The word has been everywhere for two years, and it gets attached to almost anything. An AI agent is not a slightly smarter chatbot, and it is not a classic automation repainted in AI colours. It is a different category of tool, with real strengths and equally real limits.
The distinction matters because it decides what you can delegate. A chatbot answers. An agent acts. And anything that acts on your behalf inside your business deserves to be understood before it gets plugged in.
Definition
An AI agent is a program that receives a goal rather than an instruction, breaks that goal into steps on its own, uses tools to carry them out, and keeps going until it reaches a result or establishes that it cannot.
Three words in that definition carry all the weight.
Goal. You do not say "write a follow-up email for this client". You say "follow up with every client whose invoice is more than fifteen days late". The agent has to work out who those clients are before it knows what to write.
Tools. A language model on its own can only produce text. An agent can query a database, call an API, read a file, post a message. Those accesses are what turn a conversation into work actually done.
Keeps going. The agent does not return an answer and stop. It looks at the result of each step, judges whether it moved closer to the goal, and decides what comes next. That loop is what sets it apart from everything that came before.
A simple test: if you can write the exact list of steps in advance, you do not need an agent, you need a classic automation, which is cheaper and more reliable. An agent earns its place when the steps depend on what you find along the way.
Agent, chatbot or automation: what separates them
These three tools overlap enough to be confused, and differ enough that picking the wrong one gets expensive.
| Tool | You provide | It produces | Cost and reliability |
|---|---|---|---|
| Classic automation | The exact sequence of steps | Always the same result | Low cost, very reliable |
| Chatbot | A question | An answer in text | Medium cost, variable result |
| Agent | A goal | Work actually done | High cost, variable result |
The right-hand column is the one people look at last, when it should come first. An agent running for ten minutes costs far more than a single chatbot answer, and it can get an intermediate step wrong without anything flagging it.
The three building blocks
The model
This is the decision engine. On every pass through the loop it receives the history of what has happened and chooses the next action. A more capable model makes better decisions and heads down fewer dead ends, which often offsets its higher unit price: it finishes the job in fewer steps.
The tools
These are the capabilities you grant: reading your orders, sending an email, creating a client record, searching the web. Every tool you add widens what the agent can do, and widens just as much what it can break. The tool list is the real design decision, far more than the choice of model.
The loop
The mechanism that chains it together: decide, act, observe, repeat. This is also where the stopping question lives. An agent with no cap on iterations can spend a long time on a problem it cannot solve, billing you for every attempt.
Three uses that hold up
Sorting the inbox
An agent reads incoming messages, spots the sales enquiries, creates the record in your CRM and drafts a reply for you to approve. The steps vary with the content of each message, so the agent is justified.
Documented monitoring
You ask what changed at three competitors this month. The agent visits the sites, compares against what it recorded last time, and reports only the differences. A classic automation would send you the full pages every week.
Meeting preparation
Before a call, the agent gathers the client history, their recent messages, their subscription status, and hands you a one-page brief. The work is dull, repetitive, and its shape depends on what it finds.
What these three share: a mistake by the agent is visible before it has consequences. An email drafted but not sent, a brief you read. Be wary of uses where the agent acts with nobody looking.
Best practices
Start as narrow as possible. An agent with two tools and a precise goal works. An agent with fifteen tools and a vague mission wanders. Widen the scope only once the narrow version is reliable.
Keep irreversible actions under your hand. Send, publish, delete, pay. Those four verbs deserve human approval, at least for the first few weeks. The rest can run on its own.
Set a ceiling. A maximum number of steps, a monthly budget, an alert when it is exceeded. Without a ceiling, you discover the bill at the end of the month.
Log everything. You need to be able to read back what the agent decided and why. The day it does something foolish, that trail is the only thing that will tell you whether to fix the instruction, remove a tool or change the model.
Measure against what you had. The right question is not "does the agent work". It is "does it beat what I was doing before, once I count the time spent supervising it". Plenty of impressive agents cost more than the task they replace.
Frequently asked questions
Do you need to know how to code to build an agent?
Not any more. Automation tools such as n8n let you assemble an agent by connecting blocks, without writing code. Knowing how to code is still useful for edge cases, and above all for understanding what is happening when the agent misbehaves.
How much does an agent running continuously cost?
It depends on the number of steps and the model in use, and the gap between two setups can be fiftyfold. An agent handling ten requests a day at five steps each stays very affordable. An agent watching permanently and restarting a loop every minute quickly becomes your main line of spend. Always count steps per day, never number of agents.
Can an agent be wrong without anyone noticing?
Yes, and that is its main flaw. It can invent a fact in the middle of a chain of steps and build everything after it on that basis, with complete confidence. This is why uses where a person reads the result remain the safest, and why irreversible actions should stay behind approval.
How do you learn to build your own agents?
The most effective route is to start with a simple automation and make it gradually more autonomous, rather than aiming for a complete agent on day one. Our n8n course covers that progression, from your first automation to an agent that chooses its own steps, without going through code.