What is an AI agent?

An AI agent is a software system that combines an artificial intelligence model with data, tools and rules in order to carry out a task or a sequence of tasks. What sets it apart from a simple conversational assistant is that it acts: it calls tools and produces a result inside your systems.

Definition

An AI agent is a software system combining three things: an artificial intelligence model able to interpret language, access to data, and a set of tools it can call in order to act. What characterises it is that it does not merely answer — it carries out a sequence of steps until it produces a result.

The most common confusion is with the conversational assistant. An assistant waits for a question and returns text. An agent is triggered by an event, decides the intermediate steps itself, calls tools and writes a result into a system. The difference is not the model used: it is what you allow it to do.

What an agent is made of

• The model interprets the input in natural language and chooses the next action. This is the non-deterministic part of the system.
• The context gathers what the agent needs in order to act well: instructions, company rules, and data retrieved on demand — often through RAG.
• The tools are the permitted actions, exposed in a form the model can call: read a client record, create a quote, send an email, query stock. Each tool is a door opened deliberately; everything else stays shut.
• The guardrails frame the run: what is forbidden, what requires human validation, what must be logged, when to stop.
• The loop chains observation, decision and action, until the goal is reached or a limit is hit.

How a run unfolds

1. An event triggers the agent: a message arrives, a file is dropped, a time is reached.
2. The agent assembles its context: the request itself, the applicable rules, the relevant documents.
3. It picks a tool and calls it — for instance retrieving the history of the client concerned.
4. It reads the result, adjusts, and may call another tool.
5. It produces an output: a prepared document, a completed record, a draft reply.
6. Depending on the level of risk, the output is applied directly or submitted to a person.

What an agent brings

• It absorbs unstructured input. That is its real value: where a classic workflow demands a form, an agent accepts a freely written email.
• It chains steps without a script written in advance for every combination of cases, which avoids multiplying conditional branches.
• It copes with variations in wording without every turn of phrase having to be anticipated.

Limits and risks

A language model produces a plausible answer. Plausible is not verified. That property governs everything else.

• The well-phrased error. An agent is wrong with confidence, in impeccable prose. That is harder to spot than a crash.
• Variability. Two runs on the same input can differ. For a calculation or a pricing rule, the result must be left to deterministic code, not to the model.
• Cost per use. Every step consumes model calls. A talkative agent on high volume costs a lot without giving more.
• The action surface. A tool given to an agent is real power. An agent that can delete, pay or send externally must be framed accordingly.
• Instruction injection. If the agent reads content coming from outside, that content may try to give it orders. What comes from outside is data, never an instruction.

When an agent is justified — and when it is not

Justified when the input is free text or a heterogeneous document, when cases vary too much to all be anticipated, and when a human review remains possible before the action.

Not justified when the data is already structured and the rules are fixed: a classic workflow will be faster, cheaper and verifiable line by line. See the detailed comparison.

In systems that work well, both approaches coexist: the agent interprets input and prepares, deterministic code calculates and executes, the human validates whatever commits the company.

Frequently asked questions

Related articles

• AI agent vs classic automationA practical comparison between a deterministic workflow and an AI agent: predictability, cost, maintenance, handling of exceptions, and a simple decision criterion.
• RAGRAG makes a language model answer from your own documents, retrieved at the moment the question is asked. How it works, its steps, its limits, and how it differs from fine-tuning.
• Human-in-the-loopHuman-in-the-loop places a human check at one precise point of an automated process. Where to put it, what it protects, and why it should not be everywhere.