How to build a knowledge base that AI can actually use

Companies have never produced so much documentation. Operations manuals, procedures, training materials, meeting minutes, FAQs, legal documents, presentations, how-to sheets, business guides, intranets… In a multi-site network, the amount of available information can quickly become considerable.

With the arrival of artificial intelligence, a particularly appealing promise has emerged: letting every employee ask a question and immediately get the right answer, without spending twenty minutes searching through several folders or calling someone at head office.

On paper, the principle seems simple: centralize the company’s documents, give an artificial intelligence access to this documentation and let it answer users’ questions. In reality, things are a little more complex. Because a collection of documents is not yet a knowledge base. Having digitized all of a network’s procedures does not mean an AI will be able to understand them correctly, find the right information at the right time and, above all, deliver an answer reliable enough to be used in the field.

This is often the moment when a deeper problem comes to light. The issue is not necessarily artificial intelligence: the issue is the company’s knowledge itself.

The 5 pillars at a glance

1. Clarity: explicit rules that leave no room for interpretation.
2. Structure: a common logic shared by all documents.
3. Traceability: a date, a version and an owner for every piece of content.
4. Chunking: precise blocks that keep their context.
5. Governance: clear rules for creating, approving, updating and removing content.

AI often reveals a documentation problem that already existed

Three common situations

Imagine a store manager who wants to know which procedure to follow when a customer asks for a refund in a particular situation. They ask their network’s AI assistant, which finds two documents. The first states that a refund must be approved by the store manager; the second explains that approval from the national customer service team is mandatory above a certain amount. Neither clearly states which one is the most recent. Artificial intelligence then faces the same problem as a human employee: it has several pieces of information, but does not know which one is authoritative.

Another common case: a procedure explains how to carry out an operation in “stores”, without specifying that it actually only applies to company-owned stores and not to franchisees. Or an old version of a procedure remains accessible in a document space even though a new version was released six months earlier.

In each of these cases, technology is not really to blame. It is the knowledge itself that lacks clarity, structure or governance. So the arrival of AI does not create these difficulties: it simply makes them much more visible.

What AI cannot guess

For a long time, these inconsistencies were offset by employees’ experience. Someone at head office knew the document was out of date, a network coordinator knew about the exception, a manager explained verbally that “in this specific case, we do things differently”.

An artificial intelligence does not spontaneously have this informal memory: it works from what it is given. This is an essential rule to keep in mind: an AI cannot properly invent a process that has never been formalized, and it will never be more up to date than the knowledge it has access to.

Building a knowledge base suited to AI therefore means working on five major topics: clarity, structure, traceability, information chunking and document governance.

1. Clarity: an AI should not have to guess what the author meant

The first principle may seem obvious: a procedure must be clear. Yet many internal documents have been built up gradually, sometimes by several people, at different times and with different goals. You come across phrasing such as “If needed, notify the manager.” But which manager? The store manager, the network coordinator, regional management, head office? And what exactly does “if needed” mean?

For an experienced employee, the answer seems obvious because they know how the organization works. For a new hire, much less so. And for an artificial intelligence, this ambiguity becomes a potential source of error.

A documentation base that AI can use must therefore lean towards explicit information: who does what, in which situation, at what point, within which limits, under which conditions and with which exceptions?

Take another phrasing: “Important requests must be forwarded to head office.” The intention is understandable, but the information is not usable as it stands. What is an important request? A request over 500 euros, a request involving a legal risk, a request from a strategic customer? AI works much better when it has rules precise enough that it does not need to interpret what should have been formalized.

Preparing a knowledge base therefore often starts with a rather unspectacular step: rereading existing documents and asking whether someone who does not already know the network would really be able to understand them. It is an excellent test. If a new employee systematically has to ask a colleague what a procedure means, there is a good chance an AI will run into exactly the same difficulty.

