What are AI agents and what are they for?

WHAT IS AN AI AGENT?

An AI agent is a system capable of acting autonomously within defined limits: you give it a goal—not instructions, a goal—and it decides which tools to use to achieve it.

That's the difference with everything mentioned before:

→ Classic automation follows fixed rules: if X happens, do Y.
→ A chatbot answers questions, but doesn't act.
→ An agent receives a goal ("manages product availability inquiries"), consults the systems it needs, prepares the response, and executes—or requests approval.

The analogy I always use, and which I find the most honest:

An AI agent is a very, very, very well-trained intern.

It doesn't replace a whole person. It performs specific tasks. It needs you to explain who you are, how you work, and the limits of its autonomy. And for the first few weeks, someone reviews its work.

If someone sells you an agent that "works on its own from day one, without supervision"... be suspicious.

Agents are unpredictable by nature (hence the term "hallucinations"): they need context, clear boundaries, and ongoing supervision. Just like any new hire.

WHAT TYPES OF AI AGENTS EXIST?

Agents are classified by their level of autonomy, not by their technology. There are three levels:

1. Supervised Assistants. The agent prepares the work; a person reviews and approves it before anything is sent. This is where any company should start. Example: the agent drafts a response to a customer, and your team approves it with a click.
2. Semi-Autonomous Agents. They perform low-risk tasks on their own and escalate sensitive ones to a person. Example: they automatically answer inquiries about deadlines and stock, but any complaints go directly to the team.
3. Autonomous Agents. They make decisions and execute tasks from start to finish within their limits. Agents already exist that search for suppliers, request quotes, compare proposals, and negotiate deadlines—with human review where it matters.

💡 The key that almost no one talks about:

The companies that are best implementing AI—as confirmed by major Catalan industrialists at a recent roundtable—started at step 1. Hundreds of cases as supervised assistants before granting autonomy to anything. The order is not a detail: it's the method.

WHAT ARE AI AGENTS USED FOR IN A COMPANY? (REAL EXAMPLES)

They are used to handle the third of the work that doesn't require critical thinking: repetitive tasks, with a structured approach and where the information is readily available in a system.

Examples in production today:

→Customer service:classify each incoming inquiry, answer repetitive questions (stock, lead times, technical specifications) in the customer's language, and escalate complex ones. In an international B2B e-commerce platform, 3 out of 4 inquiries no longer require team attention.
→Sales prospecting:research each target company and prepare a well-documented initial email—with approval from a client before sending. Our own agent writes these prospecting emails for KAIzen. Every email we send is a demo.
→ Administration: extract data from invoices, match them with orders and delivery notes, and leave only exceptions to the team.
→ Purchasing: request quotes from suppliers and compare proposals, with the final decision in human hands.

What they DON'T do: negotiate a sensitive claim, maintain the relationship with a key client, decide on an exception. The decision remains with your team. And that's where the value lies.

5. SUBSCRIPTION OR DEVELOPMENT: THE DECISION THAT ALMOST NO ONE TELLS YOU ABOUT

There are two main ways to automate processes with AI:

→ Rent: Subscribe to a SaaS per process. It's fast at first, but over time subscriptions accumulate, your data isn't yours, and the monthly cost only increases.

→ Develop: Build systems on top of the tools you already use, with self-managed solutions like n8n, with a documented, auditable system that you own. When the developer finishes, the system remains yours.

Large Catalan companies understand this clearly: they didn't buy tools; they built internally with governance, their own repositories, and trained teams.

HOW IS AN AI AGENT IMPLEMENTED? (AS A NEW EMPLOYEE)

Deploying an agent is much more like onboarding a person than installing software.

The process, in four steps:

1. Onboarding: who we are, how we work, how we treat clients, and what the objective is.
2. Limits: which tools and data it has access to, and what it must not touch.
3. Supervision: human review of everything that matters, from day one.
4. Continuous evaluation: measure, correct and expand autonomy only when the data supports it.

When someone tells me, “We tried AI and it didn’t work”, I almost always find the same story: they hired an intern, didn’t explain anything to them, gave them free rein over everything… and were surprised by the result.

💡 And one prerequisite that is not optional:

The agent needs a well-organised process in place. If the process exists only in someone’s head and in three separate Excel spreadsheets, the agent will simply exacerbate the chaos. The challenge is 80 per cent business, 20 per cent technology.

OWN AGENT OR RENTED AGENT?

The final question, and the one that determines the most money over three years: Do you own the agent or rent it?

→ An agent rented as SaaS is just another subscription: it grows with your usage, your data lives in its black box, and if you stop paying, it disappears.
→ An agent built on open tools (self-hosted n8n, for example), documented and auditable, belongs to your company when the project ends. Just like your ERP.

Large companies didn't buy tools: they built an asset—their own agent repositories, governance, and trained teams. You don't need their budget to apply the same principle. You need a process.

IN BRIEF

Un agente de IA es un becario altamente cualificado: trabaja para alcanzar objetivos, dentro de unos límites y bajo supervisión. Resulta útil para ese tercio del trabajo que no requiere criterio. Se implementa como un empleado, no como un programa informático. Y lo importante no es qué modelo de IA utilice, sino si se empieza con supervisión y si el sistema acabará siendo tuyo.

¿Y en tu empresa? ¿Qué tarea le asignarías mañana a un becario altamente cualificado e incansable?