A chatbot replies. An AI agent takes action. Here is the actual difference, when each fits, what each costs, and how to pick without wasting a build.
Short version: a chatbot answers questions. An AI agent takes actions. A chatbot tells your customer where to reset their password. An agent resets it, logs the ticket, and updates the CRM. That gap, from talking to doing, is the whole distinction, and it decides which one is worth building.
We get asked to "build a chatbot" most weeks, and about half the time the thing the team actually needs is an agent. The words get used interchangeably, so here is the line that matters.
A chatbot is a conversation. Input text, output text. It can be scripted (decision trees) or LLM-powered (free-form answers), but its job ends at the reply. It knows things; it does not change things.
An AI agent is a conversation plus tools. It can call an API, read a database, file a record, send an email, or trigger another system, then check whether that worked and try again. It has a goal, not just a script, and it takes steps toward that goal. The reply is often the smallest part of the job.
Put simply: if the useful outcome is information, you want a chatbot. If the useful outcome is an action in another system, you want an agent.
Take "I want to change my delivery address."
A chatbot replies: "You can change your delivery address in Account Settings under Shipping." Correct, and the customer still has to go do it.
An agent replies: "Done. I have updated the delivery address on order #4471 to the new one and confirmed it is before the cutoff." Behind that sentence it authenticated the user, called the order API, validated the address, checked the shipping cutoff, wrote the change, and logged it. Same conversation on the surface. Completely different system underneath.
A good LLM chatbot with retrieval over your docs is fast to ship and hard to get badly wrong, because the worst case is a bad answer, not a bad action.
The 2am pattern we see constantly: an operations person doing triage by hand at night because "the tool does not talk to the other tool." That gap is usually one integration and one agent away from gone.
From the field. A team came to us certain they needed "a smarter chatbot" for support. We looked at the transcripts. 80% of the volume was not questions, it was requests to do something: refunds, address changes, plan swaps. A chatbot would have described how to do each one. What they needed was an agent that did them, with guardrails on the refund amounts. We shipped it in about five weeks. The chatbot framing would have solved the wrong problem.
This is the honest catch. A chatbot that gives a slightly off answer is a minor annoyance. An agent that takes a slightly wrong action files the wrong refund. So agents need engineering a chatbot does not:
That is why an agent build runs longer than a chatbot. A single-purpose agent is usually 3 to 4 weeks; a chatbot can be days. The extra time is the reliability, and the reliability is the point.
Both run on the same kind of model, so monthly inference is similar: roughly $50 to $2,000 depending on volume, and good engineering (caching, model routing) cuts that 3 to 10 times either way. The build is where they differ. Chatbots are cheaper because there is less to make safe. Agents cost more because the tools and the guardrails and the eval suite are real work.
Is an AI agent just a chatbot with extra steps? No. The difference is actions, not steps. A chatbot outputs text; an agent calls tools and changes state in other systems, then verifies the result. The chat interface is often the same, which is why they get confused.
Can one system be both? Yes, and most useful ones are. A support agent answers questions (chatbot behaviour) and also files refunds and updates records (agent behaviour). The design question is which of those your workflow actually needs.
Which is cheaper to build? A chatbot, usually by a lot, because it needs no tool integrations, guardrails, or evaluation suite. An agent costs more because those are required to make actions safe.
How do I know which one I need? Read your real transcripts or ticket logs. If most of the volume is people asking for information, build a chatbot. If most of it is people asking you to do something, build an agent.
Send us your ticket mix or a description of the workflow. We will tell you honestly whether it is a chatbot job or an agent job, and scope it either way. Free 30-minute call, response within two business days.
See how we approach AI agent development, or start a project.
