If a bot is to answer questions about returns, deliveries, payments or order status, it needs a structured knowledge base, not just a chat interface. According to Atlassian, an AI agent is only as effective as the knowledge it relies on, and most knowledge bases have gaps.
In practice, this means that a chatbot without current articles, instructions and FAQs will either guess or refer the customer to a human. According to AskSpot, the knowledge base is the agent's “library of answers”, fed by its own questions and answers as well as content from the company's website, which allows the bot to respond consistently and in the customer's language.
According to eDesk, the first step is to analyse the most common types of requests from the last 30 days and then group them by topic, such as order status, returns, delivery, products and payments. Only then is it worth writing one article per frequent problem, in plain language and without mixing several topics in a single text.
Pegamento describes a similar process: collect the most frequent questions, assign an owner to each piece of content, choose a platform that is easy to search, and test the materials with real users. This matters because the knowledge base should be more than a collection of documents; it should also be a source of answers available during the service process.
The most common problem is not a lack of an AI model but an outdated knowledge base. If prices, the returns policy or service availability have changed, the bot may still answer based on an old version of the information and mislead the customer. HubSpot points to the same issue, writing that a knowledge base not only helps customers but also “trains” the company's own AI to give fuller answers.
The second risk is structural chaos. According to Atlassian and Pegamento, the base must be easy to search and grounded in real cases, because a document repository alone does not solve the problem of answering during a conversation. The bot needs content that can be quickly retrieved and interpreted in the context of a request.
According to eDesk, it is worth measuring once a month which articles actually resolve customer cases and which need improvement. Such a review also makes sense after every change to the offer or policy, because the knowledge base works as an operational layer for the chatbot, not as an archive.
HubSpot additionally stresses the role of the team responsible for content, because AI draws on what has been previously described and organised. If a company wants its chatbot to respond in line with service rules, it must first maintain order in the very knowledge the bot reads.
Lub System helps B2B companies implement AI, automation and IT solutions end-to-end - from strategy to deployment. See our services or get in touch to discuss your case.
Source: https://atlassian.com/pl/collections/service/guides/ai-customer-service-management