# AI agents vs chatbots

> A scripted chatbot follows a decision tree and answers a fixed set of questions. An LLM assistant understands free-form language and answers or drafts from your own documents, but a person still does the work. An AI agent is given a goal and completes the task inside your systems, with rules about when to stop and ask. Choose by what the job needs: answering, helping or finishing.

Source: https://risendigital.io/guides/ai-agents-vs-chatbots  
Publisher: Risen Digital  
Updated: 2026-09-21

## At a glance

- **Scripted chatbot:** Cheapest to build and run and the safest, and it fails the moment someone asks something that is not in the script.
- **LLM assistant:** Understands any wording, answers from your documents and never touches a live system, so its worst case is a wrong sentence a person can catch.
- **AI agent:** Completes tasks in your CRM, order system or inbox, which is where the value and the risk both live.
- **Rule of thumb:** Fixed questions want a chatbot, judgement and wording want an assistant, a task that must be finished wants an agent.

A lot of what is sold as an AI agent is a chatbot with a new name, and some of what is called a chatbot is quietly doing agent work. The label matters less than the fit. This guide sets the three kinds out plainly, compares them on the things that decide a project, and gives you a rule for matching each one to a job.

![Two colleagues reviewing a conversation on a laptop together](https://risendigital.io/images/experience/dev-two-1600.jpg)

## What is the difference between a scripted chatbot, an LLM assistant and an AI agent?

They differ in what they take in and what they are allowed to do. A chatbot matches a message to a script, an assistant understands the message and produces a response, and an agent takes a goal and acts on your systems until the job is done or a rule tells it to stop.

A scripted chatbot is a decision tree with a chat window on the front. Every path through it was written by a person, so it is predictable, cheap and easy to test, and it has no idea what to do with a question nobody anticipated. It usually ends with a button that says talk to a human.

An LLM assistant puts a large language model behind the window. It reads whatever the person writes, finds the relevant passages in the documents it has been given, and composes an answer or a draft. It is only as honest as the documents it is grounded in, and it does not act. If the answer is that a refund is due, a person still issues the refund.

An AI agent is an assistant that has been given tools and permission to use them. It can look up the order, check the policy, create the replacement and send the email, and it has explicit limits that tell it when to hand the case to a person instead. The model is the same one the assistant uses. The engineering around it is what makes it an agent.

|  | Scripted chatbot | LLM assistant | AI agent |
| --- | --- | --- | --- |
| Where is my order? | Shows a tracking link if the person picks the right menu option | Reads the question, finds the order in the context it was given, explains the status | Looks the order up, sees it is late, checks the policy, offers and applies the remedy |
| What it needs from you | A script and a set of canned answers | Your documents and a way to retrieve them | Documents, tool connections, permissions and hand-off rules |
| What can go wrong | A dead end | A wrong or made-up answer | A wrong action in a live system |

## How do chatbots, assistants and agents compare on cost, risk and fit?

Each one trades something. The chatbot is cheap and safe but brittle. The assistant is flexible and low risk but leaves the work with a person. The agent finishes the work and costs the most to build and to run responsibly.

_Interactive on the web page: Weigh the three against what matters to you. The comparison scores a scripted chatbot, an LLM assistant and an AI agent from 1 to 5 on seven criteria: handling unexpected requests, completing tasks in your systems, cost to build, cost to run, risk of a wrong action, working without a person in the loop and time to first value. The chatbot leads on cost, safety and speed to launch and trails badly on flexibility. The assistant leads on flexibility with low risk but cannot finish a task. The agent is the only option that completes work in your systems, and pays for that with the highest build cost, run cost and risk, so weight cost and risk for a bounded set of questions, flexibility for help with judgement, and task completion when the job must be finished without a person._

> **Why the agent scores low on cost and risk** Not because it is a worse tool, but because it does more. Every action it can take needs a connection, a permission, a check and a way to undo it. That work is the price of finishing a job rather than describing one.

## Which should you choose for which job?

Decide by what the job needs done: a fixed set of questions answered, a person helped with judgement and wording, or a task completed from start to finish.

