An AI agent needs approved answers with a source, owner and update process. Keep live facts in connected systems, resolve conflicting documents, and test what the agent does when information is missing.
Key takeaways
- Give each important answer an approved source and owner.
- Separate stable policy from live availability and status.
- Resolve conflicting versions before launch.
- Test unknown questions and safe escalation.
The problem and the first decision
A business launches a website assistant. The demo looks impressive—until a visitor asks about a weekend booking policy. One document says Saturday appointments are available. Another says they're closed. The assistant chooses an answer and sounds completely certain. The problem may not be the AI model. It may be the information the business gave it. Before building a customer-facing AI agent, you need a maintained source of truth that staff themselves would trust.
What is an AI agent knowledge base?
It is the collection of approved facts, policies, procedures and examples an assistant is allowed to use. A useful knowledge base tells the agent not only what to say, but also when information is outdated, missing, conditional or reserved for human review. It does not have to start as an expensive software platform. A carefully maintained set of documents or structured records may be enough for a limited pilot.
Start with the questions customers actually ask
Review recent email, WhatsApp, website chat and reception enquiries with personal data removed. Group them by intent: hours and locations; products or services; pricing policies; appointments; delivery and returns; account-specific status; exceptions. Pick the 15–25 most useful questions for a first pilot. Don't attempt to write answers to every imaginable question. For each question, identify the staff member or official system that knows the answer.
Give every important answer a source and owner
Use a simple content record: Question: Can customers reschedule an appointment? Approved answer: Yes, under the current booking policy; see the applicable conditions. Source: Current booking-policy document. Owner: Operations manager. Last checked: Actual verification date. Exceptions: Special bookings require reception approval. Escalation: Route to reception when the appointment type is unclear. This prevents an attractive but unsupported response from being treated as fact.
Separate stable facts from live information
Stable information includes address, service categories and general policies, although even these can change. Live information includes current appointment slots, stock availability, order status and assigned staff. A knowledge document cannot reliably confirm a live state simply because it was true yesterday. For a verified appointment, the agent needs an authorised, up-to-date scheduling connection and a successful booking action. If that is unavailable, it should describe the request as pending.
Resolve contradictions before the agent sees them
Two price lists with different values are not harmless. Neither is an outdated PDF circulating alongside a new website policy. Establish which source wins, who approves updates and when conflicting content is withdrawn. Don't assume the AI will make the right editorial decision. That responsibility belongs with the business. For complex conditions, use precise wording. Instead of “Returns allowed,” record the eligibility window, exclusions and which requests need staff approval.
Write answers for action, not just explanation
A good source may include a concise approved answer, relevant conditions, next-step instructions and an escalation reason. Example, clinic booking: Question: Can I visit tomorrow at 3 PM? Approved general answer: Appointment availability must be checked in the booking system. Required action: Verify live slot and permission before confirming. Fallback: Collect preferred time and pass to reception. Forbidden answer: “Your appointment is confirmed” without successful booking.
The agent's behavioural boundaries matter as much as its factual content.
Create a content-update process
Define who edits, who approves and how changes reach the assistant. Retire old versions. Record material updates. When hours or policies change, include the update in tests. A polished first version becomes unreliable if no one maintains it. A small monthly review may be useful for stable content; more dynamic information needs an appropriate update frequency based on how often it changes.
Test with real questions, missing answers and conflicts
Build a test set that includes: An exact FAQ match. A paraphrased question with the same meaning. A request whose answer depends on live data. A question not covered by the approved sources. A policy exception. Two conflicting source documents. A user asking for information the agent must not disclose. A system integration error. Check whether responses are supported, whether uncertainty is stated honestly and whether a suitable handoff occurs.
How NORVOI can help
NORVOI's AI-agent work begins by defining the task, the business knowledge the agent may use and the cases that need human review. We can help turn scattered FAQs and policies into an approved, testable foundation for a limited assistant.
Frequently asked questions
Can I train an AI agent from my website alone?
A website can be a useful source, but it may omit conditions, contain outdated details or lack live data. Review and approve the material first.
Do I need hundreds of FAQs?
No. A focused pilot can start with a small set of high-frequency questions and expand with tested evidence.
Can an agent check today's availability from a PDF?
No. A static document is not a reliable substitute for a current, authorised scheduling system.
Who should own the knowledge base?
A designated business owner or team responsible for the underlying policies and information, with an agreed update and review process.
Good answers begin with trustworthy information
An AI agent's confidence is not proof that its answer is correct. Give it approved knowledge, accurate live-system access where required, and a clear instruction to escalate uncertainty. Choose two FAQs or policy documents that might give different answers to the same customer question. We can help identify the approved source, who should maintain it, and when your AI agent must ask a person instead of guessing.
Have a similar task in your business? Start with the request, the information your team trusts and the point where a person should take over.