What an AI agent is (and how it differs from a chatbot)
A chatbot answers. An agent acts. The distinction is not a matter of vocabulary: it is four capabilities you can verify before you sign.
Over the last two years the word “agent” moved from academic papers to sales brochures, and lost its meaning along the way. Today almost anything that answers messages is sold as an AI agent: button menus from ten years ago, forms in different clothes, decision-tree chatbots.
The difference matters for the price and, above all, for the responsibility. A chatbot that gets it wrong gives a strange answer. An agent that gets it wrong cancels an appointment, sends an email or changes a record.
What an AI agent is
An AI agent is a program that uses a language model to decide what to do when faced with a request in natural language, and that can also execute real actions —check a calendar, write to a database, send a message— and verify the result before replying. A chatbot, by contrast, only produces text.
The key word is execute. A chatbot tells the customer “you can book Thursday at 3”. An agent checks the calendar, sees that Thursday at 3 is taken, offers 4, and when the customer accepts, writes the appointment in. The first informs; the second acts, and can therefore get things wrong in ways that cost money.
The four differences you can actually check
“It is smarter” is not a verifiable difference. These four are, and they can be asked about in a fifteen-minute demonstration.
1 · Tools: can it touch anything outside the conversation?
A language model, on its own, only generates text. To do anything else it needs tools: functions the program exposes to it and that the model can ask to have executed —“check Thursday's availability”, “save this appointment”, “look up this customer”—. The program executes them and returns the result to it.
Without tools there is no agent. There is a very convincing text generator, which is exactly what produces made-up answers: if the model cannot check the calendar, it imagines it.
How to check it: ask them to show you, in the demonstration, what happens when the external system does not respond. A real agent has a planned answer for that. A chatbot in disguise will carry on answering confidently and with false data.
2 · Memory: does it remember who you are?
Language models remember nothing between one conversation and the next. What looks like memory is the program storing the history and passing it back to the model on every turn.
That has two practical consequences. The first is that memory has to be built on purpose, and someone has to decide where it lives, how long it lasts and what gets deleted. The second is that each turn of the conversation costs a little more than the one before, because the history grows; that is why serious systems summarize long conversations instead of dragging them along whole.
3 · Autonomy: does it decide or follow a script?
A decision-tree chatbot walks through a script somebody drew: if the customer says A, go to B. It works perfectly as long as the customer says A, and it breaks with “sorry, are you still open on Saturdays or did the hours change?”.
An agent has no script: it has a goal, a set of tools and some rules. Faced with the same question it decides what to check and in what order. That is what lets it handle what nobody anticipated, and also what makes it harder to test: there is no diagram to audit, there is a behavior that has to be put to the test.
4 · Verification: who tells it no?
This is the one almost nobody asks about and the one that costs the most.
In a well-built system, the model proposes and the code disposes. The hard rules of the business —opening hours, closed days, how long each service takes, how many people fit at the same time— are validated in code, outside the model. If the model proposes an appointment at 8 on a Sunday evening, the code rejects it and returns the reason so that it can reformulate.
A system where the model alone decides about the calendar will work very well in the demonstration and will book an impossible appointment in the second month. We treat this as a rule of construction, not as a preference: in the anatomy of an appointment assistant it is described piece by piece.
Chatbot, automation and agent, side by side
| Scripted chatbot | Classic automation | AI agent |
|---|---|---|
| What it understands | ||
| Buttons and exact words | Nothing: it reacts to events | Natural language, with typos and roundabout phrasing |
| What it does | ||
| Replies with fixed text | Runs steps that are always the same | Decides which steps to run, and runs them |
| Faced with the unforeseen | ||
| Breaks or repeats the menu | Fails and raises an alert | Improvises — for better and for worse |
| Cost per use | ||
| Practically zero | Low and predictable | Variable: you pay the model's consumption and, if messaging is involved, each message |
| Where it shines | ||
| Repeated, closed questions | Stable, high-volume processes | Open conversation with actions behind it |
The three columns solve different problems, so the useful question is not which is better but which is better for your specific case. A good part of the work of an honest consultancy consists of telling you that your case is solved with classic automation or that, on the contrary, the complexity of the task calls for an elaborate agent. That judgment is the subject of tool or custom development.
When an agent is the wrong answer
- When the process has no variation. If the same thing always happens, an automated workflow does it more cheaply, faster and without surprises.
- When the error is not tolerable. Charges, dosages, legal or tax calculations. An agent can prepare the work; the decision is signed by a person.
- When there is nowhere to look up the truth. If the information lives in someone's head and not in a system, the agent has nothing to check. First the data is put in order, then it is automated.
- When the volume does not justify it. Twelve messages a day are answered better and more cheaply by a person.
A point of vocabulary. “Agent” is also used for systems of several models that coordinate with each other, and for programming assistants that edit files. They are relatives of the same concept —model, tools, decision loop— applied to another trade. In this article we are talking about business use: serving, checking and acting on a company's systems.
What is worth asking before signing
- Which specific tools can it use, and what happens if one of them fails?
- What does the model decide and what is validated in code?
- Where is the conversation history stored, how long does it last and who can read it?
- How does it escalate to a person, and how is control handed back to it afterwards?
- What happens the day the model provider changes its price or discontinues a version?
The full twelve questions, with the answer that should reassure you and the one that should worry you, are in the buying script.
Frequently asked questions
Is an AI agent the same as ChatGPT?
No. ChatGPT is a general-purpose conversation product; an AI agent is a system built for a specific task in a company, which uses a language model —it may be the same one behind ChatGPT, or another— as one piece among several, alongside tools, rules and memory of its own.
Can an AI agent make up data?
The model that drives it can. That is why a well-built agent does not leave the last word to the model: it looks up the real information through tools and validates the business rules in code. If you are offered an agent that “never gets anything wrong”, you are being sold a demonstration, not a system.
How much does it cost to maintain an AI agent?
It has three distinct costs: the model's consumption, which is paid by use and rises with the number of conversations and their length; the infrastructure it runs on; and maintenance, which is the most underestimated, because models change version and business rules change too. Any quote that mentions only the first is incomplete.
Do I need an AI agent or is an automation enough?
If the process always happens the same way and the input is a form or an event from another system, an automation is enough and it is cheaper and more predictable. The agent is justified when the input is free human language and someone has to decide what to do with it.
Sources and notes
- The descriptions of behavior in this article come from building and operating Navhera's own conversational agents.
- The distinction between text generation and tool execution follows the public documentation of the model providers: Anthropic, Tool use with Claude, and OpenAI, Function calling. Both describe the same mechanism: the model executes nothing, it returns a structured call that is executed by the program integrating it.
Written by the Navhera team and reviewed before publishing. If you spot an error, write to us and we will correct it with a note.