You decided, in the second demo, to hand over a contract. Drag, drop, a progress bar, and the salesperson said the agent now knows your business. The question nobody asks in a demo is where the file went, and what it sits next to.
The reading room where it landed is called the Library, and in ArkOne’s reference design for an executive agent, the Cabinet, that room has four shelves. Your contract went on one. The other three hold material you did not write, and the design never lets any of the four become another.
What the four shelves hold
When the Chair, the one voice that speaks for the whole room, needs something it has not been given, it asks the Library: a search over stored material returning the passages closest to what was asked. Every agent that reads documents has one. What varies is whether the material is one pile or several, and the reference design chooses four, recorded on the Register.
The first shelf is house knowledge: what the design knows about running an executive meeting. How a pricing decision is weighed, what a board pack contains, which questions Finance asks before Counsel does. It is the vendor’s material, the same for every company using the design.
The second shelf is your company’s documents: the contract you dropped in, your price list, your leave policy, the minutes of your own board. Nothing goes on it that your company did not produce or approve.
The third shelf is recorded failures: the moments the room got something wrong and someone said so. A discount that set a precedent; a supplier written to in the wrong tone. Each is written as a warning, and the shelf is searched for the reason a pilot reads incident reports.
The fourth shelf is recent research: anything fetched from outside during a meeting. A competitor’s public price list, a regulator’s page, a news item. Every passage carries the address it came from and the date, and it stays on this shelf however often it is read.
The shelves matter because of what comes back with a passage. A search result carries its shelf, so the Chair reads “your contract says” as one kind of sentence, “a web page says” as another, and “last time this went wrong” as a third. Put the four in one pile and the labels vanish; the model receives text in one voice and guesses which sentences are your company and which are the internet.
What is on each shelf, and what must never cross
Suppose the renewal that runs through this series, the invented client asking to keep last year’s price, and follow the four shelves through the Chair’s first round, when it asked the Library for the contract and last year’s pricing.
| Shelf | What is on it | Who put it there | What must never cross onto it |
|---|---|---|---|
| House knowledge | How to weigh a discount: margin floor, then precedent, then relationship | The design, once, for every company | Your documents. What one company taught the room never reaches another |
| Your documents | This client’s contract and last year’s price list, both returned in the first round | Your company | Web pages. A page read often is still a page, never a fact about your company |
| Recorded failures | An earlier renewal, invented for the exercise, where a discount given to one client was quoted back by another | The room, when it was corrected | Anything unverified. A warning is written only after the failure was confirmed |
| Recent research | A competitor’s public price list, fetched a month ago, dated and addressed | The Chair, during a meeting | Nothing is promoted off it. Research may be cited, never filed as knowledge |
Exhibit 2. Illustrative. The four shelves: what each holds, who put it there, and what must never cross onto it.
The first round of the imagined meeting returned two documents, and the table shows why only two. The method came off the first shelf, the contract and price list off the second, and the failure note and competitor’s list were used as what they were. The warning shaped how Finance’s floor price was read. The competitor’s list was a month old and said so.
Now imagine the same four passages arriving from one pile, unlabelled. The competitor’s price list and your own are both lists of prices. The failure note describes a discount and so does your contract. A model handed all of it as plain text does what a person does with an unlabelled folder: reads the most confident sentence as the true one.
Two limits sit beside this. The shelves say where material sits and what it is called; they do not judge truth, so a wrong document on your shelf is still read as yours, and the correction is to fix the document. And the third shelf is only as useful as the habit of writing to it; a room whose mistakes are never recorded has a failure shelf that stays empty.
What this arms you to ask
Ask where your company’s material sits relative to theirs, and to anything the agent found on the web.
A good answer names the separations, or their equivalent, and shows you one: a retrieved passage with its origin attached, a web result carrying its date, an answer saying which facts came from your documents and which from outside. An answer that changes the subject sounds like this: your data is kept separate from other customers’. That is true of any competent vendor and answers a different question; customers apart is a different thing from shelves apart. A second answer, that everything goes into the knowledge base, tells you there is one pile.
The follow-up costs nothing. Ask a question whose answer is on the web and in your documents, and differs: a price, a headcount, a delivery time. Then ask which the agent used, and whether the reply said where each fact came from. A system with shelves can point; a system with a pile can only answer.
The fourth kind of tool, from the fourth paper, is where a wrong passage does its damage. The shelves are where it is caught first, and they are the cheapest place to catch it.
Next paper: Why a weak match is worse than no match.
Asked plainly
Where does an AI agent keep the documents I upload?
In a searchable store the model can query during a conversation. The question worth asking is what else sits in that store. The reference design described here keeps a company's own documents on a shelf of their own, apart from the vendor's method, apart from recorded mistakes, and apart from anything fetched from the web, and labels every passage with the shelf it came from.
Can an AI agent confuse a web page with my company's records?
It can, if both arrive as undifferentiated text. A model reads whatever it is handed with the same confidence. The defence is provenance: each retrieved passage carries the shelf it came from, so a sentence from a web page is read as a claim from outside, and a sentence from your contract is read as your contract. Research is also dated, so an old page cannot pass as current.
Is my company's data separate from the vendor's in an AI agent?
Separate from other customers is the usual promise, and it answers a different question. Ask whether your material is stored apart from the vendor's own knowledge and from anything the agent found on the web, and whether an answer can show which of the three each fact came from. If everything sits in one store, the agent cannot tell them apart, and neither can you.
Ask the vendor
The questions this part arms, each with the good answer and the answer that arrives instead.