Knowledge and documents
What the AI knows about your business, where that knowledge comes from, and why this is the single biggest lever on answer quality.
The AI answers from your business, not from the internet
SIMCOAI grounds replies in the information you supply. If something is not in your knowledge, the right behaviour is to say so and capture the question - not to guess.
Most complaints about "the AI got it wrong" trace back to missing knowledge rather than a faulty model. Read the transcript, find the question, add the answer.
What feeds the AI
- Services
- What you actually offer, named the way customers say it. "Consultation" and "follow-up consultation" are different things.
- Opening hours
- Drives in-hours and out-of-hours behaviour everywhere else in the product.
- Policies
- Refunds, cancellations, deposits. State them as you would to a customer, not as internal shorthand.
- FAQs
- The questions you answer most on the phone today. Start here - it is the highest return for the least effort.
- Documents
- Longer reference material the AI can draw on.
- Transfer destinations
- Where a call goes when a person is needed.
Writing knowledge the AI can use
| Do | Instead of |
|---|---|
| "We refund unopened items within 30 days of delivery." | "Refunds: see policy" |
| "We are closed on bank holidays." | Leaving the hours blank and hoping |
| "Ask for the order number, then the delivery postcode." | "Verify the customer" |
| Plain sentences a new starter would understand | Internal codes, abbreviations or team shorthand |
A good test: could someone on their first day answer the phone using only this? If not, the AI cannot either.
Keeping it current
Knowledge rots. Prices change, services get retired, seasonal hours come and go.
- Review after any pricing or service change
- Check hours before every holiday period
- Add anything the transcripts show the AI could not answer
- Remove anything no longer true - stale knowledge is worse than none, because it is answered confidently
See also
- Prompts and tone - how it says things, as opposed to what it knows
- Escalations - topics that should never be answered from knowledge
- Call problems - when the AI loops or cannot answer
Running knowledge as controlled content
For larger teams, give every operational answer an owner, review date and source of truth.
| Control | Recommended practice | Why |
|---|---|---|
| Ownership | Name the team responsible for prices, hours, policy and service facts | Questions reach the person able to approve a correction |
| Precedence | Keep one approved answer for each fact and remove superseded copies | Conflicting sources cannot produce a dependable answer |
| Effective dates | State when seasonal hours, promotions and temporary rules begin and end | Old exceptions do not silently become permanent |
| Regression set | Keep 10-20 representative questions and retest them after material edits | A correction in one area does not weaken another |
| Audit cadence | Review high-risk facts monthly and the full set quarterly | Stale content is found before a customer finds it |
When two approved sources disagree, correct the source material. Do not try to resolve a policy conflict by adding more prompt wording.