Turning Senior Staff Experience into an AI Knowledge Base: Handover and Validation

Hypothetical example: a quotation coordinator's handover before retirement

The following example is fully hypothetical, for illustration only, and is not a client case.

A Hong Kong printing and packaging company has a quotation coordinator of many years who retires next quarter. The successor finds that the quotation system holds only a basic price list, while real quotes are adjusted all the time:

• Certain rush orders carry a surcharge, but "how urgent counts as rush" was never written down.
• Some long-term customers have special terms that appear only in old emails.
• Certain materials have longer lead times in particular months, so quotes must allow extra time.
• Above a certain amount, the boss must be consulted first, but people remember the threshold differently.

Handing all the old emails and quotes straight to AI risks it learning outdated terms, treating one-off exceptions as the norm, or even quoting customer personal data in drafts. The right first step is to extract the judgement rules.

Step 1: Inventory the judgements, then run structured interviews

Do not ask the senior colleague "what do you know?". Start with the 10 to 20 decisions that are asked about most, go wrong most, or affect amounts most. Then interview each with the same questions:

1. How do you usually decide? Is there a numeric threshold?
2. What are the exceptions? What happened the last time there was one?
3. Why is it done this way? Who decided, and since when?
4. Where does the information come from: price list, contract, email or personal experience?
5. Which situations always need someone else's decision?

In interviews, "why" matters more than "how". Recording only the method leaves a successor unable to adapt when a case is slightly different; recording the reason lets them judge whether the rule still applies.

Step 2: Write each rule as a knowledge card

Each knowledge card covers a single judgement rule and contains these fields:

• Situation: when it applies (for example, orders asking for less than the standard lead time).
• Rule: how it is usually handled, in clear, checkable sentences.
• Exceptions: known exceptions and how to handle them; anything unknown goes back to a person.
• Source: price-list version, contract clause, or "senior colleague experience (interview date)".
• Owner: who may change and approve this card.
• Version and effective date: so you can trace which version an AI draft relied on.
• Human approval conditions: for example amounts above a threshold, new customers or non-standard terms.
• Privacy level: public, internal or restricted. Cards hold rules only, never customer names, phone numbers or other personal data.

The number of cards does not matter. Twenty clear, owned cards are more useful than 2,000 unsorted documents. This is one practical form of turning company data into AI fuel.

Step 3: Validate on anonymised past cases

Writing the cards does not mean AI will apply them correctly. To validate:

1. Pick a batch of completed past quotes (say a dozen or so) and first remove customer names, contact details and other personal data.
2. Give AI only the approved cards and the case conditions, ask it to draft the quotation note, and require it to list which cards it cited.
3. The senior colleague or owner compares each AI draft with what was actually done and sorts differences into three types: card wrong or outdated, exception not covered by any card, AI cited incorrectly.
4. For the first two, update the card and raise its version; for the third, adjust how AI is used or make that situation human-only.

Validation is not about proving AI is clever. It finds what the cards fail to state clearly. It is most valuable while the senior colleague is still there, because only they can explain the differences.

Step 4: Human approval, versions and privacy

Once a knowledge base is live, the most common risk is not AI getting one answer wrong. It is a wrong rule being reused for months. We recommend:

• Only approved versions may be used by AI; keep draft and retired cards separate.
• Every card has a review date, and must be re-approved when prices or terms change.
• AI drafts must show which card and version they cited, so managers can check them.
• Drafts involving amounts, contractual commitments or new customers must be approved by a person before they go out; see AI agent workflow automation: from pilot to governance.
• Privacy: cards contain no personal data. If work requires customer records, look them up in the original system under its access controls rather than copying them into the knowledge base. Hong Kong companies handling personal data should follow guidance from the Office of the Privacy Commissioner for Personal Data and company policy.

How does a team learn to do this?

Knowledge handover is a skill that can be practised. Interviewing, writing cards, designing past-case tests and reviewing differences all become familiar once managers and staff have done them together once.

To build this method into team training, see the "capture experience" exercise in enterprise AI training, and use the training selection and acceptance method to set post-course checks.