Chatbot
Software that holds a conversation: it matches a question to an answer, a script or a generated reply.
- Lives in a chat window
- Answers from scripts, FAQs or documents
- Hands actions to a person or a form
- Success: the right answer
Guide · Agent harness
AI agent vs chatbot: a reply in the chat window, or the task done behind it.
A chatbot answers within a conversation. An AI agent works toward a goal: it looks things up, calls systems, takes the step and reports back — and a chat window is only one way to start it.
Many products called agents are chatbots, and many chatbots now use the same language models as agents. The difference is in what the software around the model lets it do. This guide shows how each works, walks one customer request through both, covers who is accountable for what a bot says, and gives a test for choosing.
Updated · 8 sources
AI agent
Software that pursues a goal: it decides the next step, uses tools and systems, and checks the result.
- Started by chat, email, events or schedules
- Reads and writes in business systems
- Takes actions within permissions
- Success: the task completed
What is the difference between an AI agent and a chatbot?
A chatbot is software that converses: it answers questions from scripts, a knowledge base or a language model, and leaves actions to people. An AI agent pursues a goal on someone’s behalf: it plans steps, calls tools and business systems, acts within its permissions and checks the result. In AI agent vs chatbot, the question is what happens after the reply — a chatbot stops at the answer; an agent carries on until the task is done.
Six kinds of bot
AI agent vs chatbot vs assistant: six steps from scripts to actions.
Google Cloud separates bots that follow pre-defined rules, assistants that respond and recommend while the person decides, and agents that proactively perform tasks. Most products sit somewhere on this line.
AnswersActs
123456
- Rule-based botButtons and decision trees. ELIZA, in 1966, matched keywords to scripted replies.
- Intent-based chatbotClassifies each message into an intent and returns its response.
- LLM chatbotGenerates fluent replies from a language model; still only talks.
- Chatbot with retrievalAnswers from your documents, ideally with citations.
- AssistantSuggests actions and drafts them; a person carries them out.
- AI agentChooses and takes the steps itself, within granted tools and approvals.
Mechanics
How a chatbot answers and how an AI agent acts, step by step.
A classic chatbot is a classifier with scripts behind it: Dialogflow, for example, defines an intent as “an end-user’s intention for one conversation turn”. An agent replaces the fixed script with a loop in which a model chooses the next action.
An intent-based chatbot, per message
- Message
The customer types a question.
- Match intent
The message is classified against example phrases for each intent.
- Fill details
The bot asks for missing values, such as an order number.
- Respond
A scripted answer, or one call the designer wired in advance.
- Fall back
No match: “Sorry, I didn’t get that,” or a hand-off to a person.
An AI agent, per task
- UnderstandThe model reads the request and the context it is given.
- DecideIt chooses a tool and its arguments, or answers.
- ActThe application runs the tool in the business system.
- CheckThe result is read back; risky steps wait for approval.
Back to deciding, with the new result — until the task is done or a person must decide.
One request, two ways
A refund request, handled by a chatbot and by an AI agent.
Same customer, same message. The chatbot gives a correct answer. The agent resolves the case — and pauses where the business wants a person to decide.
Chatbot
You
My order 4471 arrived damaged. Can I get a refund?
Agent
Sorry to hear that. Damaged items can be refunded within 30 days. Please fill in the returns form at the link below and attach a photo. Our team will reply within 3 working days.
You
I already sent a photo yesterday.
Agent
I can’t see previous messages. Please use the returns form so our team can help.
AI agent
You
My order 4471 arrived damaged. Can I get a refund?
Agent
I found order 4471 and your photo from yesterday. The damage is covered by the returns policy, and the item cost 64.00. I can refund 64.00 to your card now — shall I?
You
Yes please.
Agent
Done. The refund is issued, the order is marked returned and you have a confirmation email. No need to send the item back.
What the agent needed that the chatbot lacked
- ToolsOrder lookup, the message history, the refund system and email.
- PermissionA grant to issue refunds, and a rule on when to ask first.
- A recordWhat it looked up, what it proposed, what the customer agreed to.
Cost, value and liability
What moving from chatbot to AI agent costs and returns.
Agents do more per request, so they use more of the model per request. The prize forecast for customer service is large, and the courts have already shown who answers for a bot’s words.
The bot speaks for the company. In Moffatt v. Air Canada (February 2024), the tribunal rejected the argument that the chatbot was a separate legal entity responsible for its own actions, and held the airline responsible for the information on its website — chatbot included.
Controls
Six things to require before a chatbot becomes an AI agent.
A chatbot that is wrong gives a bad answer. An agent that is wrong takes a bad action. These controls belong outside the model, where a prompt cannot talk its way past them.
