Re-AI Training Programme — L4.1 AI in customer service: build and deploy your own chatbot

Welcome to lesson 4.1, AI in customer service: build and deploy your own chatbot. Many businesses now use them to help customers day and night. In this lesson you will learn what a customer-service chatbot does and how a modern AI chatbot works. You will see how to build and test one yourself, with no coding. And you will learn your legal duty to tell customers they are talking to AI.

By the end of this lesson you will be able to do four things. First, explain the difference between an older rule-based chatbot and a modern AI chatbot built on a large language model. Second, write a system prompt that sets the bot's role, tone, scope and fallback. Third, build a knowledge base and test your chatbot for wrong answers and a smooth hand-off to a person. Fourth, apply the EU AI Act rule that says users must know they are talking to AI.

Here are some key terms. A chatbot is an AI assistant that answers questions in everyday language, often for customer service. An LLM, or large language model, is the AI behind modern chatbots. It predicts natural-sounding text. A system prompt is the set of instructions that sets a chatbot's role, tone, scope and limits. The knowledge base is the documents it answers from, such as FAQs and policies. Escalation means handing a conversation to a human when the bot cannot or should not answer. Finally, the transparency duty is the EU AI Act rule to tell people clearly when they are dealing with AI.

Customer service never sleeps. A chatbot answers routine questions in plain language, at any hour. That means staff can focus on the hard cases. And this is growing fast. According to Eurostat figures published in 2025, 55 per cent of EU businesses used AI, cloud or analytics in 2023. So chatbot skills are in demand, and still quite rare. They could be your skills.

A customer-service chatbot does three main jobs. First, it answers common questions. Think of opening hours, order status, returns, or how do I do something. The bot answers these in seconds. Second, it works around the clock. A customer can get help at midnight or on a bank holiday, with nobody waiting by the keyboard. Third, it frees people for the hard cases. Staff stop repeating the same answers and focus on complex, sensitive or valuable conversations. Remember, the bot supports staff. It does not replace them. Let the bot handle the routine, so people can handle the tricky questions.

There are two main kinds of chatbot. The older kind is rule-based. It follows fixed buttons and scripts. It is predictable, but limited. If your question is not on its menu, it gets stuck. The modern kind is based on AI, using a large language model. It understands questions written in natural language and answers from the knowledge you give it. It is far more flexible, but it can hallucinate, which means it can give a confident answer that is simply wrong. The best practice is a knowledge-grounded bot. Grounded means it answers only from your own documents. That keeps its replies accurate and on-brand.

A modern AI chatbot works in four simple stages. One, you feed it knowledge. You upload FAQs, policies and product information. This becomes its source of truth. Two, you set its behaviour. You give it a name and a system prompt that defines its tone, scope and limits. Three, it answers customers. It matches each question to your content and replies in seconds, at any time of day. Four, you review and improve: check where it fails, improve the prompt and fill gaps in its knowledge. The quality of the bot comes mostly from the knowledge and the prompt you give it. You will do all four stages in exercise 4.1.

The system prompt is the bot's job description. Write it like a briefing for a new employee on day one. Take a small bakery. First, role and tone: you are the friendly support assistant for a small bakery, be warm, short and polite. Second, scope and limits: only answer questions about our products, opening hours and orders, and for anything else, offer a human. Third, the fallback, for when the bot does not know. If unsure, say so, suggest contacting the team, and never guess. Fourth, AI disclosure. Start each chat with: you're chatting with our AI assistant. A clear prompt separates a good bot from a bad one.

Your chatbot is only as good as the information you give it. So what goes into the knowledge base? Cover four areas. First, frequent questions. These are the real questions customers ask most often, with clear and correct answers. Second, policies, such as returns, refunds, delivery and privacy. These are the rules customers need to understand. Third, products and services: what you offer, the key details, the prices, and what makes each option different. Fourth, contact and escalation: how and when to reach a human, so the bot can hand over smoothly. Tip: write your FAQ the way customers really ask.

You can build a working chatbot in under an hour, with no coding at all. One, create the bot. Sign up to a free no-code tool such as Chatbase and start a new project. Two, upload your knowledge. Add an FAQ document, your policies or a website link, so the bot answers from real content. Three, name it and prompt it. Give it a friendly name and a system prompt: its role, tone, and what it should and should not do. Four, test, then deploy. Ask real customer questions, fix weak spots, then add it to a website with the code the tool gives you. This is exactly what you will do in exercise 4.1.

Test your chatbot like any product, before customers do. First, try everyday questions and edge cases. Run the ten most common questions. Then try awkward, vague or off-topic ones. Second, hunt for wrong answers. A confident but incorrect reply does the most damage, so fix these first. Third, check the hand-off. Make sure the bot offers a human when it should, and never invents information. Here is a tip. Keep a simple test log of the questions it failed. That list becomes your improvement plan.

The same technology can delight customers or annoy them. What makes the difference is behaviour. A good chatbot answers accurately from real company information. It clearly says it is an AI assistant. It admits when it does not know. And it offers a human for complex problems. A bad chatbot invents confident but wrong answers. It pretends to be a human. It traps customers in endless loops. And it never lets you reach a real person. Aim to build the good kind. It builds trust instead of frustration.

Now for something important: your legal duty to disclose the AI. Under Article 50 of the EU AI Act, people must be told clearly when they are talking to an AI system, unless it is already obvious. This rule applies from 2 August 2026. This is the law, not just good manners. Add a visible line such as: you're chatting with our AI assistant. Never let your bot pretend to be a human. And build the habit of disclosing AI into everything you make, starting now.

Time to check your understanding. Pause the video and try to answer these four questions. First, what is the difference between a rule-based chatbot and an AI chatbot? Second, why does grounding a chatbot in your own documents reduce wrong answers? Third, what three things should a good system prompt always define? Fourth, under the EU AI Act, what must you tell customers who chat with your bot? Take your time, then press play when you are ready.

If you want to explore further, there are two useful resources. Chatbase is the no-code chatbot builder you will use in the exercise. And there is Article 50 of the EU AI Act, which sets out the transparency rules. You will find the links on the lesson page.

Let's sum up the key points. Modern AI chatbots answer natural questions using the knowledge you give them. Ground your bot in your own documents to keep answers accurate and on-brand. A clear system prompt, with role, scope and fallback, is what makes a bot reliable. Always test for wrong answers and for a clean hand-off to a human. And by law, under Article 50 of the EU AI Act, you must tell users they are talking to AI.

Now it is your turn to put this into practice. On the platform, start with the activity: build your chatbot's system prompt. Then do exercise 4.1, where you build and test a live AI chatbot. It takes about one and a half hours. Finally, take the lesson quiz. It has five questions, and you need 70 per cent to pass. You can retake it as many times as you like. Good luck!

Co-funded by the European Union.