> ## Documentation Index
> Fetch the complete documentation index at: https://cntillion.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Chatbot vs AI agent: what changes for customer support

> A chatbot answers from a script and ends the chat. An AI agent works for your business inside its rules. Learn the difference and build an agent.

GuideChatbot or AI agent: who answers your customers, and who they work for.

The chatbot vs AI agent question sounds like a vocabulary debate. In practice
it decides what your customers get when they write to you, and what your
business gets out of the conversation. This guide explains the difference, the
risks on each side, and how to set up an agent that works for your business.

## What a chatbot is

A classic support chatbot is a decision tree with a text box in front of it.
Someone on your team maps the questions they expect, writes an answer for each,
and links them with buttons or keywords. When a customer's words match a
branch, the bot replies. When they do not, it shows a menu again or hands over
a link to a help article.

Most chatbots are built to deflect. The goal is to close the conversation
without a person, and success is counted in tickets avoided. That is a fair
product for a queue of simple, repeated questions such as opening hours or a
returns address. It breaks down in three places:

* **Questions nobody planned for.** Every tree is a bet on what customers will
  ask. A customer with a slightly different problem falls off the edge.
* **Anything that needs doing.** A bot that can only answer cannot change an
  order, check a delivery or book a slot. The customer still has to call.
* **Moments that decide revenue.** A deflection bot has no goal beyond ending
  the chat, so it treats a customer about to cancel the same way it treats a
  question about parking.

## What an AI agent is

An AI agent reads the customer's message in their own words, finds the
relevant facts in material you gave it, and writes an answer. It can also act:
call your systems to look up an order, update a cart, open a ticket or hand
the conversation to a person.

The useful way to think about an agent is as a representative of your
business. It knows your products, your policies and how you talk to people. It
has a goal you set, such as a resolved problem, a booking, or a customer who
stays. It works inside rules you wrote down. Every promise it makes should be
one you would sign.

That changes what you measure. A chatbot is judged on how many conversations
it ended. A representative is judged on what happened for the customer and the
business: was the issue resolved, did the handoff reach the right person, did
the customer come back.

## The differences that matter

| | Chatbot | AI agent |
| - | - | - |
| Understands | Keywords and buttons you mapped | The customer's own words |
| Answers from | Replies written in advance | Your knowledge, retrieved for each question |
| Can act | Rarely, through fixed flows | Calls the tools you enable |
| Goal | End the conversation | The outcome you define |
| When stuck | Loops or shows a menu | Says it does not know and hands off |
| You check it by | Counting deflected chats | Reading outcomes and conversations |

## What to watch for

An agent that writes its own answers and takes actions needs more care than a
tree of canned replies. Most problems come from a few causes.

**No knowledge, confident tone.** A language model with no material to draw on
still writes fluent sentences. Give the agent your real content before it
talks to anyone, and check that it says "I don't know" when the answer is not
there.

**Rules that live only in a prompt.** Instructions such as "never offer a
refund over 50 dollars" are guidance the model follows most of the time. A
rule that must hold every time belongs in the tool itself: what the tool
is allowed to do, which inputs it accepts, and whether the customer must
confirm. Write policy where it is enforced, and keep the prompt for tone and
judgement.

**Actions without confirmation.** Changing an order or an address should need
an explicit yes from the customer. Read the conversation afterwards to see
whether that step held.

**No way out.** Some customers want a person, and some problems need one. Plan
the handoff before launch: who receives it, on which channels, and what the
customer is told while they wait.

**No test before customers see it.** Run simulated conversations against the
agent, including a difficult customer and a customer who tries to get it to
break its rules, before you switch a channel on.

## The other side is getting agents too

Customers are starting to use personal AI assistants that write, compare and
follow up for them. When one of those reaches your support, it will ask
precisely, keep asking, and read your terms closely. A bot built to end
conversations does poorly against that kind of counterpart. A representative
that knows your policies and holds to them is what you want on your side of
the conversation.

## Build an agent with Centillion

[Centillion](https://cntillion.com/docs/introduction.md) builds AI agents that answer your customers from
knowledge you control, use the tools you enable, and hand off to your team
through the helpdesk. Start from
[creating an agent](https://cntillion.com/docs/agents/create-an-agent.md).

**1. Write who it works for**

Open **Agent → Instructions** and replace the role's template in the
**General** editor with who the agent works for, who it talks to, what to
do when it does not know something, and how formal to be. Facts about your
products go into knowledge, where you can update them without touching the
instructions.

![Agent configuration with the system prompt editor, Call and Chat first-message tabs, and voice and model settings](https://mintcdn.com/centillion/n9koGqutqlVBNAYF/images/agents/create-agent-02-instructions-editor.webp?fit=max\&auto=format\&n=n9koGqutqlVBNAYF\&q=85\&s=fb3e63f98578ad87bf6d2d8e97953842)

Agent configuration: the system prompt and first message, with voice and model settings on the right.

