Direct answer
Work backwards from the question list and prepare four categories: (1) high-frequency Q&A — opening hours, store addresses, return and exchange processes, price wording; (2) policies and rules — service commitments, after-sales policy, compliance requirements; (3) business facts — product catalogue, specifications, availability; (4) escalation rules — which five scenarios must transfer, and how the human side picks them up. Three things must be done alongside: mark restricted material, decide which source is authoritative when two conflict, and name an owner for each category.
Order matters: the question list comes before the material
Work in the right order: first write down 30 to 50 real questions you want the AI to answer, then work out what material answering them requires. Doing it the other way round — piling all the material in first — causes two problems: retrieval is diluted by a mass of irrelevant documents, so answer quality drops; and nobody can tell when the preparation is finished.
Four required categories
1. High-frequency Q&A. Opening hours, store addresses, contact details, booking and cancellation processes, common fault handling. This category carries the highest volume and is the easiest to get right.
2. Policies and rules. Service commitments, after-sales and refund policy, sector compliance requirements. What characterises this material is that it expires, so a review cycle has to be attached to it.
3. Business facts. Product or service catalogue, specifications, availability and price wording. Watch the question of which version is authoritative — two price lists living in the same knowledge base is the single most common source of wrong answers.
4. Escalation rules. Which five scenarios must transfer automatically, who picks them up, and what context travels with them. This maps directly onto the mandatory requirement in clause 5.2.2.6 of the national standard and cannot be omitted.
Three prerequisites that matter more than the material itself
1. Mark the classification levels. Separate out restricted material — customer lists, contracts, cost floors, personal data — and keep it out of the customer-facing answer path by default.
2. Sign off the authoritative wording. Where several documents answer the same question, designate which one governs and withdraw the others. Only the business owner can do this.
3. Name an owner. Assign a maintenance owner per category (pricing, process, technical) so that nothing is left unupdated after launch.
Usability matters more than volume
- Convert scans and images to readable text first, or retrieval will miss them.
- Split long documents into chunks that match the granularity of real questions — one chunk, one point.
- Do not keep superseded versions in the knowledge base "for reference". The AI cannot tell which one is old.
One self-check
When the material is ready, take 30 questions from your list and test against them. If fewer than half can be answered, the problem is not the AI — the material is not ready. This test is more direct than any acceptance checklist.
Key facts
| Four required categories | High-frequency Q&A / policies and rules / business facts / escalation rules |
| Prerequisites | Mark classification levels / sign off authoritative wording / name an owner |
| Material granularity | Chunk by question granularity — one chunk answers one point |
| Self-test rule | Test with 30 questions from the list; under half answered means the material is not ready |
Sources
- Material preparation and loading practice for enterprise knowledge bases
- Three-tier material classification (public / internal / restricted)
- GB/T 47746—2026, clause 5.2.2.6 (transfer rules must be defined in advance)
Follow-up questions
Is more material always better?
No. Irrelevant material dilutes retrieval. Work backwards from the question list to the material you actually need, and keep the first version lean rather than cluttered.
What about scans and old Office files?
Convert them to readable text before loading. This is usually the most time-consuming part of material preparation, and it is independent of which AI tool you use.
Can price material go into a customer-facing AI?
Published list pricing can. Cost floors and customer-specific discounts are restricted material and should not enter the customer-facing answer path.
Who should sign off on the wording?
The business owner. It can be subdivided by category — pricing to commercial, process to operations — but ultimately one person has to be able to make the call.