Short answer: a well-planned AI chatbot answers with verified information from your company, recognises what it does not know and hands the conversation to a person with all the context. What matters is not the model but the data, the rules and the supervision.
What it can (and cannot) solve
A conversational assistant is useful for frequently asked questions, order status, opening hours, requirements or first-stage qualification. It should not improvise prices, deadlines or legal commitments, nor resolve sensitive cases without a person.
How a reliable chatbot is built
- Define the scope. Which topics it handles and which it always hands over.
- Prepare the knowledge base. Up-to-date documents with a single version of the truth.
- Tie the answer to the sources (a technique called RAG) so it cites what the documentation says instead of making things up.
- Set the tone and limits: language, style, what it must not say.
- Design the handover to a person with a summary of the conversation.
- Test with real cases before launch, including hard questions.
- Monitor and improve using conversations and user ratings.
Basic legal requirements
- Transparency: tell people they are talking to an AI.
- Data protection: legal basis, information to the user, minimisation and a retention period for conversations.
- Suppliers: a data processing agreement and a review of where data is processed.
- Cookies and analytics: if the widget loads third-party resources, apply your consent policy.
This list is indicative. Consult your legal adviser or data protection officer before publishing.
Metrics that matter
| Metric | What it tells you |
|---|---|
| Resolution without human intervention | How many enquiries it closes well on its own. |
| Handovers to people | Which topics it does not master or are sensitive. |
| User ratings | Whether the experience is good. |
| Incorrect answers detected | Quality of the knowledge base and the rules. |
| Time to first response | Service improvement against the previous state. |
Common mistakes
- Launching without an organised knowledge base.
- Not allowing people to speak to a human.
- Promising that it “answers everything”.
- Not reviewing conversations after launch.
Frequently asked questions
Does the chatbot make up answers?
Models can do so if they are not anchored to sources. That is why it is designed with your own documentation, “I don’t know” rules and conversation reviews.
Does it work in several languages?
Yes, but each language should be tested with real cases and the sector terminology reviewed.
Can it integrate with WhatsApp or the CRM?
It is possible, subject to each platform’s terms. Each integration adds data and consent requirements.
Find out how we approach it in the applied AI agency or write to us.
