What Should You Put in an AI Assistant’s Business Knowledge Base?

An AI assistant’s business knowledge base should contain approved answers about your services, coverage, pricing rules, process, policies and contact routes. Each source needs a clear owner and review date. Keep confidential records out of a customer-facing assistant’s sources, and document when it must ask a person for help.

Start with what you want the assistant to answer. A website assistant handling new enquiries needs different information from an internal assistant helping staff deliver projects.

What is an AI knowledge base?

An AI knowledge base is a collection of information an assistant can consult when answering questions. It might include website pages, help articles, documents or structured records, depending on the system.

In a retrieval-based system, relevant material is found and supplied to the model when a question is asked. This is different from retraining the model. Microsoft’s retrieval-augmented generation documentation explains this approach and warns that inaccurate answers remain possible even when sources are supplied.

If you are still choosing the type of system, start with our AI agent, chatbot and automation comparison.

What business information belongs in it?

For a customer-facing service-business assistant, use this inventory as a starting point. Include only facts the business has approved for that audience.

A contractor’s assistant might need confirmed service areas, job exclusions and the steps before a site visit. A property agent’s assistant might need approved seller and buyer enquiry routes, current listing information and viewing-request rules.

Do not treat these examples as permission to invent fees, credentials or services. Start from the business’s existing documents and ask the relevant owner to resolve gaps.

Separate facts, instructions and actions

A useful source explains what is true. Behaviour instructions explain how the assistant should respond. An integration determines what the assistant can actually do.

For example, an appointment policy can explain booking conditions. It does not let an assistant check a live calendar or reserve a slot. Those actions require a connected system, appropriate permissions and a verified result.

Zendesk’s AI-agent setup guidance separates connected knowledge sources from procedures and authorised actions. Check those distinctions in whichever product you choose.

Configure the assistant to acknowledge missing information and route the question to a person. A sentence in a document is not proof that the deployed assistant follows that rule.

Make each answer usable on its own

Give each topic a descriptive heading and state the answer before the explanation. Include the service, location or customer group needed to interpret it. Replace ambiguous phrases such as “the usual fee” with an approved figure and its conditions, or explain how the fee is established.

Intercom’s content guidance for Fin recommends clear, self-contained sections, explicit limits and review by the relevant content owner. Those principles are useful when preparing sources; they do not guarantee any assistant’s behaviour.

Avoid loading several conflicting versions of a policy. Keep the current approved answer active and archive superseded material outside the assistant’s searchable sources.

Keep a source register

For each active source, record its title or link, intended audience, owner, approval date, last review and next review trigger. Note how changes reach the assistant and how you will check that they are available.

Use event-based reviews for changing information: revise the source when services, prices, hours, policies or listings change. A static document is a poor substitute for live availability. Where current data cannot be checked, the assistant should refer the enquiry to the responsible person.

Customer-facing and internal sources need enforceable access boundaries. Do not assume a file labelled “internal” is protected from customers once an assistant can retrieve it.

A practical example: a contractor’s quote-request answer

This is a hypothetical source entry, not a live business policy.

Question: Can I get a renovation quote through this chat?
Approved answer: You can send a quote request with your project location and a short description. The contractor must review the work before confirming a price. Sending the request does not book a site visit.
Escalation: If the work or service area is unclear, send the enquiry to the person handling estimates. Do not guess whether the project is accepted.

The business would complete this entry with its actual contact route, response expectation and source owner before deployment.

Notice what the answer preserves: the customer can take a useful step, while the assistant avoids promising a price or appointment. This supports the enquiry-to-assessment journey without replacing the contractor’s judgement.

Test answers before customers rely on them

Use the assistant’s actual deployed configuration to check ordinary questions and difficult cases:

• Ask about an accepted service and an excluded job.
• Request an exact price when no approved price exists.
• Ask for a date that cannot be confirmed from current data.
• Use a different phrase for a familiar customer question.
• Ask about a recently changed policy.
• Request information that belongs to another customer.
• Ask an unanswered question and check the human handoff.

Record the expected answer, actual response, source used where visible and correction needed. Re-test after source updates. This is a proposed test checklist; no particular assistant has been tested for this article.

Frequently asked questions

Can I just connect my website?

It can be a starting source if the product supports it. Check whether the website is current, complete and permitted for that audience. Marketing copy may omit exclusions, response rules and escalation details.

Should I upload old emails and customer records?

Do not upload them wholesale to a customer-facing knowledge base. Extract approved, anonymised answers where useful. Customer-specific work needs a separately controlled process and access to the correct record.

Will better documents stop every wrong answer?

No. Source quality, retrieval, instructions and system behaviour all matter. Keep human review for exceptions and commitments the assistant cannot verify.

Start with approved answers and clear ownership

Build a small, current collection around the questions the assistant must handle. Assign an owner, define its limits and test the actual responses before expanding its role.

Contact Funnelphilia with the enquiry task you want to improve and the information your customers need. We can help clarify the content and connect it to the page, enquiry route and human follow-up.