Before a bot can act
- Answers grounded in your sourcesPolicies and prices come from your documents and systems, with a citation.
- Tools granted per agentThe refund agent can refund; it cannot change prices or delete accounts.
- Confirmation before changeMoney, records and outgoing messages wait for the customer or a person.
- A clean hand-offWhen the agent stops, a person sees the whole case, not a transcript fragment.
- Every step recordedLookups, proposals, approvals and actions, kept with the conversation.
- Tested on real casesRun past conversations through the agent and compare its actions with what should have happened.
Decision guide
AI agent vs chatbot: choose by what must change, and what is at stake.
OpenAI’s guide draws the line plainly: applications that use a model but don’t let it control the workflow — “think simple chatbots” — are not agents. Decide by what a good outcome requires.
High stakes
Grounded assistantPolicy, benefits or product questions where a wrong answer has a cost: cite every source, route edge cases to people.chatbot with retrieval
Governed agentRefunds, payments, claims, account changes: tools by grant, approvals before change, a record of every step.AI agent
Low stakesFAQ chatbotOpening hours, store finder, tracking links. Scripts or a model with a short knowledge base are enough.chatbot
Self-acting agentResetting a password, rebooking a slot, updating an address — narrow tools, no approval needed.AI agent
Answers from contentActions in systemsDoes a good outcome need something to change in a system?
No, an answer is the outcomeA chatbot — ideally grounded in your documents with citations.
Yes, and a person can do itAn assistant that drafts the change for a person to make.
Yes, and it should be done for themAn AI agent with granted tools, confirmation before change and a full record.Read more →
On Botminds
How a Botminds agent goes from a chat message to a recorded action.
A Botminds agent can be reached in a chat, from a schedule or through an OpenAI-compatible API. Whichever way it starts, the same harness checks its tools and keeps its record.
- Request
A person asks in chat, or your system calls the agent through the API.
- Context
The harness assembles what this run needs, within a budget; you can preview the exact prompt.
- Grant check
Every tool call is checked against the agent’s grant; a tool outside it is refused.
- Approval
On thinking and autonomous runs, a tool set to wait pauses on a card: approve, reject or edit.
- Record
The request, each tool call, each approval, the answer and its cost are kept.
Read more in the agent harness, and build your first one with your first agent in the Academy.
Questions people ask
AI agent vs chatbot: common questions.
The technology behind itAgent harnessThe runtime around every agent→What is the difference between an AI agent and a chatbot?
In AI agent vs chatbot, a chatbot converses and answers, leaving actions to people or forms. An AI agent pursues a goal: it decides the next step, calls tools and business systems, acts within its permissions and checks the result.
Are AI agents and chatbots the same thing?
No. Many chatbots now use the same language models as agents, but a chatbot only replies. It becomes an agent when the software lets the model choose and take actions in systems toward a goal.
What is the difference between an AI agent, a chatbot and an AI assistant?
A chatbot answers. An assistant answers and suggests or drafts actions, while the person decides and acts. An agent decides and acts itself, within the tools and approvals it is given.
Is ChatGPT a chatbot or an AI agent?
It started as a chatbot built on a language model. When it is allowed to browse, run code or use tools to complete a task, it behaves as an agent. The label depends on what it is permitted to do in that use.
Can a chatbot be upgraded to an AI agent?
Yes, by giving the model tools to look things up and act, rules on which tools need confirmation, and a record of every step. The conversation can stay the same; what changes is what happens behind it.
Who is responsible when a chatbot gives wrong information?
The company that runs it. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline liable for its chatbot’s wrong description of its bereavement-fare refund policy and rejected the argument that the chatbot was responsible for its own actions.
Are AI agents more expensive than chatbots?
Per request, usually yes: Anthropic reports agents use about four times the tokens of chat interactions. The return comes from resolving the request instead of passing it to a person.
Do AI agents replace customer service teams?
They take over routine cases and prepare the rest. Well-run agents hand a person the full case — what was found, what was proposed — when a decision needs judgement or authority.
Sources
- Weizenbaum: ELIZA — a computer program for the study of natural language communication (1966)doi.org
- Google Cloud: Dialogflow ES intentsdocs.cloud.google.com
- Google Cloud: What are AI agents?cloud.google.com
- OpenAI: A practical guide to building agentscdn.openai.com
- OpenAI: Function calling guidedevelopers.openai.com
- Anthropic: How we built our multi-agent research system (June 2025)anthropic.com
- Gartner: agentic AI will resolve 80% of common customer service issues by 2029 (March 2025)gartner.com
- Moffatt v. Air Canada, 2024 BCCRT 149decisions.civilresolutionbc.ca
Read next
See it on
your own work.
Bring one process and its documents. We’ll show you the agents, the controls and the record, running on your data.