Readiness · the starting point
Is your knowledge AI-ready?
Most teams are asking which AI to buy. That is the second question. The first one is what the model will be repeating, because it will repeat whatever you already have, with the confidence of something that was checked.
Before you trust AI, trust your knowledge.
Illustrative figures. Your own come out of the scan.
The question underneath
You are not buying a model.
You are buying whatever it repeats.
So the first question is not which AI.
It is whether your knowledge can survive one.
Six questions below. They are about ownership, dates and versions, not about technology, and your own team can answer all six from memory. That is rather the point.
The scan
Six questions your knowledge base has to survive.
Not about your AI. About what it will be quoting. Answer honestly, the score moves while you go, and nobody has to give an email address to see it.
Three about the answer itself Who carries it, when it was last checked, and what it said before.
01
Can you name the team that owns your twenty most-asked answers?
Not the system they live in. The people whose job it is when one turns out to be wrong.
02
Does every answer carry a date by which it has to be checked again?
And does something actually happen on that date, to the owner rather than to the customer.
03
If an answer changed last month, can you see what it said before?
A regulator asking what you told people in March is a version question, not a search question.
Three about where it ends up The first three were about the answer itself. These are about the copies.
04
When a policy changes, how many places do you have to update?
Count the intranet, the help center, the agent scripts, the training deck and the one spreadsheet.
05
Does the same question have more than one answer somewhere in your systems?
An assistant cannot tell which of two near-identical answers is the current one. It will pick.
06
Do your chatbot, your agents and your portal give the same answer today?
Same question, three channels, this morning. Would the three replies match word for word.
What a 42 looks like
The answer is not wrong. It is just old.
This is a real shape of answer, from a real knowledge base. Nothing about it looks broken. It reads well, it has a version number, someone once approved it. And it has been served every day for eleven months since anyone last checked.
A score in the forties almost never means the content is bad. It means nothing on the outside tells the good from the stale.
Asked 1,840 times since the last review
Excess for a referred treatment
The excess of €385 applies per calendar year and is settled with the first declaration. Treatment by a contracted provider does not change the amount.
The excess went to €405 in January. Three channels have been saying €385 ever since.
Where the score leaks
Three gaps, and the first move for each.
In almost every scan the points disappear in the same three places. None of the three is a content problem, and none of them needs a tool to start on.
Ask who owns it and you get a system, not a team.
When an answer turns out to be wrong, the fix waits for someone to volunteer. Usually that is the person who happened to hear the complaint.
Put a team name on the twenty most-asked answers. One afternoon, no budget.
Everything was correct on the day it was written.
An answer without an expiry date never expires. It just gets old while it keeps being served, in every channel, to everyone who asks.
Give each one a next-review date, and let the owner hear about it instead of the customer.
The same question, answered in four places.
An assistant cannot tell which of two near-identical answers is current, so it picks one. Sometimes it blends them into a third answer that never existed.
Pick the one that is right. Let the other places point at it instead of copying it.
Now run it on your own knowledge base.
What you just did is the honest version, from memory. The real one reads what you actually have, scores it per collection, and shows you which answers an assistant would pick up today.
What it is worth
60%
lower contact center volume, right after go-live
“Polly.Help has translated into a 60% reduction in contact center volume, right after implementation.”
Already running on a foundation that was checked
