AI agent vs chatbot vs assistant comes down to who actually does the work between your request and the finished result. A chatbot only answers what you're asking right now, with no memory beyond the current session. An assistant has a fixed role and attached knowledge, but still stops at suggesting text. An agent plans on its own, uses tools, and delivers a result that goes beyond plain text.
I run my business with a workforce of eight AI employees on Claude, some of them agents, some assistants, each built for a different kind of task. Everything I write about AI agents is on the AI agents page.
What an AI agent fundamentally is, and how its loop of planning and acting works, I've already covered in What Is an AI Agent? Explained Simply. This article covers what that one leaves out: where the assistant ends and the agent begins, where the workflow sits in between, and when each form is the right choice.
AI agent vs chatbot vs assistant: five terms, one difference
One thing first: these five terms are not a binding taxonomy. No standard defines when software becomes an agent, and vendors name the same product differently depending on their marketing. What follows are working definitions I use to tell offers and tools apart. They are useful because they draw the line along the question of who plans and who operates tools, not along product names.
A chatbot is a question and answer window with no memory beyond the current conversation, every new session starts from zero. An assistant is a chatbot with a fixed role and attached knowledge, say a Claude project or a custom GPT, handling a task within a familiar frame but rarely operating tools outside the conversation itself. An agent is software that turns a goal into its own sequence of steps, uses tools such as web search, file access, or a connected account, and corrects itself along the way. A workflow is a predetermined process with an AI component at one or more points, always taking the same path regardless of the individual case. An AI employee, finally, is an agent with a personnel file, a job, and approval rules, someone you manage the way you'd manage a person, not just operate like a tool, explained in full in What Is an AI Employee?.
So the difference between an AI agent and a chatbot doesn't start at the interface, it starts with whether the software decides and acts on its own between your request and the result, or just answers.
Comparison table: chatbot, assistant, agent, workflow, employee
|
Chatbot |
Assistant |
Agent |
Workflow |
AI employee |
| Plans on its own |
no |
barely |
yes |
no, fixed process |
yes |
| Uses tools |
rarely |
a few, fixed ones |
yes, depending on the task |
yes, hard wired |
yes |
| Memory beyond the session |
no |
yes, via project and knowledge |
only what you give it |
no |
yes, via the personnel file |
| What happens on a mistake |
you notice immediately |
you check the text before using it |
it can keep planning in the wrong direction |
the process breaks or stubbornly runs wrong |
an approval gate catches it before outward effect |
| Effect outside the conversation |
none, text only |
none, text only |
possible once tools are connected |
possible, without a check |
only after your approval |
The table shows why the terms get mixed up so often in practice: an assistant and an agent are technically close, and the real difference rarely sits in the interface, it sits in whether a check is built in before anything takes effect outside the conversation.
Where the assistant stops and the agent starts
An assistant, like a Claude project with attached knowledge or a custom GPT, works within a frame you've set up beforehand: a role, a handful of documents, a conversation history. It almost always stops at a text answer, even a very well tailored one. What it doesn't do on its own is choose several tools one after another, check an intermediate result, and derive the next step from it.
That's exactly where an agent begins. Once software gets a goal instead of a single question, decides for itself which tool it needs, and evaluates an intermediate result before continuing, that line has been crossed. Inside Claude Code, this line shows up technically in the built in subagents such as Explore, Plan, and general-purpose, each with its own context window and its own job, according to Anthropic's subagent documentation (code.claude.com/docs/en/sub-agents). That ability makes an agent more capable, but also riskier: it can handle cases nobody anticipated, and along the way it can also take paths you didn't want. That's why every agent in my workforce comes with a fixed rule for what it's allowed to decide on its own and what it has to bring to me for approval before anything takes effect outside.
Olaf against Sonja: same new contact, two roles
Here's what that difference looks like for me in practice, using a new contact who reaches me through the website or a message.
Sonja is my assistant for email. She sees the new message in the inbox, knows my voice from many earlier emails, and writes a reply draft from that. She doesn't research the sender, doesn't score anything, she works with what's already in the message. The result of her work is a finished draft waiting for me. I read it, check whether the tone and the facts are right, and decide whether it goes out as is.
Olaf is my lead scout and works as an agent. Instead of waiting for a single message, he actively goes through possible new contacts, looks up publicly available information about a company, and judges whether it fits my target group. The result of his work isn't a reply, it's a scored assessment with reasoning: good fit, probably not a fit, and why. Whether that turns into an actual outreach is my call, Olaf doesn't make that decision himself.
Same trigger, a new contact, two completely different ways of working: Sonja stays with text that I approve. Olaf plans, researches, and judges on his own, before any decision that I have to make even comes up.
When a chatbot or workflow is enough, and when it isn't
Not every task needs an agent, and that's not a limitation, it's often the right call. A chatbot is enough for a single, self contained question where you check the result right away and don't need it again afterward, say a quick phrasing fix for one sentence. A workflow, a predetermined process with an AI component, is the better choice when a task genuinely runs the same way every time and every branch is known in advance, say a fixed rule for invoice numbers or a routing rule that never changes.
An agent only pays off once a task needs context and judgment, the steps differ from case to case, and you can't fully plan them out ahead of time. Where exactly that line runs is also a cost question: an agent that calls several tools on every run uses more than a fixed automation that repeats the same path over and over. For a task that genuinely never varies, an agent is often the more expensive and unnecessarily complicated option.
Frequently asked questions
Is ChatGPT an agent?
By default, ChatGPT is a chatbot with memory across the current conversation. Once you give it tools such as web search or connected actions and it decides for itself when to use them, that same interface effectively becomes a simple agent. So the label depends less on the product name than on which tools are active and who decides when they're used.
Does an agent need code?
No, not necessarily. You can set up a first agent of your own without any programming knowledge directly inside a chat app, the steps are in Build an AI Agent With Claude: A Guide. If you run agents like Olaf with their own tools and recurring roles, you'll eventually reach a point where some technical understanding helps, but it isn't required to get started.
What happens when something goes wrong?
That depends on whether an approval gate is built in. Without a check, an agent can let a wrong assessment take effect immediately, say sending an email that shouldn't have gone out. With a gate, as with Olaf and Sonja, a mistake stays a draft or an assessment that you still review before it has any outward effect, instead of an already real consequence.
Is a custom GPT an assistant or an agent?
A custom GPT is essentially an assistant: a fixed role, attached knowledge, but tools usually limited to individual, predefined actions. Once a custom GPT independently chooses several tools in a row and corrects itself along the way, it slides closer to being an agent, the line in practice is fluid rather than hard.
What does an AI agent cost per month?
That depends on the tool you build it with and how often it calls tools. For an agent built on Claude, you typically pay the same plan as for regular chat, current prices are in Claude Code Pricing: Plan or API?. A workflow with fixed automation can be cheaper if the task genuinely runs the same way every time.
Where to go from here
If you'd rather set up a first assistant before attempting an agent, three ways to do that are in Build an AI Assistant: Three Paths. What sets an agent apart from an employee with a personnel file, and why that management side ends up mattering more than the raw technology, is in What Is an AI Employee?.
All eight ready made packages, from Sonja's assistant role to Olaf's agent role, are in my community: Community.
