AI agent vs chatbot: what's the difference?
Last updated:
Three kinds of chatbot
People use the word chatbot for very different tools. The clearest way to tell them apart is to ask two questions: how does it produce an answer, and can it do anything besides talk?
- Rule-based chatbot: follows a decision tree or matches keywords. Every reply and every path is written in advance by a person.
- LLM chatbot: uses a large language model to write each reply in natural language. Good ones are grounded in your own content, so they answer from your pages and documents rather than from general knowledge.
- AI agent: an LLM chatbot that can also use tools. It decides when to call an API, look up live data or run a step, and then uses the result in its reply.
How a rule-based chatbot works
A rule-based chatbot is a flowchart with a chat window on top. You define the questions it asks, the buttons visitors can press and the reply at the end of each branch. Some versions also scan typed text for keywords, so a message containing “refund” is routed to the returns branch.
Because a person wrote every reply, nothing unexpected comes out. That predictability is the main reason to use one. It suits processes with a fixed order, such as qualifying a lead with three questions or walking someone through a form.
The cost is coverage and maintenance. The bot only knows the paths you built. A visitor who phrases a question differently, or asks two things at once, often hits a dead end or loops back to the main menu. Every new topic means another branch to design, write and test.
How an LLM chatbot works
An LLM chatbot sends the conversation to a large language model, which writes a reply. It copes with paraphrases, typos, follow-up questions and different languages without anyone designing a branch for each case.
On its own, a language model answers from what it learned during training. That knowledge is general, may be out of date and contains nothing about your prices, opening hours or return policy. When it lacks an answer, it can produce a confident guess instead. This is called hallucination.
The fix is grounding. Before the model replies, the system searches your own content, such as web pages, PDFs and Q&A pairs, and passes the most relevant passages to the model with instructions to answer from them. This pattern is called retrieval-augmented generation, or RAG. A well-grounded chatbot answers from your sources and says so when the answer is not there.
How an AI agent works
An AI agent adds tools. You describe a set of actions to the model, for example “search the product catalog” or “look up an order by number”, including the inputs each one needs. During a conversation the model decides whether an action would help, fills in the inputs from what the visitor said and asks the system to run it. The system calls the API, hands back the result, and the model writes its reply with that information. The mechanism is called tool calling or function calling.
This is what turns a chatbot from something that explains into something that can check and do. An agent can tell a visitor whether a product is in stock right now, fetch the price of a specific configuration or trigger a step on your own web page.
Rule-based chatbot vs LLM chatbot vs AI agent
The table summarizes typical behavior, not any single product. The lines between the types are not sharp: many products mix approaches, and an AI agent is really an LLM chatbot with extra abilities. Still, the three labels describe real differences in setup, behavior and risk.
| Type | How it answers | Can take actions | Setup effort | Best for | Main risk |
|---|---|---|---|---|---|
| Rule-based chatbot | Fixed replies from a decision tree or keyword rules | Only steps built into the flow | High: every path is designed by hand | Short, fixed processes and forms | Dead ends when visitors go off script |
| LLM chatbot (grounded) | Writes answers from your content, retrieved when the question is asked | No | Low to medium: add content, write instructions | Questions about products, services and policies | Wrong or invented answers if content is missing or grounding is weak |
| AI agent | Writes answers and calls tools for live data or steps | Yes, through the actions you define | Medium: content plus API setup and testing | Questions that need live data or a step in your systems | Calling the wrong action or acting on bad input |
When each one fits
Start from the conversations you already have. Collect the questions your team answers by email and phone, then sort them by what a good answer needs.
- A fixed sequence with a known end, such as a three-question eligibility check: a rule-based flow is enough.
- Questions whose answers are already written down, such as shipping costs, opening hours, what a service includes or how to reset a password: a grounded LLM chatbot.
- Questions that need live data, such as current stock or the price of a configured product: an AI agent with a lookup action.
- Requests that change something in your systems, such as creating a booking or cancelling an order: an agent can do this through your own API, but only if you build and secure that endpoint. Many businesses link to their existing booking or account page instead.
- Conversations that need judgment, empathy or authority, such as complaints or legal and medical questions: keep a person in charge and let the bot collect an email address or phone number so someone can follow up.
The risks of AI agents and how to reduce them
The more a bot can do, the more can go wrong. These are the risks that matter most on a public website, with the usual ways to contain them.
- Hallucination: the bot states something that is not in your content. Ground answers in your sources, add Q&A pairs where exact wording matters, and instruct the bot to say when it does not know.
- Wrong actions: the model calls an action at the wrong moment or with the wrong input. Offer only the actions you need, describe each one precisely, and keep actions read-only wherever you can.
- Prompt injection: a visitor, or text hidden in a page, tries to override the bot's instructions. Use rules that resist such attempts, never give the model secrets, and let your API check permissions itself instead of trusting the chat.
- Data exposure: an API returns more than the visitor should see. Choose which response fields the model receives, so internal data never reaches the conversation.
- Untested behavior: an action works in the demo and fails on real input. Test each action with realistic, unusual and hostile messages before going live, and read real transcripts regularly afterwards.
A checklist for choosing
Before you compare tools, write down answers to these questions. They usually make the choice obvious.
- Which questions come up most often, and where are their answers written today?
- Which answers depend on data that changes daily or hourly?
- Is there an API for that data, and can it be called safely from a chat?
- What should the bot never do or say, and how will you enforce that?
- Which channels do you need: only your website, or messaging apps and phone as well?
- Does a person need to take over live chats, or is collecting an email address or phone number enough?
Where intoCHAT fits
intoCHAT is an LLM chatbot grounded in your own content, with actions on top. You train it on your website (it crawls the site and you pick the pages), on uploaded files such as PDFs and Word documents, on text snippets and on Q&A pairs. Each answer is based on hybrid retrieval, semantic plus keyword search, over that content. Built-in rules keep answers grounded in your sources, resist prompt injection and make the agent say when it does not know instead of inventing an answer.
For anything your content cannot answer, you can add actions. Custom API actions let the agent call your own HTTP API during a chat, with inputs filled in from the conversation and a choice of which response fields the model sees, up to 25 per agent. Client-side actions run JavaScript functions on your own page. And if you run a shop on Shopify, WooCommerce, Magento, VTEX, WordPress or a custom site, intoCHAT can create live product-search actions from the shop's public data.
It is just as clear what intoCHAT is not. It is not a visual flow builder for rule-based decision trees. It has no human handoff or live chat. It runs on your website only, not in messaging apps. And it uses one AI model chosen and managed by intoCHAT, so you cannot pick the model yourself. If you need those things, a different tool will suit you better.
To see how a grounded chatbot answers from your own pages, you can try the free preview with your website before creating an account.