- **Choose a scripted chatbot when the questions are few and fixed** Opening hours, order tracking by number, appointment booking with a handful of options, a form that needs to be filled in a set order. The questions will not change much and a wrong answer would embarrass you, so a script you can read in full is the right tool.
- **Choose an LLM assistant when the value is in understanding and wording** Answering varied questions from a policy manual, summarising a long thread, drafting a reply for a person to send, helping staff find the right procedure. The person keeps the final action, so the assistant can be broad and still be safe.
- **Choose an AI agent when the job is only done when something has changed** A ticket resolved, an order amended, a lead enriched and routed, an invoice matched and posted. The process is documented, the systems have connections, and a mistake can be caught and reversed. Start it in draft mode and widen its permissions as it earns them.

## Why does an AI agent cost more to build and run than a chatbot?

An agent needs tools, permissions, guardrails, hand-off and monitoring, and it makes several model calls per task instead of one. None of that exists in a chatbot, because a chatbot never acts.

- **Tool connections** Each system the agent touches needs a tested integration with clear inputs and outputs, and someone who owns it when the vendor changes the API.
- **Permissions and limits** What it may read, what it may write, up to what value, and what it must never do. These are written in plain language for the business and enforced in code.
- **Checks before actions** Every proposed action is validated against the rules before it reaches a live system, and irreversible actions are held for a person.
- **Running cost** An agent reasons in steps and each step is a model call. A task that takes one call for an assistant may take many for an agent, and someone still spends time reviewing its work while trust is being built.

## Can a chatbot grow into an agent?

Yes, and that is the usual path. Start with an assistant grounded in your documents, measure which requests it handles well, then give it narrow and reversible actions for exactly those cases.

1. **Ground an assistant in your documents** Policies, procedures, product details. Let it answer and draft while people send. You learn which questions it gets right without any risk to a live system.
2. **Give it read access to your systems** Now it can look up the order or the account before it drafts. The drafts get better and you see which cases could be closed without a person.
3. **Add one reversible action** Creating a draft order, tagging a ticket, scheduling a callback. Something a person can reject in one tap. Measure how often they do.
4. **Widen the limits on evidence** Raise the approval threshold, add a second action, keep the audit trail. At this point you have an agent, and you built it one trusted step at a time.

## Frequently asked questions

**Q: Is an LLM assistant just a better chatbot?**  
A: It is a different tool that happens to live in the same window. A chatbot follows a script you wrote; an assistant understands language and answers from documents you gave it. The assistant handles far more variety, and in exchange you have to ground it carefully so it does not make things up.

**Q: Do I need an AI agent, or is a chatbot enough?**  
A: Ask what done looks like for the job. If done means the person got an answer, a chatbot or assistant is enough. If done means a record changed, an order shipped or a ticket closed without a person doing it, you need an agent, or you need to accept that a person finishes the task.

**Q: Are AI agents riskier than chatbots?**  
A: They can do more, so a mistake can cost more. That is managed by design: start read-only, hold every irreversible action for approval, validate actions against explicit limits, and keep a log. A well-built agent with narrow permissions is often safer in practice than a broad assistant whose drafts nobody checks.

**Q: Can one system be both a chatbot and an agent?**  
A: Yes, and many should be. A scripted front end can collect the basics, an assistant can understand the problem, and an agent can act on the cases that qualify. The person is brought in when none of the layers is confident. Designing the hand-offs between them is most of the work.

**Q: What does an AI agent need that a chatbot does not?**  
A: Connections to your systems, permissions that say what it may read and write, checks on every action, a hand-off route to a person and an audit trail. A chatbot needs a script and a set of answers. The gap between the two lists is the reason agents cost more and deliver more.

## Related pages

- [What is an AI agent?](https://risendigital.io/guides/what-is-an-ai-agent)
- [Conversational AI for business](https://risendigital.io/guides/conversational-ai-for-business)
- [AI voice agents for customer service](https://risendigital.io/guides/ai-voice-agents-for-customer-service)
- [AI agent use cases](https://risendigital.io/guides/ai-agent-use-cases)