# Chatbot vs AI agent

> A chatbot answers questions, either from scripted rules or from your documents with an LLM. An AI agent pursues a goal by deciding which steps to take and calling tools, such as looking up an order, updating a CRM record or booking an appointment. Choose a chatbot when people mostly need information; choose an agent when they need something done and the steps vary from case to case.

|  | Chatbot | AI agent |
| --- | --- | --- |
| Core job | Answer questions and guide users to information | Complete a task by planning steps and using tools |
| How it decides what to say or do | Scripted flows, or an LLM answering from retrieved documents | An LLM choosing among tools in a loop until the goal is met |
| Access to your systems | Usually read-only, often just a knowledge base | Reads and writes through APIs (CRM, orders, calendar, billing) |
| Typical examples | FAQ bot, policy assistant, website help widget | Refund handler, lead qualifier that books meetings, order status and change agent |
| Predictability | High for scripted bots, good for retrieval bots with citations | Lower; needs guardrails, permissions and approval steps |
| Risk if it gets something wrong | A wrong or unhelpful answer | A wrong action, such as an incorrect refund or record change |
| Build effort | Days to a few weeks | Several weeks, mostly spent on tools, permissions and testing |
| Running cost | Low; one model call per answer | Higher; several model calls and tool calls per task |
| Measurement | Answer accuracy, deflection, user satisfaction | Task completion rate, error rate, escalations, time saved |

## Choose Chatbot if

- Most requests are questions with answers in your documentation
- You want a low-risk first AI project with a quick payback
- Staff, not the bot, should take any action that changes data
- Your systems do not have usable APIs yet

## Choose AI agent if

- Requests usually end in an action someone performs by hand today
- The steps differ case to case, so a fixed script breaks often
- Your core systems expose APIs the agent can call safely
- You can define clear limits, such as refunds under a set amount
- You have volume high enough that partial automation of tasks pays back

## The short version

A chatbot talks. An agent acts. A chatbot takes a question and returns an answer, whether from a scripted decision tree or from an LLM reading your documents. An AI agent takes a goal, such as "process this return", and works out the steps: look up the order, check the policy, issue the label, update the ticket, tell the customer.

That difference changes everything downstream. An agent needs API access to your systems, rules about what it may and may not do, and testing on real cases. A chatbot needs a good knowledge base and a clear handoff to a human.

## Where a chatbot wins

**Lower risk.** The worst outcome is a wrong answer, and a retrieval-based bot that cites its sources makes those easy to spot. Nothing in your systems changes without a person.

**Faster and cheaper to build.** A [chatbot built on your docs](/guides/build-ai-faq-chatbot-from-docs/) is mostly content work: gathering, cleaning and chunking documents, then tuning answers. There are no write permissions to design. Our [AI chatbot cost page](/cost/cost-of-ai-chatbot/) breaks down where that budget goes.

**Cheap to run.** One retrieval and one model call per answer keeps per-conversation cost low and latency short.

**Works without APIs.** If your order system or practice software has no usable API, a chatbot can still answer policy and process questions while you sort out integration.

## Where an AI agent wins

**It finishes the job.** Many "questions" are really requests: change my delivery date, reschedule my appointment, send me last month's invoice. A chatbot can only explain how to do it. An agent can do it, which is where the real time savings are.

**It handles variation.** Scripted bots break when a customer's situation does not match a branch. An agent reasons about the case, picks the relevant tools, and escalates when it hits a limit.

**It works across systems.** A [lead qualification agent](/guides/ai-lead-qualification-agent/) can read a form, enrich the company, score fit, write to the CRM and book a meeting, all in one run.

**Better measurement.** Because an agent's output is a completed task, you can measure completion rates and hours saved directly, rather than inferring value from deflection.

## What we recommend

If most of what people ask is answered in documents you already have, build a chatbot first. It is cheaper, safer and teaches you what users really want.

When the logs show a clear pattern of requests that end in a manual action, add an agent for those specific actions with tight permissions and approval rules. That is the path in our guide to [upgrading a chatbot to an AI agent](/compare/migrate-chatbot-to-ai-agent/). Do not build an agent for a problem a well-written help article would solve.

Lumeor builds both. A custom agent typically sits in our $15,000–$60,000 band; see the [custom AI agent cost page](/cost/cost-of-custom-ai-agent/) for what drives the range.

## Frequently asked questions

### Is ChatGPT a chatbot or an agent?

The chat interface is a chatbot by default. When it browses the web, runs code or calls connected apps on your behalf, it is behaving as an agent. The distinction is about whether the system takes actions with tools, not about the underlying model.

### Is an LLM-powered chatbot the same as an AI agent?

No. Using an LLM makes a chatbot better at understanding questions and writing answers, but if it only retrieves and answers, it is still a chatbot. It becomes an agent when it can call tools and decide the sequence of steps to reach a goal.

### Are AI agents safe to let loose on business systems?

Only with limits. Give the agent the narrowest permissions it needs, require human approval above defined thresholds, log every tool call, and test against real historical cases before launch. Well-scoped agents are safe; open-ended ones with broad write access are not.

### Should we start with a chatbot and upgrade later?

Often, yes. A documentation chatbot builds the knowledge base, shows what people actually ask, and reveals which requests end in an action. Those action-ending requests become the first agent tools. Our upgrade guide covers that path.

### How much more does an agent cost than a chatbot?

At Lumeor, a chatbot trained on your documents typically runs $5,000–$20,000 to build, and a custom AI agent $15,000–$60,000. The difference is mostly integration work, permissions and testing for the actions the agent takes.

---

Source: https://www.lumeor.studio/compare/chatbot-vs-ai-agent/
Published by Lumeor Studio, Newtown, PA. hello@lumeor.studio. Get a written scope: https://www.lumeor.studio/start/