Guide

What is an AI agent? Definition, examples and cases in production

By Fabien Cavanna, Going for Growth · June 29, 2026 · 8 min read

In shortAn AI agent is a software system that pursues a goal autonomously: it reads an input (document, message, web page), analyzes it, decides on an action, produces a result and executes it, often by calling external tools. Unlike a simple chatbot that only replies, an agent chains several steps and takes action. A well-designed agent keeps a control point at every step to stay reliable.

What is the difference between an AI agent, a chatbot and an automation?

An AI agent pursues a goal and acts, a chatbot replies, an automation applies fixed rules: three different things. The three terms are often used as synonyms, and the confusion leads to poorly scoped projects.

  • A classic automation runs a series of fixed steps, written in advance. If the input data falls outside the planned scenario, it breaks or ignores the case. It decides nothing: it applies rules.
  • A chatbot replies to a message. With a language model, it can answer fluently and in context, but it stays inside the conversation: it informs, it will not modify a file, send a letter or update a system.
  • An AI agent pursues a goal. It interprets a situation, chooses the steps itself depending on the case, calls tools (a database, an API, a sending service) and produces a concrete action, not just text.

In other words: the automation follows a path, the chatbot holds a conversation, the agent makes decisions to reach a goal. These boundaries are detailed in the guide AI agent, chatbot and automation: what are the differences.

How an AI agent works

An AI agent, in production, comes down to five steps. The key point is not the sophistication of the model, but the control placed at every step. That is what separates a demo from a system that actually runs.

  1. Read: the agent retrieves the raw input (a scanned PDF, an email, the content of a page, a database row) and extracts it into a usable format. Control: check that the input is readable and complete before going further.
  2. Analyze: it understands what it has in front of it, identifies the useful elements, classifies, spots anomalies. Control: measure a confidence level, and flag for human review whatever is ambiguous.
  3. Decide: based on the analysis, it chooses the appropriate action among those allowed. Control: bound the possible decisions, and require validation for high-stakes cases.
  4. Produce: it generates the deliverable (a letter, a summary, a record, a structured response). Control: validate the expected format and review the content before delivery.
  5. Act: it executes, meaning it sends, saves, triggers the next step. Control: log every action, and plan a rollback in case of error.

A reliable agent is not the one that never makes a mistake: it is the one that detects that it might be wrong and hands off at the right moment. Control at every step is not an option, it is the heart of the job.

What are the types of AI agents?

There are four families of AI agents, from the simplest to the most elaborate: reactive, tool-oriented, orchestrated, voice. Not all agents are equal, and not all needs call for the same level of complexity.

Reactive agent (simple)

A reactive agent responds to an event with an action, without memory or planning: it receives an input, applies reasoning, returns a result. Enough to classify messages, extract a field, sort requests. No need for anything heavier when the need fits in a single decision.

Tool-oriented agent

A tool-oriented agent calls external tools to accomplish its task: query a database, run a search, write to a system, send an email. It is the most common type in business, because the value is almost always in the action on a real system, not in the conversation.

Orchestrated agent (multi-agent)

Several specialized agents split the steps of a complex process between them, coordinated by an orchestrator agent. One reads, another analyzes, a third writes, a fourth checks. This separation makes the system more readable and easier to make reliable: each agent has a clear responsibility and its own control point.

Voice agent

An AI voice agent adds speech in input and output: it transcribes what the caller says, processes the request like a classic agent, then replies in a synthetic voice. Useful for phone reception or appointment booking. The underlying logic stays the same: read, analyze, decide, produce, act. Voice is just one more channel.

Concrete examples in production

The best way to understand an AI agent is to look at what it actually does, not at what is promised about it.

Opposio: contesting a traffic fine

Opposio is a service for contesting French traffic fines, and a good example of an orchestrated agent in production. The flow follows exactly the five steps:

  • It reads the document provided by the user (the fine notice) and extracts the key information from it.
  • It analyzes the situation to identify the grounds for contestation that are admissible given the case.
  • It decides on the appropriate contestation strategy, or flags that there is no serious ground.
  • It produces the argued contestation letter.
  • It checks at every step, because a flawed legal letter is useless: the chain of agents reviews itself before delivering.

This is not a chatbot that chats about traffic law: it is a system that produces an actionable document, with a check placed at every link.

Anonymized business cases

Beyond this public example, the same pattern shows up in varied contexts:

  • Document processing: read incoming items (invoices, contracts, forms), extract the data, classify it and route it to the right system, with human review on doubtful cases.
  • Monitoring and takedown requests: watch sources, detect what matches defined criteria, prepare then send takedown requests, logging every action.
  • Content at scale: produce structured content in series from data, with a quality and format check before publication.

In each case, what matters is not that the agent exists, but that it runs every day without surprises. Systems that run, not promises.

Do you need to be technical, and how much does it cost?

Honest question, honest answer: not everything needs an AI agent. If a task is fully predictable and unambiguous, a classic automation will do the job, more simply and more cheaply. An AI agent makes sense when you need to interpret, judge varied cases, decide instead of applying a fixed rule.

On the technical side, you can build an agent without coding everything from scratch. A backbone like n8n lets you orchestrate the steps, connect the models and tools, and place the control points, all while self-hosting. That does not mean it is trivial: thinking through the breakdown into steps, handling edge cases and making the control reliable remains real work. The "how" is detailed in the practical guide Build an AI agent with n8n.

The cost of an AI agent depends on three measurable factors: the number and complexity of the process steps, the volume handled (which determines the language model consumption), and the level of reliability required (a legal agent demands more control than sorting emails). A simple, self-hosted agent on n8n mostly costs design time; the recurring costs come from model usage and hosting. Be wary of round numbers quoted out of context: the real trade-off is not the price of the tool, but the time it saves compared to the manual work it replaces.

If you are unsure whether an agent fits your case, a 30-minute call is often enough to settle it: sometimes the answer is that an automation is enough, and that is a useful outcome too.

Frequently asked questions

What is an AI agent?
An AI agent is a software system that pursues a goal autonomously: it reads an input, analyzes it, decides on an action, produces a result and executes it, often by calling external tools. It differs from a chatbot, which only replies, and from a classic automation, which follows fixed steps without deciding anything.
What is the best AI agent?
There is no best AI agent in absolute terms: the right agent is the one that solves your specific case with the right level of control. A simple agent is enough to sort messages, whereas a legal or document process requires an orchestrated agent with a check at every step. The useful question is not "which one is the best" but "which one fits my problem".
What is the price of an AI agent?
The cost depends on the complexity of the steps, the volume handled (which determines the model consumption) and the level of reliability required. A simple self-hosted agent mostly costs design time, with recurring costs tied to model usage and hosting. The real trade-off is the time saved against the manual work it replaces, not the price of the tool alone.
What is the difference between an AI agent and a chatbot?
A chatbot replies to a message and stays inside the conversation: it informs. An AI agent pursues a goal and takes action: it decides on the steps, calls tools and produces a concrete action (modify a file, send a letter, update a system), not just text.

Further reading

An AI agent built and maintained for you, in production?

Describe your process and I will tell you honestly whether it can be automated. A first 30-minute call, free and with no commitment.

Book a free 30-min call