How does an AI chatbot differ from a classic chatbot?
A classic chatbot follows fixed decision trees, an AI chatbot understands freely worded questions and answers in full sentences.
| Classic chatbot | AI chatbot | |
|---|---|---|
| How it works | Fixed flows and keywords | A language model understands free questions |
| Knowledge | Predefined answers | Your website, FAQ and documents |
| Languages | Usually just one | Questions in German, French, Italian or English |
| Maintenance | Every answer entered individually | Keep the knowledge base up to date |
| Risk | Feels rigid, does not understand variations | Without guardrails, it can phrase wrong answers |
What is a chatbot useful for in an SME?
- Recurring questions about opening hours, delivery terms, warranty or process
- Product advice in an online store, for example on sizes or compatibility
- Pre-qualifying quote and appointment requests and capturing them completely
- Checking order status, if there is an interface to the store
- Answering internal questions from employees about processes and templates
How does the chatbot know what to answer?
From a knowledge base that you define. The method is called retrieval augmented generation (RAG): for each question, the system looks up the matching passages in your content and writes an answer from them. Guardrails define how it behaves:
- Answer only from the knowledge base, not from the general knowledge of the model
- When unsure, say openly that it does not know and hand over to the team
- Make no binding commitments on prices, dates or legal questions
- Use a tone and form of address that suits your company
Where are the limits of an AI chatbot?
- It can be wrong. Good guardrails and testing reduce the risk but do not eliminate it completely
- It is only as good as the content it can access
- Complex advice and sensitive situations belong with a person
- It needs maintenance: new offers, prices and processes must go into the knowledge base
What does data protection require for a chatbot?
Personal data in the chat is governed by the revised Federal Act on Data Protection (revFADP). In practice, this means:
- State clearly that an AI is answering
- Name the chatbot provider, the country of processing and the purpose in the privacy policy
- Ask only for the data needed for the enquiry, and no sensitive personal data
- Decide how long chat histories are stored
- Sign a contract with the provider that excludes the use of your data for training
- For sensitive content, consider hosting in Switzerland or the EU, for example with the open Swiss model Apertus
How is a chatbot introduced?
| Step | Content |
|---|---|
| Collect questions | Gather the most frequent questions from emails, phone calls and sales |
| Build the knowledge base | Prepare website content, FAQ and documents and fill gaps |
| Set up | Define tone, guardrails, handover to the team and design |
| Test | Check with real questions and deliberately difficult ones |
| Go live and maintain | Review conversations regularly and improve answers |
How do I know whether a chatbot pays off?
A chatbot pays off when your team answers many similar questions or enquiries come in outside office hours. After launch, measure how many conversations are resolved without handover, how many enquiries the chatbot captures and whether routine emails and calls go down.