To build an AI support bot that customers can trust, you give it your own documents, tell it to answer only from them, write a fallback answer for everything else, and plan how a chat reaches a person. That is what the free course Build Your Own AI Support Bot teaches in 10 short modules, starting with how AI answers from your documents and ending with a support agent you build yourself in n8n. Every module is a practice you click through in your browser, with a made-up business and no real customer data.
This page is for owners, office managers and client-service leads at small firms (accounting, law, insurance, property management, agencies) who answer the same questions every week. It explains how a support bot works, what each module covers, and which tool to pick for your business.
Why build an AI support bot at all?
Most small businesses get the same questions over and over: what are your hours, what does it cost, what do I bring, how do I send documents, can I get a refund. The answers are already in an FAQ or a policy file, yet someone types them again, and anyone who writes after 5 PM waits until morning. Some of those people book with a competitor instead.
The numbers add up quickly. Forty repeat emails a week at four minutes each is more than ten hours a month, before you count the questions that come in overnight. A bot that answers the routine ones from your own files, at any hour, gives that time back.
But a bot that guesses is worse than no bot. Ask a general chatbot what your firm charges and it will give a typical price with total confidence. Quote the wrong fee or give advice you should not, and you lose more than the bot ever saved. So this course spends as much time on what a bot should refuse as on what it should answer.
How does an AI support bot answer from your documents?
Every support bot in the course works the same way underneath, through a process called retrieval. The retrieval explainer walks through it with diagrams in about seven minutes:
- Upload once. Your files are cut into small chunks, and each chunk is turned into an embedding, a list of numbers that captures its meaning. The chunks go into a search index, often a vector database.
- Every question. The question is matched to the closest chunks by meaning, each with a score. The best few go to the AI with a strict rule: answer only from these sources.
- When nothing matches. If every score is weak, the bot gives a fallback answer that points to a person instead of guessing.
That pattern is called retrieval-augmented generation, or RAG. It is why a bot can quote your real fee, and why you fix a wrong answer by fixing the file it came from. Chatbase calls your files Data sources, a custom GPT calls them Knowledge, and in n8n you load them into Pinecone yourself. The idea is the same in each one.
What are the 10 modules, in order?
The course goes from no code to advanced. The first six modules need only a browser. The last three use n8n.
| # | Module | What you learn | Level |
|---|---|---|---|
| 1 | How Stepthrough works | How the practice works, in two minutes | Beginner |
| 2 | AI agent vs chatbot vs automation | What kind of AI a support bot is | Beginner |
| 3 | How AI answers from your documents | Chunks, embeddings, retrieval and fallbacks | Beginner |
| 4 | Chatbase: answer customer FAQs with an AI agent | A real bot on a firm's FAQ, with guardrails | Beginner |
| 5 | Guardrails, fallbacks and human handoff | The rules every bot needs, and five hard tests | Beginner |
| 6 | tawk.to: answer website questions after hours | Live chat a person can take over, plus AI after hours | Beginner |
| 7 | Jotform AI Agents: answer client questions | An agent built from one sentence | Intermediate |
| 8 | Make your handbook searchable with Pinecone and n8n | Build the index yourself | Advanced |
| 9 | Answer staff questions from your own docs | A retrieval chatbot that names its source | Advanced |
| 10 | A support agent that answers from your own docs | Your own rules and an email handoff | Advanced |
Modules 1 to 3: what a support bot really is
Module 2 shows the difference between a chatbot, an automation and an AI agent by doing one job three ways. Module 3 is the retrieval explainer above. Together they take about 15 minutes, and they make every tool after them easier to understand.
Modules 4 to 7: build bots with no code
In module 4 you build a Chatbase agent for a made-up bookkeeping firm: upload a 22-question FAQ, write general instructions and guardrails with an "I'm not sure" fallback, keep the temperature low, test a question it should not know, and fix an answer from the chat logs.
