
The question
We want an AI chatbot on our website that answers visitors. What can it concretely do, and how do we make sure it doesn’t tell customers wrong things?
An AI chatbot on a website seems simple: a chat window in the corner of the page that answers questions. The difference between a useful one and one that creates problems is not the window, but what information it uses, what it is allowed to say and what it does when it doesn’t know the answer.
Language models phrase things fluently even when they don’t have the correct information. That is why the important question is not whether the chatbot can answer, but how it is limited to the company’s sources and how the situations it cannot resolve reach a person. Below you will find the criteria, the real risks and the steps to get started.
If you want us to handle this for you, see our service Website chatbot or our service Lead capture and qualification.
What a website chatbot can realistically do
A well-configured chatbot takes over repetitive questions whose answer already exists somewhere: on the website’s pages, in the catalogue, in the delivery and returns policy or in the company’s documents. Its value comes from the visitor getting the answer immediately, without searching through menus or waiting for an email.
It does not replace the sales or support team. It works best as a first line: it resolves what is simple and well documented, and passes the rest to a person, with the question already clarified.
- Answers questions about services, products, opening hours, delivery areas or return conditions, based on existing materials.
- Directs the visitor to the right page, instead of repeating all the content in the chat.
- Collects the details of a request (what the customer wants, by when, how they can be contacted), so that the team receives a complete enquiry.
- Shows the team what questions visitors ask and which of them don’t yet have an answer on the website.
Where its answers come from: sources matter more than the model
A language model used without restrictions answers from its general knowledge, which may be out of date or simply wrong for your company. For a company chatbot, the right approach is for the model to first search the company’s approved materials and to phrase the answer only on the basis of what it finds there. The technique is also known as RAG, short for retrieval-augmented generation, meaning generation based on searching sources.
The practical consequence is simple: the chatbot cannot be more accurate than its sources. If the delivery page is outdated or two documents say different things, the chatbot will repeat the confusion. Before configuration, it is worth tidying up the content.
- Choose explicitly which sources it uses: specific pages, catalogues, policies, FAQs.
- Remove or update old documents and contradictory versions.
- Decide who updates the sources when a rule, a schedule or an offer changes.
- Check whether the answers can include a link to the page the information comes from, so that the visitor can verify it themselves.
How to prevent made-up or wrong answers
No AI-based chatbot can guarantee that it will never make a mistake. The risk can, however, be greatly reduced through clear rules and testing. The most important rule: when the information is not in the sources, the chatbot says it doesn’t know and offers a route to a person, instead of improvising.
Equally important are the topics the chatbot should not handle on its own. A wrong answer about a price, a warranty or a contractual condition can create expectations that the company will then have to manage.
- An explicit instruction to answer only from the approved sources and to admit when it doesn’t have the information.
- A list of sensitive topics handed directly to a person: complaints, legal or medical situations, negotiations, exceptions to the rules.
- Cautious wording for information that changes often, with a pointer to the up-to-date page or to confirmation from a colleague.
- No promises on the company’s behalf: the chatbot informs, but does not grant discounts, confirm exceptions or commit to deadlines.
- A set of test questions, including trick questions and unusual phrasings, run before launch and after every major change.
Handover to a person: the part most often overlooked
A chatbot that only says it doesn’t know and ends the conversation leaves the visitor more frustrated than a simple form. Handover to a person must be designed from the start, not added after the first complaints.
Decide what happens during working hours and what happens outside them. If nobody can take over the conversation immediately, the chatbot can ask for contact details and say honestly when the visitor will receive an answer, depending on what the team can ensure.
- The visitor can ask to speak to a person at any time, and the request is not ignored.
- The colleague who takes over sees the whole conversation and does not ask the same questions again.
- Requests arrive somewhere that is actually monitored: a team email, a CRM, a shared inbox.
- There is a person responsible for the requests received through chat.
Personal data and transparency towards visitors
Chat conversations may contain names, phone numbers, addresses or order details. This data must be handled as carefully as that from contact forms: what is kept, for how long, who has access and on which servers it is processed, including by the AI service used.
The visitor must know that they are talking to an automated system, not a person. Besides being fair to the customer, transparency also reduces unrealistic expectations. Legal obligations concerning data protection and the use of artificial intelligence should be checked in the regulations in force, at official sources, and for your company’s situation they are worth discussing with whoever handles data protection.
- A clear message at the start of the conversation that the answers are generated automatically.
- An updated privacy policy covering the data collected through chat.
- The AI service provider’s terms checked: where the data is processed and whether it is used for other purposes.
- A rule on what must not be requested in chat, for example card details or identity documents.
When it is worth it and when an FAQ page is enough
A chatbot makes sense when the website has a lot of content, when visitors ask varied questions that require combining several pieces of information, or when the team constantly receives the same questions by email and phone. It is also useful for finding out what people are looking for and cannot find.
If the questions are few and predictable, a good FAQ page, a clear form and a visible phone number can solve the same problem with less maintenance. A chatbot added on top of a website with incomplete information does not fix the missing information, but makes it more visible.
Getting started and checking after launch
Most problems appear in the first few weeks, when visitors’ real questions differ from those imagined during configuration. Plan time from the start for reading the conversations and making adjustments.
- Gather the real questions received by email, phone and social media and group them by topic.
- Check whether each topic has a correct and up-to-date answer on the website or in documents; fill in the gaps.
- Define the permitted sources, the topics handed to a person and the tone of the answers.
- Test with ordinary questions, trick questions and off-topic questions.
- Start with a limited set of topics and expand once you see how it behaves.
- Read the conversations regularly, correct the sources and keep track of the questions left unanswered.
