Chatbot vs AI agent: the difference, and which one you need
The short answer: a chatbot answers questions, an agent finishes work. If your pain is customers asking the same twenty questions over and over, a chatbot solves it. If the pain is that after the answer somebody still has to go and do something in three systems, you need an agent. Most of what we actually ship is a combination of the two.
Last updated: August 27, 2026
The difference in one table
| Chatbot | AI agent | |
|---|---|---|
| What it does | Answers a question | Carries a task to completion |
| What it can reach | Text and knowledge | Systems, permissions and tools |
| What starts it | A person typing a message | A trigger, a schedule, or a person |
| How you know it worked | The user is satisfied with the answer | There is a result you can check in a system |
| Where it fails | A question outside what it was given | Permissions or tools that are not quite right |
| When to pick it | Repeat questions, information that exists | Repeat work that touches systems |
What a chatbot is genuinely good at
A good chatbot is a useful thing, and the word has been worn down enough that people undervalue it. It takes the information you already have, the price list, the policies, the common questions, the opening hours, and answers in the customer's own words instead of pointing them at a page.
It fits when most enquiries are requests for information, when that information exists and is written down, and when the answer does not change depending on who is asking. In that situation it takes real load off a team within days.
Where it stops: the moment an answer requires looking at a live system. "What is the status of my order" is a query rather than a question. And the moment the customer wants something to happen, the chatbot has reached its edge.
What an agent adds
An agent has three things a chatbot does not: tools, permissions and state.
Tools are the actions it is allowed to take. Read from the CRM, draft an invoice, book a meeting, open a service ticket. A closed list, defined in advance.
Permissions are the scope of those tools. An agent that needs to read a calendar gets read access to a calendar and nothing more. That is what stops it doing something expensive by mistake.
State is knowing where it is inside a task that runs longer than one message. It can wait for an approval, carry on tomorrow, and know what it has already done.
Because of those three, an agent is measured differently. A chatbot succeeds when the user is happy. An agent succeeds when there is a record in a system you can open and check.
How to tell which one you need
Look at what happens after the answer. If the conversation ends with the customer having information and leaving, a chatbot is enough. If it ends with somebody on your team opening three screens and typing, you are in agent territory, and the chatbot has only moved the work along.
A second question worth asking: how much of your team's time goes on the questions themselves, and how much on what follows them. The answer usually surprises people, and we wrote about measuring it in the automation project process guide.
The combination most companies end up with
In practice most of what we put live is both. A conversation layer that talks to the customer, and an agent behind it doing the work. The customer does not see the seam, and they do feel it: they ask when the order arrives, get a real date out of the system, ask to change the address, and it actually changes.
The rule we work to: anything that reaches a customer, moves money, or writes a record other systems read goes through a person, at least at first. Reading, classifying, drafting and enriching usually do not. As the team watches it get things right, that checkpoint moves or comes out.
What goes wrong
An agent with too many permissions. If it can do anything, eventually it will do something you did not want. A closed list of actions is what lets you sleep.
A chatbot pretending to be an agent. It sounds confident, gives an answer that sounds right, and tells the customer something it never checked against a system. That is the most common damage.
No route to a person. Every system needs a clear way to hand a conversation to a human with all the context attached. Without it, the angry customer starts from scratch and gets angrier.
FAQ
How long until an agent is running?
Most go live 4 to 8 weeks from the first session. We build in production conditions from week 1, on real data and real integrations, because the distance between a demo and a working system is where these projects die.
Can we start with a chatbot and grow into an agent?
Yes, and it is often a good path. The chatbot teaches you what customers actually ask, which is exactly the information you need to choose the right agent to build next.
What stops an agent inventing answers?
It should not be answering from its own knowledge when a system holds the answer. It is sent to fetch the value, and when it has no source it hands off to a person rather than guessing. That is a design decision rather than a property of the model.
Do we need a developer on staff to maintain one?
Not for day-to-day operation. Changing wording, a routing rule or an approval threshold is configuration. A structural change to the process itself is real work, and we either do it or teach your team to.
What does it cost to run?
Model costs on a typical business process are usually the small line, measured in tens to hundreds of dollars a month depending on volume. The meaningful cost is building and maintaining it.
Next step
We build AI agents wired into the systems you already run, with a human approval point wherever judgment matters. If you are mid-decision between the two, talk to us and we will tell you which fits the pain you described.
