The difference in one table
| AI chatbot | AI agent | |
|---|---|---|
| What it does | Answers questions in conversation | Carries out a task, often in several steps |
| What it works from | Documents and knowledge you provide | Knowledge, plus tools that read and write in your systems |
| Where the result lands | In the chat window | In your CRM, order system, calendar or inbox |
| What it needs | Good content, kept up to date | Integrations, permissions, limits and logs |
| What can go wrong | A wrong or invented answer | A wrong action, which may be harder to undo |
| How it is judged | Were the answers correct and useful? | Was the task completed correctly, and were the exceptions handed over? |
The two words are used loosely, and many products sit between them. A chatbot that can look up an order status has taken its first step toward being an agent. What matters is not the label. It is whether the system only talks or also acts.
When each one fits
Choose a chatbot when
- People ask questions that your documents already answer
- The goal is faster answers for customers or staff
- No system needs to change as a result of the conversation
- You are new to AI and want a contained first project
- A wrong answer is inconvenient, not costly
Choose an agent when
- The request is only resolved when something is done, such as a refund, a booking or an update
- Staff spend their time moving information between systems
- The steps follow a pattern that a person could explain
- Your systems have APIs or another safe way in
- You can define what the agent may do alone and what needs approval
What an agent needs that a chatbot does not
Moving from answers to actions changes the project. The language model is the smaller part. The larger part is the engineering around it.
- Tools. A defined set of actions the agent may call, such as find customer, read order or create ticket, each built and tested like any other integration
- Permissions. The agent acts with the rights of a role, never with more access than a person in that job would have
- Limits. Amounts, record types and volumes it may touch without approval
- Approval steps. Points where a person confirms before the action is taken
- Logs. A record of what it read, what it decided and what it did, kept for review
- A way out. Clear handover to a person when the agent is unsure or the case is outside its scope
This is why an agent project is closer to a systems integration project than to setting up a chat widget. See how to identify business workflows suitable for AI for choosing where to begin.
Risk and control
| Question | Chatbot | Agent |
|---|---|---|
| What is the worst mistake? | A wrong statement to a customer | A wrong change in a system of record |
| How is it prevented? | Answers limited to approved sources, with the source shown | Narrow tools, limits and approval for anything costly |
| How is it caught? | Sampling of conversations | Logs of every action, reviewed, with alerts on unusual patterns |
| Can it be undone? | A correction is sent | Only if actions were designed to be reversible |
| Who is accountable? | The owner of the content | The owner of the process the agent works in |
A worked example
The same customer message, handled two ways
An online retailer, invented for this example, receives the message: my order has not arrived, can I get a refund?
- ChatbotExplains the refund policy, tells the customer where to find the order number and gives a link to the refund form. The customer still has to do the work, and so does the support team.
- AgentIdentifies the customer, finds the order, checks the carrier status and sees that the parcel is marked lost. The refund is within its limit, so it issues it, updates the order, confirms to the customer and logs each step.
- Agent, a different caseThe order is above the limit and the carrier shows it as delivered. The agent does not decide. It prepares a summary and hands the case to a person.
The agent saved work in the first case and stayed inside its limits in the second. Both behaviors had to be designed.
A checklist to decide
Tick what is true of the work you want to hand over. With none or one ticked, a chatbot is likely enough. With several, you are describing an agent.
Signs you need an agent
0 of 7 checkedA sensible path from one to the other
- 01
Answer
Start with a chatbot over your own content. You learn what people ask and where the content is weak.
- 02
Look up
Give it read-only access to one system, such as order status. Nothing can be damaged, and answers become personal.
- 03
Propose
Let it prepare actions for a person to approve with one click.
- 04
Act within limits
Allow it to complete the low-risk cases alone, with logs and sampling, and keep approval for the rest.
Each step is useful by itself, and each shows whether the next is worth taking. Our guide to planning an AI pilot explains how to measure a step before moving on.
What to do next
Collect twenty real requests from your inbox and mark each as needing an answer or needing an action. The split tells you which you need first. Our AI agent solutions catalogue shows agents by job, each with its status stated, and AI agents and automation explains how we build them.
Or send us the twenty. We will tell you what we would automate, what we would leave to people and where a simple rule would do.
Common questions
Is an AI agent just a chatbot with integrations?
In part. Both use a language model. An agent adds the ability to choose and call tools, to work through several steps toward a goal and to check its own progress. That addition brings permissions, limits, logging and testing with it, which is most of the engineering work.
Are AI agents safe to use with customer data?
They can be, when designed for it. The agent should act with the least access the task needs, within limits, with every action logged and with approval for anything costly or irreversible. Ask how data is handled by the model provider as well. Safety comes from this design, not from the model.
Do we need a chatbot before we can build an agent?
No, but it is a sound first step. A chatbot shows what people ask and exposes gaps in your content at low risk. Some agents have no chat at all. They are triggered by an email, a form or a schedule and work in the background.
Which costs more to build and run?
An agent, in most cases. It needs integrations with your systems, permission and approval design, and more testing. Running costs are also higher, because completing a task takes several model calls. The return is larger too, since finished work saves more time than answered questions.