**2. Add guardrails**

In the **Guardrails** editor on the same page, list the topics the agent
should stay away from, such as competitors, or legal or medical advice.
Guardrails are added to the agent's prompt; nothing checks a reply against
them afterwards. The separate **Guardrails** page sets the rate limit and
spam detection.

![Guardrails page with the rate limit set to 20 messages every 240 seconds and the spam detection switch](https://mintcdn.com/centillion/iOhd2RAouod3JP9I/images/agents/guardrails-01.webp?fit=max\&auto=format\&n=iOhd2RAouod3JP9I\&q=85\&s=60062445b5d7ee827227399cd28fe355)

Guardrails: the rate limit for one visitor and spam detection, with the saved settings on the right.

**3. Give it your knowledge**

Under **Knowledge base → Data sources**, add your website, documents,
Q\&A and Notion pages. Training runs on its own after every change. See
[how to train an AI chatbot on your own data](https://cntillion.com/docs/guides/train-ai-agent-on-website.md).

**4. Let it act**

Open **Tools**, choose **Add tool**, then **New tool → Integration**, and
choose what the agent can do:
create a ticket, collect leads, call your API, or use a connected
integration such as Shopify or Stripe. An instruction to look up orders
does not grant that access by itself; the tool does.

![Integration catalog with the tool search box, a type filter and the Escalations, Custom tools and Collect leads cards](https://mintcdn.com/centillion/WLpDA0mJ1ryQ1IZ5/images/agents/custom-actions-02-catalog.webp?fit=max\&auto=format\&n=WLpDA0mJ1ryQ1IZ5\&q=85\&s=569de53cd2c0d87b9df9b3964df59e35)

The integration catalog of escalations, custom tools, lead collection and integrations.

**5. Plan the handoff**

Under **Tools**, **Escalations** creates a ticket for follow-up, and
**Centillion live chat** pauses AI replies so a teammate continues the
conversation from the [helpdesk inbox](https://cntillion.com/docs/helpdesk.md). Both need Helpdesk
enabled and the customer's channel selected.

**6. Test, then measure**

Under **More → Tests**, describe a customer and one clear outcome, and
run simulated conversations. After launch, **Evaluations** in the agent
sidebar checks real
conversations against criteria such as Issue resolved, Accurate handoff or
Goal completed.

For tasks with several steps, such as a return or a booking, write a
**Procedure**: a trigger, the steps and the tools each step needs.
Every agent has **Version control**. A version saves the prompts, model, knowledge,
tools, procedures and evaluation criteria together, so you can see which
rules a conversation ran on.

## Questions

### Is an AI agent just a chatbot with a language model?

It uses a language model, and the bigger change is the job it does. A chatbot
picks a reply from a list. An agent reads the question, retrieves the relevant
part of your knowledge, writes an answer, and can call tools you enabled.

### Will an AI agent make things up?

It can, when it has nothing to answer from. Give it your real content, tell it
in the instructions to say when it does not know, and test with questions it
should not be able to answer. In Centillion, the agent answers from the six
most relevant passages it finds for each question.

### Do guardrails block forbidden replies?

In Centillion, guardrails are instructions added to the agent's prompt.
Nothing checks a reply against them afterwards. For a rule that must always
hold, enable only the tools the agent needs, and let your own API behind a
custom tool refuse what your policy does not allow. **Only use in
procedures** keeps a tool out of reach until a procedure calls for it.

### Can the agent change orders or accounts?

Only through tools you enable under **Tools**. The Shopify tools, for
example, ask the customer to confirm before they create an order, make a
return or exchange, or change a profile or address.

### When should I keep a person in the loop?

When the customer asks for one, when the agent cannot find the answer, and
for decisions you would not delegate, such as exceptions to your policies.
Set up **Escalations** or **Centillion live chat** so those conversations
reach your team.

## Try it

Create an agent, write two paragraphs about who it works for, and give it your
website. Then ask it the question your customers ask most, and one it should
not be able to answer.

> [card] [Start free](https://app.cntillion.com/sign-up?utm_source=docs\&utm_medium=article\&utm_campaign=guides-chatbot-vs-ai-agent)
>
> Start on the Free plan, or take a seven-day trial of a paid plan. Your agent is live on your site in an afternoon.

> [card] [Book a demo](https://cal.com/patrick-turricelli-smgfks/15min?utm_source=docs\&utm_medium=article\&utm_campaign=guides-chatbot-vs-ai-agent)
>
> Fifteen minutes with the team, on your own use case.

Or ask the assistant on this page.

## Related

> [card] [Creating an agent](https://cntillion.com/docs/agents/create-an-agent.md)
>
> Instructions, guardrails, the model and Compare.

> [card] [Train an agent on your data](https://cntillion.com/docs/guides/train-ai-agent-on-website.md)
>
> How retrieval works, and why answers go wrong.

> [card] [Custom tools](https://cntillion.com/docs/agents/custom-actions.md)
>
> Call your API, run code on your site, or show a card in chat.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.