2. A consistent structure: helping AI understand where information is

The second step is to give content a coherent organization. In many companies, documents have piled up over time and in very different forms:

• a procedure written in Word;
• another in a PowerPoint presentation;
• a third in an FAQ;
• a fourth in a sixty-page PDF;
• an important piece of information that exists only in the minutes of a meeting held two years ago;
• another in an email that only three people have kept.

For a human, this scattering is already a problem. That is precisely the challenge of centralizing information across a franchise network without making management more complex. For an AI, it also makes it harder to identify the right information and understand the context in which it should be used.

The goal is not to turn every document into perfectly identical forms. Rather, it is about establishing a common logic. A procedure could, for example, always specify:

• its objective;
• the people concerned;
• its scope;
• the different steps;
• any exceptions;
• related documents.

This consistency brings a twofold benefit. It helps employees navigate the documentation, and it makes the knowledge easier for AI to use, because the information is organized in a more predictable way.

Take the example of a network with different procedures depending on store formats. If each document clearly states, in the same place, which locations it applies to, AI will be able to select the right answer much more easily. If this information sometimes appears in the title, sometimes in a footnote and sometimes nowhere, the risk of error increases.

Structuring knowledge therefore does not simply mean filing documents in folders. It means giving each piece of information enough context to understand what it means and when it applies.

3. Traceability: knowing which information is truly authoritative

This is probably one of the most common problems. A company updates a procedure and releases the new version, but the old one remains available somewhere. A few months later, both documents come up in a search: the employee hesitates, and so does the AI.

To build a reliable knowledge base, every important piece of content should be able to answer three simple questions:

• When was this information last updated?
• What is its version?
• Who is responsible for it?

These elements may seem administrative, but they become essential as soon as AI is used to access the company’s knowledge. A procedure dating from 2023 and one updated in 2026 should not be treated as equivalent. You need to be able to identify the authoritative version and, when a piece of information seems inconsistent, know who can check or correct it.

Traceability matters even more in networks whose procedures change often: a new regulation, a change in commercial offer, the rollout of new software, a modified opening procedure or approval rule. In each of these cases, the AI assistant must not only know the new information: you also need to make sure the old information no longer interferes with its answers.

Documentation work therefore does not end when a new version is published. You also need to organize the removal, archiving or deactivation of content that has become obsolete.

4. Smart chunking: finding exactly the right information

Another mistake is to believe that giving an AI complete documents is enough for it to always find the relevant information. Imagine a 250-page operations manual covering opening, sales, human resources, customer relations, local marketing, orders, disputes and closing the store. All the answers may be in there, but they are not necessarily organized precisely enough to be found easily.

Building a good knowledge base therefore means thinking about the level at which information should be structured:

• a complex procedure can be split into clearly identifiable steps;
• a long FAQ can be organized around specific questions;
• a manual can be divided into topics self-contained enough for an answer to be found without interpreting dozens of pages.

This chunking should not go too far, however, because information taken out of context can become misleading. Take the rule “Approval from the regional manager is not required.” If this sentence is separated from the paragraph specifying that it only applies to orders below a certain amount, it can lead to a completely incorrect answer.

Good chunking is therefore a matter of balance: each block of knowledge must be precise enough to be found easily, yet complete enough to retain the context essential to understanding it. This work is invisible to the end user, but it is often what makes the difference between an assistant that gives approximate answers and one that actually finds the relevant information.

Discover the Cerca knowledge base

5. Document governance: the condition for the knowledge base to stay useful over time

This is probably the least spectacular point, and yet perhaps the most important. A knowledge base is never finished. The company evolves, processes change, new products appear, teams reorganize, rules become obsolete and exceptions are created. Building a quality documentation base at a given moment is therefore not enough: you need to know how it will live on.

A few questions help lay the foundations of this governance:

• Who can create a new procedure?
• Who must approve it?
• Who is responsible for updating it?
• How often should certain information be reviewed?
• What happens to old versions?
• How do you report information that has become incorrect?
• Who decides when two documents contradict each other?

