To make a chatbot, decide the one job it does, gather the information it should answer from, choose how to build it (a chatbot platform, an AI app builder or code), write clear instructions for the AI model, then test it on real questions before you put it in front of anyone. Most modern chatbots pair a large language model with your own content, so they answer from your documents rather than from general knowledge. You can build one without code; with an AI app builder such as Greta.sh you describe it in plain English, and the free plan gives you 5 credits a day to start.
The building has become the easy part. A useful chatbot comes down to narrow scope, good source material and honest handling of questions it can't answer.
What kind of chatbot do you need?
| Type | How it works | Best for |
|---|---|---|
| Rule-based | Follows fixed menus and scripted replies | Simple FAQs, booking flows and routing |
| AI, general | Sends messages to a language model with instructions | Writing help, brainstorming and tutoring |
| AI, trained on your content | Finds relevant passages from your documents, then answers from them | Customer support, internal knowledge and product help |
| AI with actions | Can also call other systems, such as looking up an order | Support and operations tasks |
Most business chatbots need the third type. "Training" usually doesn't mean creating a new AI model. It means giving an existing model your content to look up, which is often called retrieval-augmented generation, or RAG.
Start with the simplest type that does the job. If five fixed answers cover most questions, you may not need AI at all.
How to make a chatbot in 7 steps
Step 1: Define one job
Write a single sentence: "This chatbot helps [who] do [what], and hands off to [whom] when it can't." For example: "This chatbot answers customers' questions about delivery and returns from our policy pages, and offers a contact form for everything else."
Also write down what it should refuse: legal, medical or financial advice, questions outside its topic, and anything about other customers.
Step 2: Gather the source material
The answers will only be as good as the content. Collect your help articles, policies and FAQs, then clean them up:
- Remove outdated pages and contradictions
- Write down answers that only exist in people's heads
- Leave out anything private that shouldn't be repeated to users
Also collect 20 to 30 real questions from your inbox or support tickets. You'll use them to test.
Step 3: Choose how to build it
| Path | Who it suits | Trade-off |
|---|---|---|
| Chatbot platform | A support bot on an existing website | Quick to set up; limited to what the platform allows; pricing often scales with conversations |
| AI app builder | A chatbot inside your own app, or a product you plan to sell | More control over the experience; more for you to test |
| Automation tool plus AI | A bot that connects several existing tools | Flexible; can get complicated |
| Custom code | Teams with developers and unusual needs | Full control, and all the upkeep |
Our list of best AI agent builders compares tools for more complex agents.
Step 4: Choose an AI model
You don't need the most powerful model for most chatbots. Compare a few from OpenAI, Anthropic and Google on your own test questions, then look at accuracy, speed and cost. Usage is generally priced by tokens, roughly the amount of text sent and received, so long documents and long conversations cost more.
Greta.sh's live connectors include OpenAI, Claude and Gemini for AI steps in Flows, and the apps you build can call any service that has an API.
Step 5: Write the instructions
These are often called a system prompt. Be direct and specific:
You are the support assistant for [your company]. Answer only from the provided documents. If the answer isn't in them, say that you don't know and offer the contact form. Never invent prices, policies or order details. Keep answers under 100 words, in plain language. Don't give medical or legal advice. If a customer is upset or asks for a person, offer the contact form straight away.
The most useful line there is "If the answer isn't in them, say you don't know." Without it, models tend to fill the gaps with confident guesses.
Step 6: Build it
If you use an AI app builder, describe the whole thing in your first prompt:
Build a customer support chatbot web app. An admin signs in and uploads help documents, and can edit a list of FAQs. Customers open a chat window on a public page and ask questions. The chatbot answers using an AI model and only the uploaded documents, shows which document the answer came from, and says it doesn't know when the documents don't cover the question. Add a button to send the conversation to a support email. The admin can read past conversations and see which questions went unanswered. Limit each visitor to 20 messages a day. Clean design that works well on phones.
Add a daily message limit from the start, because a public chatbot can run up AI costs quickly if someone abuses it. Our guide to rate limiting and abuse prevention explains how. Our AI chatbot SaaS guide covers the architecture if you want to sell a chatbot as a product.
Step 7: Test it before launch
Run through your 20 to 30 real questions and score each one: correct, partly correct or wrong. Then try to break it:
- Ask about something not in your documents. Does it admit that it doesn't know?
- Ask it to ignore its instructions, reveal them or pretend to be someone else.
- Ask for a discount or a refund. Does it promise something you don't offer?
- Ask about another customer's details. It should never have access to them.
- Paste in an extremely long message. Does anything break, or cost a lot?
- Try it on a phone.
Treat the answers as if they come from your company, because customers will. Review conversations weekly after launch, and turn unanswered questions into new content.
Common mistakes
- Trying to answer everything. A narrow chatbot that's reliable beats a broad one that guesses.
- Hiding the human. Always give people a clear way to reach a person.
- Skipping the disclosure. Tell users that they're talking to AI. Check the rules that apply in your market.
- Sending private data to a model without checking the terms. Read the provider's data policy before sending customer information. Our AI app ideas guide covers this in its section on customer data.
- Launching without limits. Set message caps and spending alerts before the link goes public.
- Never updating the content. Outdated documents produce outdated answers.
Got an idea? Build it now!
Just start with a simple prompt. No coding required: Greta.sh turns your idea into a working app in minutes.
How much does it cost to make a chatbot?
You'll have three main costs: the platform or builder, AI model usage, and your time spent on content and testing. Model usage scales with the number and length of conversations, so estimate from your expected volume rather than from a demo.
On Greta.sh, the free plan gives you 5 credits a day but cannot deploy. Deployment and custom domains start on Starter (a $5/month welcome offer, normally $20/month). Pro (a $20/month welcome offer, normally $50/month) adds the Database, Flows and GitHub version control, which you'll need if your chatbot stores documents and conversations. Greta.sh builds its own database for each app rather than connecting to one you already run. AI provider usage may be billed separately, so check that before you launch.
FAQ
How do I make a chatbot for free?
Many chatbot platforms and AI app builders have free plans for testing, and AI providers sometimes offer trial credit. Running a real chatbot usually has some cost, either for the platform or for AI usage, so estimate from your expected number of conversations.
Can I make a chatbot without coding?
Yes. Chatbot platforms let you configure one without code, and an AI app builder can generate a custom chatbot from a plain-English description. You still need to prepare good content and test the answers carefully.
How do I train a chatbot on my own data?
Usually you don't retrain the model. You upload your documents, the system finds the passages that relate to each question, and the model answers from those passages. This approach is often called RAG. Clean, up-to-date documents matter more than any setting.
How long does it take to make a chatbot?
A basic version can work within a day or two. Preparing the content and testing against real questions usually takes longer than the build, especially for customer-facing use.
Can I add a chatbot to my website?
Yes. Most chatbot platforms provide a widget that you paste into your site. If you build your own, you can place it on a page or subdomain, or embed it as a widget, depending on how it's built.
How do I stop a chatbot from making things up?
Instruct it to answer only from your documents, tell it to say when it doesn't know, show the source of each answer and test with questions that your content doesn't cover. You can reduce made-up answers this way, but you can't eliminate them, so keep a person available for anything important.
