Knowledge base
A knowledge base is the store of articles a support team answers from, covering policies, procedures, and product information. In an AI deployment it is also the content the agent retrieves from, which makes its accuracy a live operational concern rather than a documentation one.
A customer support knowledge base usually exists in two versions: the public help centre customers read, and the internal material that both human agents and AI agents work from. The internal version carries the detail that matters and is normally the messier of the two.
What goes in a customer support knowledge base?
Most knowledge bases hold four kinds of content, and they age at very different rates.
Product and policy information: fees, limits, eligibility, and terms. Changes rarely, and matters enormously when it does
Process documentation: how a dispute is raised, what a KYC review needs, how long a transfer takes
Response guidance: tone, required wording, and what must never be said
Edge cases and exceptions: the material that usually lives in someone's head rather than in the system
The fourth category is where most AI deployments run into trouble. It is the content least likely to be written down and the content an AI agent most needs, because the straightforward cases were already covered. A useful audit question: when an experienced human agent handles an unusual case well, where did they learn how?
Why is a knowledge base not enough for an AI agent?
A knowledge base describes what is true in general. Most customer contacts are about something specific, and the gap between those two is where deflection stalls.
A customer asking why their transfer has not arrived cannot be answered from an article about transfer times. Answering properly needs the transfer record, its status, and the policy applied to that status, together. The article explains the rule, and the customer wants the ruling. Three limits follow:
No account data: anything customer-specific sits out of reach
No actions: the agent can describe how to raise a dispute and cannot raise one
Documentation drift: the written process and the actual process diverge, and agents follow the actual one
Our blog on why your knowledge base is not enough goes further into that gap.
Frequently asked questions
Does an AI agent replace our knowledge base?
No, it raises the stakes on it. An agent retrieving from your knowledge base inherits whatever is wrong or missing in it, and does so at volume. Most teams find that deployment surfaces content problems they already had.
How much knowledge base work is needed before deploying an AI agent?
Less rewriting than teams expect, and more gap-filling. Modern retrieval copes with imperfect structure and inconsistent formatting reasonably well. What it cannot do is retrieve content nobody ever wrote down, so the effort belongs on missing material rather than on tidying existing articles.
Should our public help centre and internal knowledge base be the same?
They serve different readers and should stay separate. The public version is written for customers, while the internal version carries procedural detail, thresholds, and exception handling customers should not see. An agent usually needs both, with clear rules about which content may be quoted to a customer.