AI agent vs chatbot: what's the difference, and what is agentic AI?

"We want an AI agent" is one of the most common things I hear from clients. Often what they need is a good chatbot. Sometimes it's the other way round. The words are used loosely, so here is a plain explanation of AI agents vs chatbots, what "agentic AI" adds, and how to decide which one your business needs.

AI agent vs chatbot: the short answer

• A chatbot answers. You send a message, it replies. It may look things up in your documents, but its job ends with the reply.
• An AI agent acts. You give it a goal, and it decides on steps, uses tools (your CRM, calendar, database, email, a browser, code), checks the results and keeps going until the goal is done or it needs help.

The difference isn't the model. The same large language model can power both. The difference is what the system around the model allows it to do: only talk, or also take actions over several steps.

What a modern chatbot can do

Chatbots have come a long way from scripted decision trees. A well-built LLM chatbot can:

• answer questions in natural language, including in dialects and mixed languages,
• ground its answers in your own content (policies, catalogue, past tickets) so it doesn't make things up,
• remember the conversation and ask clarifying questions,
• hand over to a person with a summary when it's out of its depth.

For many businesses this covers most customer conversations. A WhatsApp assistant for a Moroccan shop or an Arabic customer-service chatbot is mostly a chatbot, with a few carefully chosen actions.

What makes something an AI agent

A system starts being an agent when it can do three things:

1. Use tools. Call an API, query a database, send an email, update a record.
2. Plan over multiple steps. Decide what to do next based on what just happened, not follow a fixed script.
3. Work towards a goal, not just respond to the last message.

A simple example: a customer writes "I want to move my appointment to next week." A chatbot can explain how to reschedule. An agent can check the calendar, find free slots, offer two options, book the one the customer picks, and send a confirmation.

In between, there's a large grey zone. A chatbot that can look up an order status is using a tool, but it's still mostly answering questions. That's fine. Labels matter less than what the system is allowed to do.

Agentic AI vs AI agent

These two terms are close, which is why people search for the difference.

• An AI agent is a specific piece of software: one model in a loop with instructions and tools.
• Agentic AI describes the broader approach or property: systems that act with some autonomy towards goals. An agentic system might be one agent, or several agents working together with handoffs, reviewers and human checkpoints.

So "agentic AI" is the category, and "an AI agent" is one building block in it. When a system has several agents with different roles, it's a multi-agent system, and the design questions shift to how they coordinate.

Business examples side by side

• Customer support. Chatbot: answers FAQs about delivery and returns. Agent: also processes the return, issues the label and updates the order.
• Sales. Chatbot: answers questions about pricing on your site. Agent: qualifies the lead, writes a personalised follow-up, logs it in the CRM and schedules the next touch. My Sendline platform works in this direction: replies picked up by email can trigger AI-written follow-ups inside automation flows.
• Internal knowledge. Chatbot: answers staff questions from your manuals. Agent: also opens the IT ticket or fills in the expense form.
• Operations. Chatbot: tells you what a report says. Agent: gathers data from several systems, writes the weekly report and flags what changed.

Risks and guardrails

Agents are more useful because they can act. They're riskier for the same reason. A chatbot that gets something wrong produces a bad reply. An agent that gets something wrong can send the wrong email, book the wrong slot or change the wrong record.

The guardrails I put on every agent:

• Least privilege. Give it only the tools and permissions the task needs. Read-only where possible.
• Human approval for irreversible actions. Payments, outgoing messages to customers, deletions. Start with the agent drafting and a person approving.
• Limits. Cap the number of steps and retries so it can't loop or run up a bill.
• Logging. Every tool call recorded, so you can see why it did what it did.
• A clear handover. When it's unsure, it stops and asks a person.

Cost

Chatbots are usually cheaper per conversation: one or two model calls per reply. Agents make several calls per task, re-reading context each time they plan, call a tool and check a result. That doesn't make agents expensive by default, but it does mean cost should be measured per completed task, not per message. There are good ways to keep it down, like routing routine steps to smaller or local models; I cover them in how to reduce LLM API costs.

How to choose

Ask three questions:

1. Does the job end with information, or with something done? Information: chatbot. Something done in another system: agent.
2. How bad is a mistake? If an error is cheap to catch and undo, autonomy is fine. If not, keep a person approving, whatever you call the system.
3. Is the process clear enough to write down? Agents need clear rules and good tool access. If the process lives in someone's head, fix that first.

A good path for most businesses: start with a chatbot that answers well and hands over cleanly, then add actions one at a time where the data shows they're safe.

Working together

I build both: grounded chatbots and AI agents that act on real systems, from the model to the production software around it. See my case studies.

If you're not sure which one you need, send me a short description of the workflow and I'll tell you honestly.