Where does my company’s material sit relative to yours?
You handed over a contract in the second demonstration. Drag, drop, a progress bar, and somebody said the agent now knows your business. What nobody said is where the file landed, or what it is now sitting beside. If the answer is that everything goes into one store, then your contract, the vendor’s own guidance and a web page fetched last month are, from the model’s side, three passages of undifferentiated text.
Origins survive, or they do not
Every agent that reads documents searches some store. The setting worth asking about is how many stores there are, and what each one is allowed to hold.
The reference design keeps four, recorded on the Register, the standing page of every setting in ArkOne’s reference design for an executive agent. House knowledge, meaning the vendor’s method, holds how a pricing decision is weighed. Your documents holds what your company produced or approved. Recorded failures holds the occasions the room was wrong and somebody said so. Recent research holds whatever was fetched from outside, carrying its address and its date, permanently.
The separation buys one thing, and it is the thing your question is really after. A passage arrives labelled with where it came from. Your contract says becomes a different kind of sentence from a web page says, and a model that receives the labels can pass them into the reply. Pour the four into one and the labels are the first thing lost.
Four passages, two filings
Suppose the renewal invented for this work, in which a client asks to keep last year’s price. Four passages reach the agent. The middle column files them by origin. The right-hand column is the same four in one store, which is what the model then reads.
| Passage reaching the agent | Filed by origin | Filed in one store |
|---|---|---|
| This client’s contract | Your documents | Text about a price |
| How a discount is weighed | House knowledge, the vendor’s method | Text about a price |
| A competitor’s public list, fetched a month ago | Recent research, dated and addressed | Text about a price |
| An earlier discount that was quoted back at you | Recorded failures, written after the event | Text about a price |
Exhibit 1. Illustrative. The same four passages, once labelled by origin and once poured into a single pile.
Four distinguishable things become one on the right, and the count is the whole argument. Every passage in the right-hand column is a sentence about pricing, and the model has no way to tell which of them describes your business.
What it does then is what a person does with an unlabelled folder. It reads the most confidently worded sentence as the true one. The competitor’s list is a month old and reads exactly as firmly as your own contract. If those two prices disagree, one of them shapes a reply going to your client, and nobody in the room can say which.
The answer that can point at a source
A good answer names the separations, or whatever the vendor’s equivalent is, then demonstrates one. A retrieved passage with its origin attached. A web result carrying the date it was fetched. A reply that says which figures came from your documents and which came from outside.
The ordinary answer changes the question without seeming to. Your data is kept separate from other customers’. That is true of any competent vendor, and it is about tenancy, not about origins. It answers where other companies cannot reach, and you asked where the web is filed relative to your board minutes. The second ordinary answer is that everything goes into the knowledge base, which is one pile stated plainly.
The follow-up costs nothing and needs no access. Find a figure that exists in both your documents and on a public page, and where the two disagree. Ask a question that needs the figure. Then ask which source the agent used. A design that files by origin can point at one. A design with a single store can only give you the answer again, in the same voice, which is exactly the failure you were testing for.
One limit belongs in the same sentence as the capability. Shelves record where a passage came from; they say nothing about whether it is true. A wrong document filed correctly under your own material is still read as yours, and no separation catches that. What separation catches is a stranger’s sentence wearing your company’s authority, which is a different and commoner fault. The relevance question sits beneath this one.
Asked plainly
Where do the documents I give an AI agent actually sit?
In some store the system searches when it needs something it was not given. What differs between designs is whether that store is one pile or several separated by origin. Separation is what allows a retrieved passage to arrive labelled as your document, the vendor's own method, or a page fetched from the web, and a label is what lets a reply say where each fact came from.
Is keeping my data separate from other customers the same as separating it from the vendor's material?
No, and the two are often confused in a sales meeting. Keeping customers apart stops your contract reaching another company. Keeping origins apart stops your contract being read in the same breath as a web article or the vendor's generic guidance inside your own answers. A vendor can do the first perfectly and still pour everything into one store for you.
How do I test whether an AI agent knows where a fact came from?
Pick a figure that appears both in your own documents and somewhere public, and where the two disagree: a price, a headcount, a delivery time. Ask a question that needs it, then ask the agent which source it used. A design that files material by origin can point at one. A design with a single pile can only repeat the answer.