Most complaints about website chatbots come down to the same thing: the bot said something the company would never say. It invented a price, promised a delivery date, or confidently described a feature that does not exist. The model is rarely the real problem. The problem is what it was allowed to answer from.
If you sell to customers in several countries and want an AI assistant on your site to handle first-line questions around the clock, the knowledge base is where most of the work, and most of the value, sits. This guide shows how to build a first version without waiting months for a perfect documentation project.
Start with one high-frequency scenario
Do not try to cover everything on day one. Look at your inbox, live chat history or sales call notes and pick the scenario that generates the most repetitive questions. For a going-global agency that might be "Which ad platforms do you open accounts on, and what do I need to prepare?" For a SaaS company it might be "Does it integrate with X, and what does the trial include?"
Build the knowledge for that scenario first, launch it as a pilot, and expand based on real conversations.
The six content types you need
1. Service or product scope. What you offer, for whom, in which markets and languages. Write it as plain statements a model can quote: "We support account opening on Facebook, Google and TikTok for registered businesses in the markets listed below" is better than a marketing headline.
2. Who it is not for. Boundaries are as important as capabilities. If you do not serve certain regions, industries or budget levels, say so. This lets the assistant decline politely instead of making something up.
3. Process and preparation. The steps a customer goes through and the documents or information they need at each step. This is usually the most frequently asked category.
4. Pricing logic, not price lists. Unless your prices are fixed and public, give the assistant the factors that drive cost and a rule to hand off pricing questions to a human. Quoting numbers that change by client is a fast way to create disputes.
5. FAQs with verified answers. Collect real questions from sales and support, and have a subject-matter owner approve each answer.
6. Escalation rules. A written list of topics that must go to a human: contracts, payments, complaints, legal or medical advice, anything about a specific account's status, and any request the knowledge base does not cover.
Make every answer traceable
For each knowledge item, record:
- Source (web page, internal doc, policy)
- Owner (who can confirm it is correct)
- Last reviewed date
- Scope (which markets or products it applies to)
When the assistant answers an important question, it should be able to point to its source. When a source is updated, the owner knows which knowledge items to revise. This also makes audits possible if a customer later disputes what the bot said.
Handle languages deliberately
Global teams often ask whether to maintain the knowledge base in every language. A practical approach:
- Maintain one master version in the language your subject-matter experts work in.
- Let the assistant answer in the visitor's language, but test key answers in each priority language with a native speaker.
- Keep market-specific facts (local regulations, availability, payment methods) as separate items tagged by market, so they are not applied globally by mistake.
Test before you launch
Write 30–50 test questions covering:
- Normal questions the knowledge base should answer
- Edge cases with partial information
- Questions it should refuse or escalate
- Attempts to get it to promise outcomes or invent facts
Run the tests, review answers with the knowledge owners, fix gaps, and re-run. Keep the test set; re-run it whenever you change the knowledge base or the underlying model.
Maintain it like a product
After launch, review a sample of real conversations weekly. Look for:
- Questions the assistant could not answer (new knowledge needed)
- Answers that were technically correct but confused visitors (rewrite needed)
- Escalations that could have been handled automatically, or vice versa
Feed findings back into the knowledge base. Over a few cycles, the assistant becomes noticeably more useful, not because the model changed, but because the knowledge did.
A note on expectations
A well-built assistant can reduce repetitive back-and-forth and give visitors in other time zones a consistent first response. It cannot guarantee lead volume or replace human judgment on complex deals, and it will still make occasional mistakes, which is why escalation rules and human review stay in place.
Further reading
PandaWM's AiChatSale service page outlines a full workflow from knowledge preparation to sales handoff, and its article on overseas marketing automation with an AI sales assistant looks at where chat fits into the wider funnel. More services are listed on the PandaWM homepage.
Top comments (0)