Every business has someone who just knows. At a hydraulics supplier in Singapore it was the people who answer enquiries: they read a few lines of part description, sometimes only a photograph, and find the right item in a catalogue of more than fifteen thousand rows. That match is the whole job. It is slow, and almost none of it was written down.
This is how we turned that knowledge into something the business keeps.
Where does the knowledge come from?
From the mailbox, not from interviews. Every enquiry the team has answered sits next to the quotation they actually sent, and that pair is the ground truth: what the customer wrote, and what an experienced person decided it meant.
Talking to the staff still matters, because people explain their exceptions well. But what they say is checked against what they did, and when the two disagree, what they did wins.
How is it written down?
As rules a person can read: each has a number and a plain sentence, and a rule taken from staff practice says how many past quotations it rests on. "When a line says steel for a valve, staff quote carbon steel" is a rule only if the past quotations show it, and the rule says how many of them do.
The rules live on Knowledge pages inside the client's own CRM. The staff read them, and they are the ones who say when a rule is wrong. A rule nobody can read is a rule nobody can correct.
What does the AI do with it?
QuotationAI reads a new enquiry, splits it into lines, and for each line suggests the catalogue item it thinks is meant, with the reason: the words it read, the words each candidate explains, and what colleagues quoted before for the same words. A person picks. Nothing is sent on its own.
When a suggestion is wrong, the panel shows which word was misread. The fix is a new or better rule, which every later enquiry benefits from, rather than an override that teaches nothing. The full tour is in QuotationAI, opened up.
How do you know it got better?
By counting, round after round, on the same answered mailbox. A round is a set of rule changes; each is run on every past enquiry, saved, and compared case by case with the round before. The article Measured, not claimed describes the method.
In round 12, the first suggestion was the item staff had actually quoted on 1,507 of 3,483 answered items. That is under half, and it is a real count: each of those items can be opened and checked.
Whose knowledge is it?
The client's. The rules and the measured runs sit in their own workspace, written in their own trade words, and their staff are the ones who confirm or correct them. What their people know is now written down, which was not true before.
Is this only for quoting?
No. Quoting is where the knowledge is densest and a wrong answer costs most, so it is where we started. The same approach suits any work where experienced people make many small judgements against a large reference: matching, classifying, replying. The first conversation is about how your team works today and what it knows that nobody has written down.