For years, knowledge bases in customer service have been relatively straightforward. People write procedures, instructions, FAQs and internal notes. Other people read them. The structure may not always be perfect, information may occasionally be duplicated and some documents may be outdated. But experienced employees know how to navigate around these imperfections.

AI changes this.

When generative AI becomes part of customer service, the knowledge base is no longer just an internal documentation tool. It becomes part of the infrastructure that determines what AI knows, what it can retrieve and, eventually, what it can tell a customer.

AI does not eliminate the need for a knowledge base. It turns it into infrastructure.

From a human knowledge base to an AI knowledge architecture

This question quickly becomes very practical. Imagine a customer relations knowledge base containing around 230,000 words, organized into 17 main sections with 3 to 24 subsections each, and continuously maintained and updated by experienced colleagues.

Traditionally, such a knowledge base was designed for one specific user: a human employee.

Now the same knowledge may also be used by AI to prepare draft responses, assist agents during conversations or eventually answer customers through e-mail, chat or voice. This creates a different problem. A human employee can usually recognize outdated information, understand an internal note or notice when two instructions contradict each other. AI needs much clearer rules.

The quality of generative AI does not depend only on the model. It also depends on the quality, structure and governance of the knowledge behind it.

But quality alone is not enough.

Knowledge also needs context. AI needs to understand not only what a piece of information says, but also what it relates to, who it is intended for and where it can be used.

This is where a very practical question appears:

How do we make an existing knowledge base usable by AI without creating a maintenance nightmare?

Three possible approaches

Looking at the problem, I see three broad ways to organize the knowledge itself.

1. One knowledge base with metadata

Employees and AI continue to use the same knowledge base. Individual pieces of content receive metadata defining how they can be used: PUBLIC or INTERNAL, EMAIL or VOICE, product, intent, owner, validity date and so on.

Advantage: one version of the knowledge to maintain.

Challenge: filtering and access rules become critical when internal and customer-facing information exist in the same environment.

2. Internal knowledge base with a separate AI knowledge base

Employees continue using the internal knowledge base, while AI receives a separate, curated knowledge base containing only approved information.

This clearly separates internal knowledge from information that AI may use with customers. It also makes it possible to use different AI systems, models or rules for different purposes.

The price: two versions of knowledge that must remain synchronized over time.

3. Master knowledge base with controlled AI knowledge layers

One master knowledge base — the official source of truth. From that master source, different versions of the knowledge can be prepared for customer-facing AI, internal agent-assist, e-mail or voice.

This requires more work initially: metadata, publishing processes, versioning, ownership and technical integration.

One source of truth, with different versions of the knowledge prepared for different AI uses.

For me, this is where the discussion shifts from documentation to architecture.

RAG does not replace the knowledge architecture

Choosing how to organize the knowledge is only one part of the problem. AI still needs a way to retrieve the relevant information.

One common approach is RAG — Retrieval-Augmented Generation. Documents or selected knowledge are indexed, broken into smaller pieces and made searchable for generative AI. In the simplest implementation, RAG can be placed directly on top of existing documentation without fundamentally redesigning the underlying knowledge base.

This often requires changes to the content itself — clearer structure, smaller self-contained topics, consistent terminology and enough context for AI to interpret the information correctly. These changes can make existing documentation easier for AI to use, but they do not fix the underlying knowledge management problem.

If the source documentation contains duplicated information, outdated instructions, internal notes or content that should not be communicated directly to a customer, those issues remain part of the source knowledge.

RAG can improve how AI retrieves knowledge, but it does not by itself solve the quality, context and governance of the underlying knowledge.

The real problem is maintenance

Whichever approach an organization chooses, I think one principle should remain constant:

Knowledge management should require as little duplication, synchronization and manual maintenance as possible.

This becomes particularly important once several AI use cases appear. Today it may be an e-mail assistant. Tomorrow it may be agent-assist, chat or voice. If every new channel creates another copy of the same knowledge, keeping everything aligned becomes harder each time AI expands.

Where should the knowledge actually live?

There is also another practical question.

Organizations can build AI on top of their own knowledge base or make part of that knowledge available through solutions provided by voice, e-mail or chat vendors. Vendor solutions can simplify implementation. But they may also introduce another knowledge layer that needs to be maintained or synchronized. And vendors change. Models change. Communication channels change.

The company should therefore keep its core organizational knowledge unified and under its own control.

Technology may change quickly. Organizational knowledge should not have to move with every technology change.

From document management to knowledge architecture

For many years, companies could treat a knowledge base primarily as documentation. With AI becoming part of customer service, that is no longer sufficient. The question is no longer only whether the information is correct.

It is also whether the company knows:

That is the transition I find most interesting. From maintaining documents to designing a knowledge architecture.

And infrastructure needs architecture.

How is your organization making that transition?