Small teams do not need a governance committee to run a useful website assistant. They need a repeatable way to decide what the assistant may say, who can change that decision, and what happens when the evidence is incomplete. That is AI knowledge base governance: a working agreement for customer-facing information.
The risk is rarely that a team has no content. It is that several pages answer the same question differently: an old campaign promises three-day shipping, the current policy says five, and a sales deck describes an exception that was never made public. A chatbot cannot resolve a business disagreement safely. Someone on the team has to decide which source represents the current promise.
Start with a source hierarchy
Before adding material, sort it by authority. Give each important customer topic one preferred source: the page that wins if another source disagrees. This does not require a large inventory. A shared register with a URL, owner, last-reviewed date, and scope is enough to begin.
| Topic | Preferred source | If sources conflict | Owner |
|---|---|---|---|
| Returns | Current returns policy | Pause the older wording and ask the policy owner | Operations lead |
| Product compatibility | Current specification or product page | Require the product owner to confirm the applicable variant | Product lead |
| Prices and promotions | Current public pricing page | Remove expired campaigns; do not infer an exception | Commercial owner |
| Unusual customer requests | No automatic source | Route to a person who can make a case-by-case decision | Support lead |
“Preferred” does not mean every other page is wrong. It means the team has agreed where an answer should come from. Keep drafts, internal notes, retired offers, and unreviewed uploads out of the customer-facing source set until they are ready. A shorter, current source set is easier to govern than a large archive.
Give every high-impact topic an owner and reviewer
One person can hold both roles in a very small business, but the responsibilities should still be explicit. The owner decides what the policy or product information means and keeps the source current. The reviewer checks whether the source is clear enough for a visitor-facing answer and whether it has the right qualifiers: region, date, product variant, eligibility rule, or exception.
Start with information that can create a costly promise: prices, availability, delivery, returns, cancellation, service scope, contact details, and compliance-sensitive statements. Marketing can improve readability, but it should not silently become the final authority on a fulfilment commitment.
A fictional returns policy conflict
Fictional example: a shop’s Help page says “returns within 30 days,” while a discontinued spring promotion says “60 days.” The operations owner confirms that 30 days is the standing policy. The reviewer removes the promotion from the assistant’s sources, checks that the policy states any condition that matters, and records the decision: “Returns policy wins; promotional page retired; reviewed 30 September.” The assistant should not try to blend the two periods into a confident answer.
This small record matters when the question returns later. It explains why a source was excluded and gives the next reviewer a place to start.
Use a checklist when information changes
Any event that changes a customer promise should trigger the same short routine. Put the checklist beside the page editor or source register:
- Update the preferred public source and state the effective date where useful.
- Search for duplicate or older wording in FAQs, campaign pages, downloads, and product pages.
- Confirm whether the assistant’s source set should include, replace, or exclude each affected item.
- Refresh and review the affected knowledge. Confirm that the updated information is available to the assistant.
- Have the owner or reviewer test a few visitor-style questions, including one that should receive a limited answer.
- Record the change, owner, and next review date.
RobiFox is built around this kind of reviewable process: it can read selected website pages and documents, present conflicting information for a decision, and preserve an owner’s correction over later extracted text. Its overview of the workflow also describes reviewing what the assistant learned before relying on it.
Make unanswered questions part of the operating process
An unanswered question is not automatically a defect. Some questions need an order record, discretion, a specialist, or information the business should not publish. In those cases, the correct response is a clear limit and a route to the right person.
Other unanswered questions are evidence. If visitors repeatedly ask whether a service is available in a particular region, the team can decide whether to add a public service-area page, clarify an existing source, or keep the question for human support. RobiFox records questions it could not answer so they can become a practical content backlog; its answer-quality approach also treats correct refusals as a result worth measuring.
For a related look at the review loop, see AI Chatbot Answer Approval Workflow That Works.
Run a short weekly review
Set aside 20 minutes each week. Review changes to high-impact sources, a small sample of answers on those topics, and repeated unanswered questions. Ask four questions: Was the answer supported by the right source? Did it preserve the important conditions? Did it point to a person when a judgment was required? Does the same question keep exposing a gap in the website?
Once a month, check the source register for pages that have changed ownership, expired promotions, products that no longer exist, and policies approaching a review date. When a major policy changes, do not wait for the weekly slot; use the change checklist immediately.
Keep the limits visible
Governance does not make an assistant correct by default. Sources can be ambiguous, pages can be missed, and a customer’s situation can require a human decision. It does give a small team a way to detect those limits and respond deliberately. The useful target is not an assistant that answers every question. It is one that helps visitors find information the business can stand behind, while making uncertainty visible when it cannot.
Cover image: AI-generated editorial illustration, not a product screenshot.
