GuidesAI chatbot
AI chatbot for your website: getting a customer service assistant right
Yesterday’s chatbot and today’s
Old chatbots followed a decision tree: buttons, keywords, and a dead end as soon as the question left the script. An AI chatbot understands a freely worded question and writes its answer.
That flexibility has a price: without a frame, it also answers off the mark. The whole job is to give it the right sources and the right limits.
Defining the scope
A chatbot that can do everything does not exist. You choose:
- Topics coveredDelivery, returns, how the product works, pricing.
- Topics excludedLegal or medical advice, anything binding.
- Actions allowedLooking up an order, opening a ticket. Nothing irreversible without confirmation.
Sources: the heart of the matter
The chatbot does not answer from memory. For each question, it fetches the useful passages from your content, then writes from them.
- Documentation, FAQ, terms of sale, product pages.
- Up-to-date content: an outdated FAQ gives wrong answers with confidence.
- A source cited in every answer, so the visitor can check.
That is what the Votons assistant does: it answers questions about the manifestos by relying on the exact text. The technique is explained in our guide to RAG.
Handing over to a human
This is what separates a good assistant from a wall.
- The chatbot offers a human as soon as it finds no source.
- It also offers one when the customer insists or gets angry.
- It passes on the whole conversation, so the customer repeats nothing.
- Outside opening hours, it takes the message and states a realistic delay.
Tone and rules
- It presents itself as an AI from the first message.
- It answers briefly, in the visitor’s language.
- It promises nothing the company will not keep: a delay, a goodwill gesture.
- It politely declines what falls outside its scope.
Personal data
- Ask only for what is needed to answer.
- Say what happens to conversations, and how long they are kept.
- Do not send sensitive data to the model without need.
Measuring
| Indicator | What it tells you |
|---|---|
| Share of conversations resolved without a human | Real usefulness |
| Share of “I don’t know” answers | The gaps in your content |
| Rating given by visitors | Perceived quality |
| Most frequent questions | What to fix on the site |
The last indicator is often the most valuable: a chatbot reveals what your site does not explain.
Before going live
Have it tested by people who do not know the product, with real questions. Keep a set of a hundred questions and replay it at every change. To go beyond conversation, see AI agents.
Frequently asked questions
- Can an AI chatbot replace customer service?
- It handles repetitive questions, which are often the bulk of the volume. Disputes, refunds and unhappy customers must reach a human, with the conversation history.
- How do you stop the chatbot from inventing an answer?
- By forcing it to answer only from your content, to cite the page used, and to say it does not know when nothing matches. That is the principle of RAG.
- Do you have to tell visitors they are talking to an AI?
- Yes. It is good practice and a transparency obligation under the European AI regulation. A chatbot must present itself as one.
