AI agents for companies: what they are, what they are not, and when one is worth it

An AI agent is a program that receives a goal, uses a language model (LLM) to decide the next step, runs tools (query a database, send an email, call an API) and repeats the loop until it finishes or a person stops it. The difference from a chatbot is not the model: an agent acts on real systems.

What an AI agent is

An agent has four parts: a goal written in plain language, a model that reasons, a set of tools with defined permissions, and a limit on steps or budget. If any of the four is missing, it is not an agent: it is a long prompt or a script with no control.

Chatbot versus agent: the table that prevents bad purchases

The practical rule: if the task ends in an answer, a chatbot with the company's knowledge is enough. If the task ends in an action on another system (create, move, send, approve), you need an agent.

When building an agent pays off

• The task repeats several times a day and today a person does it by copying data between screens.
• The rules fit in a document of one or two pages; if nobody can write them down, the agent will not guess them either.
• There is an API or an integration for every system it touches; the agent should not drive screens blindly.
• A mistake is reversible or goes through human approval before it takes effect.

If a task fails any of the four points, fix the process or the integration first; automate afterwards.

Minimum controls before going to production

1. Least privilege. The agent uses its own credentials, with access only to what it needs, and they can be revoked without touching anyone else.
2. A log of every action. What it was asked, which tool it used, what it answered and what it cost. That log is the evidence ISO/IEC 42001 asks for.
3. Hard limits. Maximum number of steps, money and time per task. When one is exceeded, it stops and notifies.
4. Human approval for every irreversible action: payments, deletions, bulk sends.
5. Tests with anonymized real cases before connecting it to live data.

How to start without risking operations

Pick one boring, measurable task, for example classifying support emails and opening the ticket with the data already extracted. Measure how long it takes today, run the agent in read-only mode for a week (it proposes, it does not execute), compare its proposals with what the team did and, once accuracy is stable, grant it write permission. That same log later serves as quality evidence.

An agent that leaves no trace cannot be audited, and what cannot be audited cannot be certified.