checked instead of assumed. How that works, and what happens when the count does not match, is in the conversation is with the AI.

Interfaces

On the vault screen, in the AI client you already use, or through your team's API

The same conversation happens in three places, and the vault behind all three is one. The channel through which an AI client talks to it is an open standard: it answers 401 without a credential, which is the correct behavior, and publishes discovery for any client to read on its own.

What this enables is not plumbing: it is your team building dashboards, flows and automation on the same substrate, with no one from us in the room. The three doors, and why changing doors never changes who owns the vault, are in interfaces.

Capabilities

What it does, what it refuses, and the contract that keeps the set single

When something is not in your documents, the agent says it is not there. It does not fill in, does not summarize what it has not read, does not fill the silence with plausible fiction. For anyone who answers to an auditor, it is not in the material is a feature, not a failure.

And the set of operations does not change size with the door: what the agent exposes is a subset of what the API exposes, by contract. What it does today, what it refuses, and the three questions a capability answers before it becomes a served sentence, are in capabilities.

What the agent reads to answer, it also triggers when the question becomes a request. The next one is the first of them: translate, with a receipt.