What is an AI knowledge base, and what must your agent find there?
An AI knowledge base is the place where the facts about your company that an agent must know are written once: who your customers are and how they are named in each tool, which documents are authoritative, which rules apply, and where to find everything else. It begins with a decision: write down what three members of your team know without telling one another. No software will do it for you.
The technical term you will encounter is “RAG”: the agent retrieves information from your documents before answering instead of answering from memory. What matters is the order of operations: first write down what is authoritative, then connect the system. A knowledge base built on contradictory documents produces contradictory answers, faster than before.
What it costs to leave documents scattered
The cost appears on no dashboard, and that is what makes it dangerous. It takes three forms.
The agent rediscovers everything, every time. Because it does not know where to look, it searches: several queries where one would be enough, waiting at each step, and a bill that rises with the number of documents reviewed. This is the problem OpenAI describes in the case published on September 21, 2026 with publisher V7: agents must “rediscover context for every request.”
That context, reread on every turn, also determines your bill: see what an AI agent really costs.
The answers are plausible and wrong. The most expensive case. An agent that builds a price from an obsolete price list produces a credible, polished, sendable, and incorrect quote. No model can guess that a 2024 file was replaced.
The task returns to a human. After the third answer that needs correction, your team takes over again. The system remains connected, and no one uses it anymore.
What a platform like V7 does, and what it teaches us
The case published by OpenAI is valuable for its mechanism, not its figures. V7 Go connects to a company's document repositories (SharePoint, Google Drive), extracts entities, relationships, and facts, and organizes everything in a “Context Graph”: a structured register the agent queries directly, with sources preserved. Alberto Rizzoli, V7's cofounder and CEO, summarizes the argument this way: “to solve difficult enterprise use cases in finance and insurance, AI needs to learn how your business works as well as it learned from the Internet.” The figures highlighted, 89% accuracy on the hardest queries and 78% lower cost per document, are claims from the provider and its customer, not independently measured.
This platform targets finance, insurance, and real estate, across millions of files: it is not for you. What it demonstrates is: performance comes from organizing context, not the power of the model. At your scale, that work takes a few pages.
The context sheet: what to write before any tool
One sheet per activity (incoming inquiries, quotes, prospecting), maintained by the person who does the work. Five sections, and one rule: each line states its source, or it does not appear.
| Section | What to write | Example question it answers |
|---|---|---|
| What is authoritative | For each type of information, the single document or software tool that is correct, and who keeps it current | “The applicable price is in management software, not the salesperson's spreadsheet” |
| Names | Companies, products, and references that have several names depending on the tool, with the selected name | “Dupont SARL, DUPONT S.A.R.L., and Ets Dupont are the same customer” |
| Rules | What the team applies without thinking and no one has written down | “When the lead time exceeds three weeks, we warn the customer before pricing” |
| What is obsolete | Documents that remain around and must no longer be used as references | “The 2024 schedule is archived” |
| Who decides | The person to whom the case is escalated when two sources conflict | “A discount beyond the schedule goes to the founder” |
Allow one to two hours per activity, interviewing the person who performs it. This is where our automation work begins, and it is often when the founder discovers two contradictory rules applied by two employees. The same principle applies inside our company: our agents never copy a name or address from memory; they read it from a single reference file. This is also what our inquiry and quote agents read when they prepare a sourced proposal before a person approves sending it.
Separate what the agent must decide from what it must write
A development from the same week helps explain this point. On September 21, 2026, Simon Willison described Jev, unveiled the previous week by TypeSafe AI: a model that writes nothing. It accepts text as input and returns numbers: yes or no with a confidence level, a choice among options, or a score on a scale. That is all, and it is very inexpensive: the provider charges $0.042 per million input tokens, with output free. Willison raises a clear concern: while a conventional model can at least be asked to justify its answer, Jev returns only a number, making its biases difficult to detect, which is why he recommends systematic evaluations. We draw another conclusion: keep a person on cases that commit the company.
In your process, some tasks are decisions (is this inquiry in our scope, should this quote be followed up) and others are writing. The costs differ. Separating them before choosing a tool prevents you from paying a writing model to sort messages, and it is done on paper, with the context sheet in front of you.
Its limits: when it is not worth it
This preparation does not replace a technical choice: at some point the tools must be connected, and some models are better than others for certain tasks. It also does not fix an inconsistent process: if the same inquiry follows three paths depending on the person handling it, you must decide first.
And sometimes the work is not justified. If your activity fits in one well-maintained software tool, you already have your knowledge base. If the target process occurs three times a month, the preparation will cost more than the task. This is what our engagement framework says: when there is nothing useful to build, we say so before starting.
FAQ
Do you need a tool to create an AI knowledge base?
Not to begin. The first context sheets fit in a shared document or spreadsheet. A tool becomes useful when the volume exceeds what one person can maintain by hand, and you will then know exactly what you are asking it to do.
Will changing models solve the problem?
Not this one. A more capable model reasons better over what it is given; it does not guess which price list is current or that two spellings refer to the same customer. When an agent is wrong, first look at what it had in front of it.
Must everything be centralized in one place?
No, and that is rarely possible. What matters is that it is written which source is authoritative and where to find it. Documents can remain where they are.
We review one task, your tools, and the decisions that must remain human. The written conclusion is a simple automation, an AI agent, or nothing useful to build.
Sources: “How V7 gives AI agents institutional memory,” OpenAI, September 21, 2026 (openai.com); the Context Graph description and Alberto Rizzoli quotation come from it; the 89% and 78% figures are claims from the provider and its customer, not independently measured. · “Jev introduces a new shape of LLM, System One, aka Decision Models,” Simon Willison, September 21, 2026 (simonwillison.net), for the nature of the outputs, input price, and concern about the absence of justification. · Search volumes: Ahrefs, France, recorded September 22, 2026. Practices described as ours come from our team charter and job descriptions. No customer is named and no result figure is claimed.