Why assistants make things up
Generative AI can state false information confidently, and NIST treats this as a known risk. A knowledge base cuts down avoidable gaps, but it does not remove the risk. That is why the approach below also uses testing, monitoring and a person who handles exceptions, in line with the NIST AI Risk Management Framework.
Start with one controlled document
A folder of old PDFs and conflicting messages is not a knowledge base. Use one document or database with an owner and a review date. Separate general rules from product-specific exceptions. A simple starting structure:
STORE FACTS
Owner: [person responsible]
Last checked: [date]
Channels and markets: [where each rule applies]
SHIPPING
Processing time by product type:
Shipping services and destinations:
Where the official ship-by date is found:
What we can and cannot promise about delivery:
Who handles a delay:
RETURNS AND PROBLEMS
Published return policy (link):
Custom-item exceptions:
Damaged or missing item procedure:
What requires human approval:
PRODUCTS
Product / SKU:
Materials, size, colors, contents:
Personalization limits:
Care instructions:
Current listing link:
ESCALATION
Questions the assistant must not decide:
Person to contact:
Expected internal response time:Fill it from your live listings, published policies and order system, not from memory.
Give every answer a source and a date
A shipping answer should point to the current processing profile or shop policy. If the source is missing, outdated or contradictory, the assistant should stop and ask a person. That one rule prevents most overpromises.
Keep customer data out of it
Do not put names, addresses or whole conversations into the general knowledge document. Keep only what a defined purpose needs. This follows the data-minimization guidance of the European Data Protection Board and the US Federal Trade Commission. For the legal side, see my customer data checklist.
Test it before customers see it
Ask the questions customers actually ask:
- "Will this arrive by Friday?"
- "Can I return a personalized item?"
- "What if the package is damaged?"
Record every place the draft answer overpromises, then fix the document, not just the answer. NIST recommends testing before deployment and monitoring behavior in use. For the bigger picture, read AI in customer service: save time without losing trust.
Keep reading
Want help building this for your shop?
If you would like help setting this up, reach me through the contact section.
Contact MeSources
Rules, numbers and dates change often. This page reflects the sources below as checked on Oct 9, 2026, so confirm the current version before you act.