Definition of the AI agent
An AI agent is a software system that is given a goal rather than a sequence of instructions. To reach it, it relies on a large language model (LLM) which chooses for itself, at each step, the next action: look up information, call an application, ask for clarification or stop. The definition of the AI agent comes down to this loop: reason, act, observe the result, then start again until the goal is reached or a limit is met.
The term covers uneven realities. It refers as much to an assistant that chains two tool calls as to a system that works for a long time without intervention. The useful criterion is the degree of autonomy: who chooses the steps, the designer or the model? The more the model decides the path, the more justified the word agent is. The expression agentic AI refers to this family of systems.
What an AI agent is for
An agent is used to handle multi-step tasks whose sequence cannot be entirely written in advance: processing a case file from heterogeneous documents, qualifying and then routing a request, reconciling information spread across several applications, diagnosing a technical incident, writing and then testing code.
Three criteria help judge whether an agent is the right choice.
- The variability of the task. If all cases look alike, conventional automation costs less and behaves predictably.
- The ability to verify. It must be possible to check the result with a rule, a test or a person.
- Access to systems. An agent can only act through the interfaces opened to it. Without a usable programming interface, it is limited to writing.
The four components of an AI agent
- The model. It interprets the goal, reasons and produces the decisions. Its limitations (factual errors, sensitivity to the wording of instructions) become those of the agent.
- The tools. These are the functions the agent can call: document search, a database query, creating a ticket, sending a message. Each tool is described to the model by its name, its role and its parameters. The list of tools defines what the agent can do.
- The memory. In the short term, it corresponds to the context of the current task. In the long term, it relies on external storage that the agent consults, often following the principle of retrieval-augmented generation (RAG).
- The planning. The agent breaks the goal down into sub-tasks, chooses an order and revises its plan when a step fails. In multi-agent architectures, an orchestrator distributes the work between specialised agents.
AI agent, chatbot, automation: not to be confused
| Chatbot | Conventional automation | AI agent | |
|---|---|---|---|
| What it produces | An answer | The execution of a scenario | A result obtained through several actions |
| Who sets the steps | Not applicable | The designer, in advance | The model, along the way |
| Faced with an unforeseen case | Answers, without acting | Stops or fails | Looks for another path |
| Predictability | Medium | High | Lower, to be kept within bounds |
A chatbot converses. Even when backed by a language model, it changes nothing in the information system. Conventional automation (script, workflow engine, RPA software robot) executes a sequence set in advance: it is reliable as long as the case was foreseen. The agent occupies the space between the two. It acts like automation, but decides its path as an operator would. This flexibility comes at the price of lower predictability.
Human oversight of an agent
Because it acts, an agent is designed together with its limits. Four mechanisms recur in robust set-ups.
- Validation before action for anything irreversible or binding: a payment, an external message, a deletion.
- Permissions reduced to what is necessary. The agent has its own account, limited to the tools and data of its task.
- Execution limits. Number of steps, duration, cost, conditions for stopping and for handing over to a person.
- Traceability. Every decision and every tool call is logged to allow checking after the fact.
These safeguards also address a risk specific to agents: prompt injection, whereby content read by the agent (an email, a web page) seeks to divert its behaviour. For uses classified as high risk, human oversight is moreover a requirement of the European regulation on AI, or AI Act.
A concrete example: payment follow-up requests
Take an accounts department that receives emails from suppliers every day asking about the status of an invoice payment. A chatbot could explain the procedure. Automation could answer if every email contained an invoice number, always in the same place.
An agent proceeds differently. It reads the message, identifies the supplier, looks up the invoice in the accounting software package, finds for example that it is blocked for lack of a purchase order, drafts a reply explaining this and creates a task for the buyer concerned. Until trust is established, the reply is sent after validation by an accountant. Cases the agent cannot resolve are passed on to a person, with details of the searches already carried out.
For the design and production release of such systems, see our page dedicated to AI agents in the enterprise.