Without governance, a knowledge base gradually deteriorates. It starts with clean, consistent documents, then versions multiply, updates are no longer systematic and users lose trust. And trust is fundamental: if an employee gets three wrong answers within a few weeks, they are very likely to stop using the assistant.

The quality of an AI experience therefore depends as much on the model used as on the discipline with which the company’s knowledge is maintained. It also depends on your teams’ ability to use the tool day to day: that is the subject of our article on training your teams in AI.

The hardest case: what everyone knows but nobody has written down

Finally, there is a particularly important topic in networks: informal knowledge. Every organization has a huge amount of it. It is all the information that experienced employees know but that does not appear in any document:

“In this situation, you’d better call Sophie.”

“For this type of store, we usually do it like this.”

“This procedure works except during the Christmas period.”

“Officially, the document says this, but since last year we’ve been applying a new rule.”

This knowledge often keeps the organization running, but it is also a weakness, because it depends on the people who hold it. When an employee leaves the company, part of this know-how can disappear with them. And when you deploy an AI assistant, the problem becomes immediately visible: AI does not know what the company itself has never taken the time to formalize.

Deploying an AI-assisted knowledge base can therefore have an unexpected benefit: forcing the organization to identify what it really knows. Which information is documented? Which practices are only passed on verbally? Which official procedures no longer really match reality? Which people currently hold critical knowledge? This work goes far beyond the AI project: it strengthens knowledge transfer across the entire network.

Before connecting AI, test the quality of your knowledge

A simple exercise can be particularly revealing, and it can be carried out before you even choose a tool.

Try it in your network: the ten-question exercise

1. Pick ten questions frequently asked by your frontline teams. Not theoretical questions, but real everyday ones: “What should I do when a customer disputes this situation?”, “Which procedure should I follow to order this product?”, “Who needs to approve this expense?”, “How do I report this incident?”, “Which rule applies to my type of location?”
2. Try to find each answer using only the available documentation.
3. Analyze the result:
   • three people reach three different answers? The problem exists before AI is even brought in;
   • the answer systematically requires calling someone? Information is probably missing;
   • two documents contradict each other? A decision will have to be made between them;
   • nobody knows which document is the most recent? Traceability needs to be improved.

This approach turns the logic around. Instead of only asking “How do we connect our documentation to an AI?”, you start by asking “Does our documentation really allow us to answer our teams’ questions correctly?” That is probably the best starting point.

A good knowledge base benefits humans first

This may be one of the most important lessons: the efforts made to make knowledge usable by AI also improve how employees use it.

• A clearer procedure is easier to understand.
• Better-structured documentation is easier to browse.
• A clearly identified version reduces errors.
• Better-defined governance prevents obsolete documents from piling up.
• Formalizing informal knowledge makes onboarding new employees easier and reduces dependence on a few long-standing experts.

In other words, preparing your documentation for AI often comes down to simply improving how your documentation is organized. That is excellent news, because the work done keeps its value even as technologies evolve. AI models will change, so will interfaces, and assistants will become more powerful. But clear, reliable, structured and up-to-date knowledge will always remain an essential asset.

AI does not automatically turn documents into reliable knowledge

We will hear more and more about assistants capable of querying all of a company’s documentation. The technology will progress quickly and make accessing information much easier. But the quality of the answer will always depend in part on the quality of the available knowledge:

• if two procedures contradict each other, AI should not have to decide on its own which one reflects reality;
• if a rule is out of date, it may return outdated information;
• if a process has never been written down, it will not be able to guess it reliably;
• if a document does not specify its scope, it risks applying it in the wrong situation.

That is why, at Cerca, we believe that thinking about AI Agents cannot be separated from thinking about the network’s knowledge. The goal is not simply to let an artificial intelligence read more documents, but to let it access the right knowledge, in the right version, with its context, and for the right person.

Artificial intelligence can considerably simplify access to a network’s know-how, but it does not automatically turn confusing documentation into reliable knowledge. Before asking AI to better understand the company, you sometimes need to start by helping the company better organize what it knows.

Request your 30-day free trial with no commitment