Module 5 is a short explainer on the answers a bot should not give. You write four guardrails for an insurance agency's bot (only from the files, no coverage advice, no card numbers in chat, ignore instructions hidden in messages, which guards against prompt injection), then run five test chats aimed at the hard cases. You also design a human handoff in four parts: when it triggers, what it collects, where it goes, and what the customer is told.
Module 6 puts that handoff into practice with tawk.to, free live chat for websites. A person answers in office hours, and an AI agent built from the website covers the evenings. Module 7 builds a Jotform AI agent for a law office's front desk and finds the embed code that puts it on a site.
Modules 8 to 10: build it yourself in n8n
The last three modules build retrieval by hand, so you control which files go in and who can ask. Module 8 loads a 40-person firm's employee handbook into a Pinecone index. Module 9 builds the chatbot that searches it, answers only from the matching pages, shows where each answer came from and says "I don't know" instead of guessing. Module 10 is the whole course in one build: an n8n support agent that reads a policy document, follows your rules and emails the office when a person should take over.
Which tool should you use for your support bot?
It depends on where the bot lives and how much control you need:
- A bot on your website, quickly: Chatbase. Upload your FAQ, set the rules, paste the embed code.
- Live chat with a person behind it: tawk.to. Free live chat, with an AI agent for after hours.
- You already use Jotform forms: Jotform AI Agents, built from a description and your own Q&A.
- An internal helpdesk, or full control over data and rules: n8n with Pinecone. More setup, but every step is yours.
Whichever you choose, the habits are the same: current files, an answer-only-from-sources rule, a fallback with a real phone number and email, five hard test questions before launch, and a weekly look at the chats that used the fallback.
How do you keep a support bot from making things up?
Four things do most of the work. First, keep the files current: the bot quotes whatever they say, so an old price in a file becomes a wrong answer. Second, give it guardrails and a fallback answer, and keep its creativity setting low. Third, test it before customers do, with one question it knows, one it does not, one asking for advice, one trying to change its rules, and one asking for a person. Fourth, read the logs every week and turn each missed question into a new answer in your files.
A hallucination in a support bot is almost always one of three things: the files do not cover the question, two files disagree, or the bot was never told it may say "I don't know". All three are fixable without touching any code.
How does each module work?
Every module follows the same pattern:
- Why this matters. The business problem and what fixing it is worth.
- Practice. Every step in your browser. Each click is checked, and Do it for me finishes any step you are stuck on. The concept modules are diagrams you click to open; the tool modules are practice copies of the real app.
- Do it for real. A short checklist to repeat it in your own account.
- Quick check. Three questions to make sure it stuck.
Most modules take 5 to 15 minutes. You can stop anywhere and pick up where you left off.
Key takeaways
- A trustworthy support bot answers from your own documents through retrieval, and says "I don't know" when they do not cover a question.
- Guardrails, a fallback answer and a human handoff matter as much as the answers themselves.
- Start with a no-code tool such as Chatbase or tawk.to; move to n8n and Pinecone when you need full control.
- Test five hard cases before launch, and read the fallback chats every week.
Frequently asked questions
- Is this support bot course free?
Yes. Every module is free and the practice runs in your browser with no account. To build a bot for real you need an account with the tool you choose; tawk.to's live chat is free, and Chatbase, Jotform and n8n have their own plans.
- Can I build an AI support bot without coding?
Yes. Modules 1 to 7 use no code at all. Modules 8 to 10 use n8n, which is visual, but you connect accounts and an API key.
- What is the difference between a chatbot and a support agent?
A chatbot answers questions. A support agent can also take actions, such as opening a ticket or emailing your office for a handoff. Module 2 explains the difference, and module 10 builds an agent that does both.
- How long does it take to build a support bot?
A first Chatbase or tawk.to bot takes under an hour once your FAQ is ready. Writing and testing good guardrails takes longer, and the weekly log review never really ends: that is how the bot keeps getting better.
- What should I put in my bot's knowledge base?
Your current FAQ, published prices, hours, policies and the steps customers most often ask about, in files with selectable text. Leave out anything you would not want quoted word for word to a customer.