Knowledge base
What is a knowledge base and why does it so often go stale?
A knowledge base is the recorded version of what your team knows. Without an owner and a rhythm it's partly wrong within six months.
In short
A knowledge base is an organised collection of articles with the answers to recurring questions: policy, procedures, product information and exceptions. It serves two readers at once, your customer searching on their own and your agent or AI pulling an answer from it, and that dual audience determines how you write it.
Why a knowledge base goes out of date
Knowledge bases rarely decay all at once. The return policy changes and the article doesn't. A carrier is added and the delivery page still names one. After six months there's information that's partly right, and that's harder than information that's clearly missing.
What helps is an owner per article and a review date. Not a big overhaul project, but ten minutes a week where the longest-untouched articles come by.
Why contradictory information is worse than missing information
If a 30-day return window sits in your knowledge base and 100 days on your product page, a customer can find both. What they do next we can't see, but the question of which of the two applies then reaches you, and it needn't have.
For an AI agent this is the same problem, just visible sooner. Two sources contradicting each other produce an answer nobody can say is right.
How to write for two readers
Write a complete answer per article that can be read on its own, without pointing to another article for the core. Name amounts, windows and exceptions explicitly instead of writing that conditions apply.
That isn't an SEO trick and not an AI trick either. An article a person understands in one go is exactly the article an AI agent pulls a correct answer from.
Internal knowledge base or public help centre?
| Internal | Public | |
|---|---|---|
| Reader | Your own agents and the AI agent | The customer, and search engines |
| Tone | Short, with exceptions and edge cases | Complete, without internal jargon |
| May contain | Margins, goodwill rules, supplier agreements | Only what you want to say publicly |
| Effect when stale | Agents give wrong answers | Customers contact you to verify |
Keep one source and two views. Write the policy once with the exceptions attached, and publish the part the customer may see. Two separate texts diverge within six months, and then the question is which of the two the AI read.
Where it goes wrong
Putting everything in one large article
A page with fourteen topics is read to the end by nobody and, in search, produces a hit the visitor then has to search within. One topic per article makes both searching and quoting possible.
Mixing internal procedures with customer information
What your team should do with a complaint is different from what your customer needs to know. Mixed together, you read customers internal steps while agents search fruitlessly among public texts.
What to do this week
Take your five most common questions and look up the answer the way a customer would, through your own site search. If you can't get there in two steps, or you find two answers contradicting each other, you've got your first five jobs.
What this looks like in a system
In Cuego the knowledge base is the source the AI agent draws on and the one your agents search during a conversation, without a second system. A customer version can sit as a help centre on its own address. When a question arrives for which no article exists, that shows up as a gap, so your knowledge base grows where your customers miss it.
Knowledge base & AIRelated terms
Self-service
Self-service is letting the customer arrange themselves what would otherwise be a conversation. It only works when it takes less effort than firing off an email.
Read onDeflection
Deflection is the share of questions that never reached an agent. It's the easiest number to overstate and the most useful to measure honestly.
Read onCanned response
A canned response saves typing on questions that keep returning. It becomes harmful the moment it answers a question that was slightly different.
Read on