A chatbot is a system, not just a model
The simplest useful mental model is a conversation-shaped application.
You see a message box and a stream of replies. Behind that interface, several parts work together:
Your message
↓
Chat application loads instructions and conversation state
↓
AI model interprets the request or drafts a response
↓
Application may retrieve data or use another service
↓
Rules and safety checks run
↓
Reply appears and selected state is savedThe LLM used by many modern chatbots is one component in this path. It generates or interprets language. The chatbot is the larger product that decides what the model sees, connects it to information, displays its output, and handles what happens when the model cannot safely complete a request.
Not every AI chatbot uses an LLM. A task-focused chatbot can use machine learning to classify what a person wants and extract details from the message. An LLM-based chatbot can generate a new response from instructions and context. Many production chatbots combine generated language with fixed rules for tasks that need predictable behavior.
How an AI chatbot works
Each message starts a new turn. A typical turn has six stages.
1. Receive the message
The channel might be a website, mobile app, messaging service, or voice interface. If the person speaks, speech-recognition software first turns the audio into text.
2. Load the conversation state
The application loads the information it has chosen to keep. That may include recent messages, a short summary of older messages, the user's account status, or progress through a task.
This is how a chatbot appears to remember. The model does not automatically retain a private memory of every earlier call. The application or model service must store relevant state and supply it again on a later turn.
3. Interpret the request
An intent-based chatbot maps the message to a known goal, such as track_order, and extracts needed details, such as an order number. If a required detail is missing, it follows a dialog rule and asks for it.
An LLM-based chatbot receives a bundle of input that can include product instructions, the user's message, selected conversation history, and relevant records. It then generates the next response. The application, not the user-facing chat box, assembles that bundle.
4. Get information or perform a bounded action
A chatbot may search an approved knowledge base, look up an order, create a support ticket, or call another service. These connections are separate capabilities. The language model does not gain access to a database merely because the chatbot can discuss its contents.
The surrounding software should check who the user is and what they are allowed to do before exposing data or changing anything.
5. Check and return the response
The application can reject unsafe requests, remove unsupported content, require a fixed confirmation, or route the conversation to a person. Some replies are written by a model. Others are templates filled with verified data. A single conversation can use both.
6. Save what the next turn needs
The chatbot saves selected messages, task progress, or a summary for later. It may also log quality signals for review. Saving a conversation for possible future improvement is not the same as retraining the model during the conversation.
Worked example: changing a delivery address
Suppose you write:
You: Where is order 4821?
Chatbot: It has been packed but has not shipped.
The chatbot may have authenticated your account, extracted 4821, retrieved the order record, and turned the verified status into a short reply.
You then write:
You: Can I change its delivery address?
“Its” only makes sense because of the previous turn. The chatbot application supplies the relevant state: the current order is 4821. It also retrieves the rule that an address can change only before shipment.
The system now has the order and its status, but it does not have the new address. A well-designed chatbot does not guess or claim success. It asks for the missing information:
Chatbot: Yes. Order 4821 has not shipped, so its address can still be changed. What address should I use?
After you provide the address, the application can validate it and show a final confirmation before submitting the update. If the order had already shipped, it could explain the limit or hand the conversation to support.
This exchange feels like one continuous conversation. In the software, it is a sequence of separate turns joined by stored state, data lookups, business rules, and explicit failure paths.
Why AI chatbots matter
A chatbot turns a rigid interface into a flexible language interface. Instead of finding the right menu, form, or search phrase, you can describe what you need in your own words. Follow-up messages let you refine the request without starting over.
That flexibility is useful for support, information retrieval, intake, and guided tasks. It also creates risk. Generated language can sound certain when it is wrong. The chatbot may misunderstand what a pronoun refers to, retrieve the wrong record, or try to handle a request outside its permitted scope.
The quality of a chatbot therefore depends on more than the model. It depends on the information it can access, how state is managed, which actions are allowed, how uncertainty is handled, and whether there is a clear route to a person.
Common misconceptions
“The chatbot is the LLM”
An LLM can produce text from input. A chatbot adds the conversation interface, instructions, state, data connections, permissions, checks, and product behavior. Changing the model can alter the chatbot's performance without replacing the whole chatbot.
“It remembers everything I said”
A chatbot remembers only what its design and data policy preserve and reintroduce. It may receive the full history, a shortened summary, selected facts, or no earlier context at all. A new session may behave differently from a continuing one.
“It learns from every conversation”
Usually, it does not update its model during your chat. Operators may review logs, collect ratings, change instructions, update source material, or train a later model. Those are separate processes.
“A natural answer is a verified answer”
Fluency is a property of the wording, not proof of the claim. A chatbot can generate a plausible statement without checking a source. For high-stakes or account-specific answers, useful evidence includes citations, verified records, confirmations, and clear limits.
“Every AI chatbot is an AI agent”
The labels overlap, and vendors use them differently. A chatbot is defined mainly by conversation. An agent is usually expected to choose and carry out multiple steps toward a goal with more independence. A chatbot can perform a bounded action without becoming an autonomous agent.
How a chatbot fits into a wider AI system
The model is the language engine. The chatbot is the conversational product around it. A harness coordinates instructions, state, data, model calls, and permitted actions. An agent may use the same parts while taking more initiative over a multi-step task.
These are layers, not mutually exclusive product labels. One product can present a chatbot interface, use an LLM as its model, rely on a harness to coordinate each turn, and expose a few agent-like actions. The useful question is not what the marketing page calls it. Ask where state lives, what data the system can access, what actions it can take, and which decisions require confirmation.
Where to go next
Read What Is an LLM? for the language model used inside many current chatbots. Read What Is an AI Harness? for the software layer that manages context, tools, and controls around that model. Then use AI Model vs. Chatbot vs. Harness vs. Agent to separate the layers that product descriptions often blend together.