Saltar al contenido principal
    Skip to main content

    AI agent

    An AI agent is a system that uses a language model to decide which steps to take, and acts in your software through tools: it reads an email, checks the CRM or drafts a reply. It follows instructions and a memory of the context. It does more than chat: it chooses the path case by case.

    In practice

    OpenAI, in its guide to building agents, defines them as "systems that independently accomplish tasks on your behalf" and boils them down to three parts: the language model, the tools and the instructions. Anthropic adds memory and separates two things people mix up: in a workflow, the model follows predefined code paths; in an agent, the model directs its own steps.

    In a small business, realistic uses are specific:

    • Sorting the mail in a shared inbox and leaving a draft reply or a task for the right person.
    • Pulling the details out of invoices or delivery notes and putting them in a spreadsheet or the ERP, with fixed rules that check, for example, that the net amount plus VAT equals the total.
    • Spotting quotes in the CRM that nobody has answered and drafting a reminder for a person to send.
    • Giving the first reply in your website chat or on WhatsApp from your FAQs, and handing the case to a person when it is unsure.

    OpenAI's guide keeps agents for decisions full of exceptions, rules that are hard to maintain and processes that depend on unstructured text. If the work always follows the same steps, an ordinary automation is enough, for example a workflow in n8n. If the agent talks to customers, since 2 August 2026 the EU AI Act requires that they know they are dealing with an AI, unless it is obvious (Article 50).

    When it pays off: AI agents for business. What it costs: how much an AI chatbot costs.

    Common mistakes

    1. Giving it more access than it needs. Anthropic recommends least privilege so that a trick that works does as little damage as possible. INCIBE, Spain's national cybersecurity institute, adds that an AI should never pay, delete, grant permissions or send sensitive data on its own.
    2. Letting it act without approval. OpenAI advises that sensitive, irreversible or high-impact actions should wait for a person to approve them. In n8n you can pause the workflow until someone approves the use of a specific tool.
    3. Forgetting that it reads other people's content. An email or a PDF can carry hidden orders for the model: that is prompt injection. The Spanish Data Protection Agency (AEPD) proposes its rule of 2: an agent that reads content from anyone, accesses sensitive information and acts on its own is a set-up that should not be allowed.
    4. Trusting what it says. Anthropic admits that even the most advanced models can generate incorrect text, and that its techniques reduce this without eliminating it. Feeding it your documents with RAG helps, but check what goes out.
    5. Not signing with the provider. Whoever supplies the model processes data on your behalf, so it is a data processor and the GDPR requires a contract (Article 28).

    Sources

    Articles that use this term